GLM models compared cost vs quality at every effort level

Data analysis · aggregated from 82 sources · 422 benchmarks · 706 measurements
GitHub
Alexandre Gensse

What each recent GLM model actually costs, at every effort level, and the quality it delivers — fused from public benchmark measurements for every (model, effort) pair, on the same scales as every other maker's on this site.

The best value today is GLM-5.3-Flash at max effort — 4.6× cheaper than the price trend for its quality (expected score 58.7 % on the benchmark panel).

Quality is the expected score across the whole benchmark panel, not one benchmark: the type of task can move it, see what moves everything.

Measurements
82 sources · 422 benchmarks · 706 measurements
Updated
19 Nov 2026
👑 Best overall
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Grunt work
GLM-5.3-Flash · Low Cost 1.0× · Score 48.7 %
115
Classification, tagging, extraction, routing, log/PR triage — run at scale, where throughput and unit cost dominate.
Everyday tasks
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Routine coding, refactors, unit tests, summaries, first-draft agent steps — solid work that doesn't need the frontier.
Advanced reasoning
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Production code review, architecture, hard debugging, customer-facing reasoning — you need essentially flagship quality.
Cutting-Edge thinking
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Research-grade reasoning, novel or ambiguous problems, the hardest agentic runs — a few extra points of capability are worth a premium.

Quality vs. cost of every GLM model, at each effort level

One point per couple: a model at one configuration (thinking off, minimal, low → max, thinking on); no line joins a model's couples. Only the couples of the Pareto frontier are labelled (model · effort), where there is room; hover or tap any point for its model, effort, cost and score. Cost is a multiple of the cheapest couple shown; quality is placed on the model's latent scale and labelled with the score it predicts on the benchmark panel. Below the weakest tier target the quality axis is compressed, so the models short of today's level sit together at the bottom. Tier bands (off by default) shade the four usage tiers of the tier picker below : each band runs from that tier's target up to the next tier's — the qualities that clear its bar but not the next one. Uncertainty ovals (off by default) show each couple's 16–84 % interval, width = cost, height = quality; hover to bring the one under the cursor forward.

Best GLM model for the price — the Pareto frontier

The non-dominated (model, effort) couples : nothing offers more quality at equal-or-lower cost. Solid = frontier, faded = dominated; the frontier couples are labelled (model · effort) where there is room — hover or tap any point for its model, effort, cost and score. No line joins a model's couples. The frontier couples are the candidates of every pick below. The faint grey line is the price trend — what a given quality costs today, fitted on every couple of the latest model of each family. Tier bands (off by default) shade the same four usage tiers as on the landscape above.

Value — distance to the price trend

The price trend — ln(cost) = a + λ·θ, θ the latent quality — is what a given quality costs today across the 6 couples of the latest model of each family, not just the frontier: it is the going rate of the models on sale. Earlier generations, dearer per point than their successors, would flatten it and make quality look cheap; it does not move when older models are shown (R² ≈ 0.57, drawn faint grey above). Each couple weighs by its uncertainty across the line, both axes counted. On the latent scale the trend is a straight line: each step of quality costs the same ratio more, wherever it sits. A couple's distance to it says how many times cheaper or dearer it is than the trend charges for its quality — the same gap read on the other axis is how much more quality it gives than the trend at its cost. No couple serves as a reference. Only the frontier couples are listed.

Model · effortcostexpected scorevalue index
trend = 100
GLM-5.3-Flash · Max3.0×58.7 %456
GLM-5.3-Flash · High1.6×53.1 %223
GLM-5.3-Flash · Low1.0×48.7 %115
GLM-5.3 · Max33×60.5 %62

Value index = 100 × how many times cheaper than the trend the couple is for its quality (100 = the trend). Green = above 100, red = below; intensity = distance to the trend in decades, capped at one. A couple on the Pareto frontier can still be red : being non-dominated does not mean being good value for what it delivers.

Best value by task tier — what to actually run

Consolidating the frontier into a decision. Four target qualities are derived from the data, not fixed: they spread evenly between the frontier's weakest and strongest couples, each end drawn 5 % of the span inward, so that the bottom tier follows the weakest couple available and the top tier the best — as models improve, the targets move with them. The width follows the spacing on one rule : adjacent windows cross at half weight midway between their targets. For each tier we pick, among the frontier couples, the one that maximises window × eβ·λθ ⁄ cost, λ the slope of the price trend and β the tier's price of quality as a multiple of the market's : grunt work pays a quarter of it (β = 0.25), everyday tasks 0.63, advanced reasoning 1.59, cutting-edge thinking four times (β = 4) — evenly spread in ratio, symmetric about the market. Below its target a couple is penalised by a Gaussian window ; at or above it the window is 1, and eβ·λθ ⁄ cost rewards smoothly a couple that brings more quality than the tier is ready to pay for its extra cost. Cost counts in ratios, as people perceive prices : twice as dear weighs the same at every price. The big figure is the pick's value index : 100 = the price trend, 500 means five times cheaper than the trend charges for its quality, 50 twice as dear. The crown is the frontier couple that sits furthest below the trend. Sliders below override the targets, widths and prices of quality.

Target 49.2 % – Grunt work
GLM-5.3-Flash · Low Cost 1.0× · Score 48.7 %
115
Classification, tagging, extraction, routing, log/PR triage — run at scale, where throughput and unit cost dominate.
Target 52.6 % – Everyday tasks
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Routine coding, refactors, unit tests, summaries, first-draft agent steps — solid work that doesn't need the frontier.
Target 56.2 % – Advanced reasoning
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Production code review, architecture, hard debugging, customer-facing reasoning — you need essentially flagship quality.
Target 59.8 % – Cutting-Edge thinking
GLM-5.3-Flash · Max Cost 3.0× · Score 58.7 %
456
Research-grade reasoning, novel or ambiguous problems, the hardest agentic runs — a few extra points of capability are worth a premium.
👑 Best overall
GLM-5.3-Flash · Max
Cost 3.0× · Score 58.7 % · Value index 456

Picked as the frontier couple that sits furthest below the price trend : it costs 4.6× less than the trend charges for its quality — or, read on the other axis, it scores 6.3 points of expected score above what the trend gives for its cost. Its value index is that ratio times 100: 100 = the trend, fitted on every couple of the latest model of each family, so today's going rate of the models, not the frontier; no couple serves as a reference.

Tune the tiers (target, σ, price of quality β)
Normalized cost matrix — GLM effort levels compared

Cost per model × effort as a multiple of the cheapest couple, sorted by decreasing expected score. Every cell is computed from the measurements by the fusion model described in the method below: a cost on a task of typical size, with the couple's own 16–84 % interval. Intensity = cost.

Modellowmediumhighxhighmaxthinking onExpected score
top effort
GLM-5.3
9.47.4–12
—
2115–28
—
3328–39
—60.5 %
GLM-5.3-Flash
1.00.78–1.3
—
1.61.2–2.0
—
3.02.3–3.8
—58.7 %

"Expected score · top effort" = the score the model at its top effort is expected to reach on the benchmark panel, from the same fit as the landscape. A model measured at a single setting (solo : no effort parameter, see the measurement sources) fills one merged cell. A thinking off column appears when a model is measured with its thinking switched off, a minimal one for the lowest level some models have below low; each is a configuration of its own, read from its own measurements. n/a : effort not exposed, or measured by a single publisher (a couple is shown only once two publishers measured it).

Value-index matrix — every model and effort against the price trend

The value index of every couple shown, model × effort, in the same order as the cost matrix : 100 = the price trend, what a given quality costs today across the latest model of each family ; 500 means five times cheaper than the trend charges for that quality, 50 twice as dear. Under each value, its 16–84 % range.

Modellowmediumhighxhighmaxthinking on
GLM-5.3
2716–46
—
4828–84
—
6246–83
—
GLM-5.3-Flash
11570–188
—
223137–362
—
456299–694
—

Green = above 100, cheaper than the trend for its quality ; red = below ; the deepest green is the best index shown, the deepest red the lowest, on a log scale. A high index does not make a couple a pick : the tiers weigh it against their target quality.

What moves everything — four factors behind the numbers

The tier picks above are an order of magnitude. Four factors, each measured elsewhere, modulate them strongly — keep them in mind when reading the tiers and the matrix.

Task type

Effort and model choice don't weigh the same depending on the nature of the work — review, reasoning and pure coding reward effort very differently.

Complexity

The harder the task, the more effort costs — up to the point where extra thinking starts to hurt (the inverted-U).

Generation

Each release buys the same quality for less: turn on Older models to set each family's earlier models beside its current one, on the same scales.

Cache & harness

Two invisible factors — cache-read rate and the surrounding agent harness — that can dominate both cost and score.

Data Sources

Every source below measured at least two (model, effort) couples on the same task; 10115 cost and 23548 score measurements enter the fusion. Each row below is one such source, with the actual configuration we verified (harness · effort), its type, and the couples it links.

SourceActual config (harness · effort)TypeLinked couples (same task)
AA CritPt
artificialanalysis
AA model payload · CritPt physics, 4 models × ladder + 3 at max; floor scores blank ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low · GLM-5·think · GLM-4.7·think
AA GDP-PDF
artificialanalysis
AA model payload · GDP-PDF all-pass, 4 models × ladder + 3 at max; noisy ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low
AA GDPval v2
artificialanalysis
AA evaluation page · 220 tasks, Elo; cost derived from AA's published token counts ✓sweepGLM-5.3·low · GLM-5.3·max · GLM-5.1·think · GLM-5.3-Flash·max · GLM-4.7·think
AA HLE
artificialanalysis
AA model payload · HLE no tools, 4 models × ladder + 3 at max ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low · GLM-5·think · GLM-4.7·think
AA Index
artificialanalysis
AA-Index · cross-model at max + Opus 5 sweep low→max ✓sweepGLM-4.7·think
AA MLCR
artificialanalysis
AA model payload · MLCR-AA, score-only; Sonnet 5.5 max-only, Opus 5 full ladder, 5 others at max ✓sweepGLM-5.3·max · GLM-5.3-Flash·max
AA Omniscience
artificialanalysis
AA model payload · Omniscience accuracy (not the penalised index) ✓sweepGLM-5.3·low · GLM-5.3·max · GLM-4.7·think · GLM-5·think · GLM-5.1·think · GLM-5.3-Flash·max
AA SciCode
artificialanalysis
AA model payload · SciCode, 4 models × ladder + 3 at max ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low
AA TB 4.0
artificialanalysis
AA model payload · Terminal-Bench 4.0 (AA harness), 4 models × ladder + 3 at max; floor scores blank ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low
AA Terminal-Bench-Science
artificialanalysis
AA model payload · Terminal-Bench-Science 0.1 (AA harness, separate from Vals's board), score-only, 4 models × ladder + 2 at max ✓sweepGLM-5.3·low · GLM-5.3·max · GLM-5.3-Flash·max
AA-Briefcase
artificialanalysis
AA model payload 27 Sep · agentic knowledge work, Elo; 4 models × full ladder + 3 at max; costPerTask per eval ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low
AA-LCR
artificialanalysis
AA model payload · long-context reasoning, input-dominated cost ✓sweepGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low · GLM-5·think · GLM-4.7·think
Agents on Rails · features
rubyonrails
rubyonrails.org/ai + rails/ai-evals runs · 20 tickets ×3, miniswen; Opus 5 & Fable 5.1 high→max, Opus 5.5 default ✓sweepGLM-5.3-Flash·max
ARC-AGI-1
arcprize
ARC Prize evaluations.json · semi-private, costPerTask per model × effort; scores ≥97.5 % flagged gen5-saturated ✓sweepGLM-5.3-Flash·max · GLM-5.3-Flash·high · GLM-5.3-Flash·low
ARC-AGI-2
arcprize
ARC Prize evaluations.json · semi-private, costPerTask per model × effort; Opus 5.5 & Fable 5.1 full ladders ✓sweepGLM-5.3-Flash·max · GLM-5.3-Flash·high · GLM-5.3-Flash·low
BullshitBench v2
github-bullshitbench
petergpt/bullshit-benchmark c1d4ab6 · 100 nonsense prompts, share of clear push-backs (green), effort from the board's reasoning column; primary leaderboard ✓sweepGLM-5.3·low · GLM-5.3·max · GLM-5.2·high
CursorBench 4.0
cursorbench
cursor.com payload · v4.0 long-horizon suite, Opus 5/5.5, Fable 5.1, Sonnet 5/5.5 × low→max, measured tokens × list incl. cache ✓sweepGLM-5.3·max · GLM-5.3·high · GLM-5.3-Flash·max · GLM-5.3·low · GLM-5.3-Flash·high · GLM-5.3-Flash·low
FrontierCode main
cognition
Cognition leaderboard JSON · v1.1 main (100 tasks), Claude models × low→max incl. Sonnet 5.5, measured USD/rollout ✓sweepGLM-5.3·max · GLM-5.3-Flash·max
FrontierCode-Ext
cognition
Cognition leaderboard JSON · v1.1 Extended (150 tasks), Claude models × low→max incl. Sonnet 5.5, measured USD/rollout ✓sweepGLM-5.3·max · GLM-5.3-Flash·max
harnesseval
github-harnesseval
Claude Code · 6 PRs common to every cell, CLI-metered cost ✓sweepGLM-5.3·low · GLM-5.3·high · GLM-5.3-Flash·low · GLM-5.3-Flash·high
ibragim.dev · Pi
ibragim-dev
Personal coding evals core10 v1.3 · 10 tasks × 3 runs, Harbor 0.21.0, Pi 0.85.1, effort per row, pass@1 and $/task; board of 28 Sep ✓sweepGLM-5.3-Flash·max
Playcode MacBook SVG
playcode
Playcode · one-shot MacBook SVG, high/xhigh/max per model, provider-billed cost (quality ordinal, not scored) ✓sweepGLM-5.3-Flash·low · GLM-5.3-Flash·high · GLM-5.3·low · GLM-5.3·high · GLM-5.3·max
Vals VibeCode
vals.ai
Vals AI · Opus 5 sweep low→max (scores), $ cost at leaderboard setting; board entries — snapshot 27 Sep, re-priced ✓sweepGLM-5.2·max · GLM-5.3·max
vlm-exam
github-vlmexam
roboflow · identical 133-image set, 3 models × low/high, measured tokens ✓sweepGLM-5.3-Flash·low · GLM-5.3-Flash·high
WeirdML
epoch-archive
Epoch archive · 3 models × high/max, cost per run ✓sweepGLM-5.3·max · GLM-5.2·high
deepseek-v41-techreport: deepswe
deepseek
DeepSWE v1.1 (Resolved), mini-SWE harness; DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness mini-SWE; all columns at max effort for report/card E; 1017 card effort not stated; + effort sweep 25..100 of deepseek-v4.1-flash from Fig 9/11 (same runs; effort-100 point = Table 3 cell, kept once), tokens_out per response/trajectory; read 2026-10-19 ✓sweep+xmodelGLM-5.3·max
deepseek-v41-techreport: gpqa
deepseek
GPQA Diamond (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness none(reasoning); all columns at max effort for report/card E; 1017 card effort not stated; + effort sweep 25..100 of deepseek-v4.1-flash from Fig 9/11 (same runs; effort-100 point = Table 3 cell, kept once), tokens_out per response/trajectory; read 2026-10-19 ✓sweep+xmodelGLM-5.3·max
deepseek-v41-techreport: tb21
deepseek
Terminal-Bench 2.1 (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness DeepSeek-Harness-minimal; all columns at max effort for report/card E; 1017 card effort not stated; + effort sweep 25..100 of deepseek-v4.1-flash from Fig 9/11 (same runs; effort-100 point = Table 3 cell, kept once), tokens_out per response/trajectory; read 2026-10-19 ✓sweep+xmodelGLM-5.3·max
FeatherBench
ed-is-ai/featherbench
OpenRouter(Anthropic) · 28-task multi-domain suite (coding/data/realworld/security/tool-use), effort=high (Haiku=solo), mean cost/tokens per task-run, consolidated results/summary.json; only Opus 5.5 rung saturated ✓xgenGLM-5.2·high · GLM-5.3·max
AA /task v3
artificialanalysis
AA-Index v3 · max ✓xmodelGLM-4.7·think
AA /task v4
artificialanalysis
AA launch article · per-task $, Fable 5.1 xhigh/max vs Fable 5/Opus 5 max ✓xmodelGLM-4.7·think
AA AutomationB.
artificialanalysis
AA evaluation page · 657 tasks partial-credit, 3 models at max ✓xmodelGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max · GLM-5.3·low
AA FrontierCode 1.1 (AA-run, weighted score)
artificialanalysis
Artificial Analysis AA FrontierCode 1.1 (AA-run, weighted score); score only from the model-page payload; effort from payload effort.slug; read 2026-10-15 ✓xmodelGLM-5.3·low · GLM-5.3·max · GLM-5.3-Flash·max
AA-AnalystAgent
artificialanalysis
Artificial Analysis AA-AnalystAgent; score only from the model-page payload; effort from payload effort.slug; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
aa-index411
tabbit.ai
—xmodelGLM-5.3·max
aa-index5
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aa5-aa-briefcase
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aa5-code2video-aa
artificialanalysis
—xmodelGLM-5.3·max
aa5-critpt-v2-hard
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aa5-frontiercode-2-0
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aa5-harvey-lab-aa
artificialanalysis
—xmodelGLM-5.3·max
aa5-healthbench
artificialanalysis
—xmodelGLM-5.3·max
aa5-healthbench-professional
artificialanalysis
—xmodelGLM-5.3·max
aa5-hle-diamond
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aa5-mlcr-aa
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aa5-omniscience-non-hallucination
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max
aaapex
artificialanalysis
—xmodelGLM-5·think
aagpqa
artificialanalysis
—xmodelGLM-5.3·max · GLM-5.3-Flash·max · GLM-4.7·think · GLM-5·think · GLM-5.1·think
aaifbench
artificialanalysis
—xmodelGLM-4.7·think · GLM-5·think · GLM-5.1·think
aaime25
artificialanalysis
—xmodelGLM-4.7·think
aalcb
artificialanalysis
—xmodelGLM-4.7·think
aaomni-index
artificialanalysis
—xmodelGLM-5.3·low · GLM-5.3·max · GLM-4.7·think · GLM-5·think · GLM-5.1·think · GLM-5.3-Flash·max
aatau-banking
artificialanalysis
—xmodelGLM-4.7·think · GLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max
aatau2
artificialanalysis
—xmodelGLM-4.7·think · GLM-5·think · GLM-5.1·think
aatb21
artificialanalysis
—xmodelGLM-5.3·max · GLM-5.3-Flash·max · GLM-4.7·think · GLM-5.1·think
aatbhard
artificialanalysis
—xmodelGLM-4.7·think · GLM-5·think · GLM-5.1·think
aibenchy.com private benchmark suite — Haiku 5.5 (xhigh) vs Opus 5.5 (high)
aibenchy
aibenchy.com undisclosed-task private suite, score 1-10 scale (site's own verdict text quotes '9.1 vs 9.4'); page states 'Last updated at: 2026-10-07'; Haiku-5.5-xhigh row is the same run repeated across two /compare/ pages (identical score/cost/time/rank), not two independent measurements; read 2026-10-08 ✓xmodelGLM-5.3-Flash·max · GLM-5.2·high · GLM-5.3-FlashX·max · GLM-5.3·high · GLM-5.3-Flash·high · GLM-5.3-Flash·low · GLM-5.3·low
ALE-Bench
epoch-archive
Epoch archive · 3 models at matched high, measured token splits ✓xmodelGLM-5.3·high · GLM-5.2·high · GLM-5.3-Flash·high
APEX-Harm mean task score under attack (all 6 categories)
apexharm-2610.13623
APEX-Harm 180 attack task instances per model (15 underlying APEX-Agents tasks x 12 attack variants), APEX-Agents harness with MCP tools, 150 steps, no timer, judge/rules per paper; metric mean task score 0-1 (utility under attack); ASR/ER safety rates not extracted; arXiv 2610.13623 Table 4 (Oct 2026); cost = Table 1 total for Gemini 3.8 Flash only; read 2026-10-20 ✓xmodelGLM-5.3-Flash·max
apexharm-asr
apexharm-2610.13623
—xmodelGLM-5.3-Flash·max
ArtifactArena model leaderboard, Design Lab Harness (DLH) Elo
arxiv-2610.06511
Best artifact per (model, harness) in a cross-model Bradley-Terry round robin, MuJoCo robot arena; Figure 2 of arXiv 2610.06511 (read from image); read 2026-10-15 ✓xmodelGLM-5.2·high · GLM-5.3·high
ArtifactArena model leaderboard, Sampling Harness (SH) Elo
arxiv-2610.06511
Best artifact per (model, harness) in a cross-model Bradley-Terry round robin, MuJoCo robot arena; Figure 2 of arXiv 2610.06511 (read from image); read 2026-10-15 ✓xmodelGLM-5.2·high
ArtifactArena model leaderboard, Verifier-Grounded Refinement Harness (VGH) Elo
arxiv-2610.06511
Best artifact per (model, harness) in a cross-model Bradley-Terry round robin, MuJoCo robot arena; Figure 2 of arXiv 2610.06511 (read from image); read 2026-10-15 ✓xmodelGLM-5.2·high · GLM-5.3·high
arxiv-2610.04976-pricedguidance-test-median-bits
arxiv-2610.04976
—xmodelGLM-5.3·max
arxiv-2610.12596-spirebench-acts
arxiv-2610.12596
—xmodelGLM-5.3·low · GLM-5.3·max · GLM-5.3-Flash·max
arxiv-2610.13075-longbench-v2-direct
arxiv-2610.13075
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
arxiv-2610.13075-officeqa-v1-direct
arxiv-2610.13075
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
arxiv-2610.13075-officeqa-v2-direct
arxiv-2610.13075
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
arxiv-2610.22599-direction
arxiv-2610.22599
—xmodelGLM-5.3·max
arxiv-2610.22599-r2
arxiv-2610.22599
—xmodelGLM-5.3·max
arxiv-2610.22599-slope
arxiv-2610.22599
—xmodelGLM-5.3·max
arxiv-2610.22599-top-idea-gain
arxiv-2610.22599
—xmodelGLM-5.3·max
arxiv-2610.25078-homebench-im-f1
arxiv-2610.25078
—xmodelGLM-5.3·high
arxiv-2610.25078-homebench-im-succ
arxiv-2610.25078
—xmodelGLM-5.3·high
arxiv-2610.25078-homebench-vm-f1
arxiv-2610.25078
—xmodelGLM-5.3·high
arxiv-2610.25078-homebench-vm-succ
arxiv-2610.25078
—xmodelGLM-5.3·high
arxiv-2610.25078-traceelephant-when
arxiv-2610.25078
—xmodelGLM-5.3·high
arxiv-2610.25078-traceelephant-who
arxiv-2610.25078
—xmodelGLM-5.3·high
arxiv-2610.39438-t3-cr
arxiv-2610.39438
—xmodelGLM-5.3·max
arxiv-2610.39438-t3-csr
arxiv-2610.39438
—xmodelGLM-5.3·max
arxiv-2610.39438-t3-evidence
arxiv-2610.39438
—xmodelGLM-5.3·max
arxiv-2610.39438-t3-intent
arxiv-2610.39438
—xmodelGLM-5.3·max
arxiv-2610.39438-t3-strategy
arxiv-2610.39438
—xmodelGLM-5.3·max
bsb-v1-green
github-bullshitbench-v1
—xmodelGLM-5.3·low · GLM-5.3·max · GLM-5·think
CausalGame Agent mode (structured tool calling): mean survival rate over 14 experiments
causalgame-2607.04293
CausalGame (30 models, 14 game experiments with causal traps; antenna/deployment-zone/weather SCMs), Table 16 mean survival rate % over the 14 experiments (95% CI over runs), 3 runs, Agent mode (tool calls); the other mode is a separate group; arXiv 2607.04293 (Jul 2026); read 2026-10-20 ✓xmodelGLM-5.2·high
CausalGame Prompting mode (Python code): mean survival rate over 14 experiments
causalgame-2607.04293
CausalGame (30 models, 14 game experiments with causal traps; antenna/deployment-zone/weather SCMs), Table 16 mean survival rate % over the 14 experiments (95% CI over runs), 3 runs, Prompting mode (Python code); the other mode is a separate group; arXiv 2607.04293 (Jul 2026); read 2026-10-20 ✓xmodelGLM-5.2·high
Code4Scene private set (95 cases), overall S_model
arxiv-2609.36777
Code4Scene: 190 Unreal Engine cases (Text-to-Scene + Image-to-Scene), coding agents with 25 shared tools and a metering proxy; score in [0,1], S_model weights T2S and I2S equally; a second attempt only after a failed first; cost exists only in Figure 1(b) (not extracted: no cost); Sep 2026; private set (95 cases), overall S_model; read 2026-10-19 ✓xmodelGLM-5.3-Flash·max
Code4Scene public set (95 cases), overall S_model
arxiv-2609.36777
Code4Scene: 190 Unreal Engine cases (Text-to-Scene + Image-to-Scene), coding agents with 25 shared tools and a metering proxy; score in [0,1], S_model weights T2S and I2S equally; a second attempt only after a failed first; cost exists only in Figure 1(b) (not extracted: no cost); Sep 2026; public set (95 cases), overall S_model; read 2026-10-19 ✓xmodelGLM-5.3-Flash·max
Codex Radar intelligence-efficiency board, Codex CLI harness
codexradar
Codex Radar recurring repository-level coding eval (per-task pass over valid tasks; payload mode equal_latest_3: equal-weight latest three valid samples per task); harness codex; snapshot payload stamped 2026-10-01T23:00:27+08:00 read 2026-10-20; score=passed/valid_tasks (derived; payload gives passed, valid_tasks, iq=150*pass_rate); cost=average_price_usd per run (estimated USD-equivalent from token prices); valid_tasks differ per row (36..336); read 2026-10-20 ✓xmodelGLM-5.3·low · GLM-5.3·high · GLM-5.3·max · GLM-5.3-Flash·low · GLM-5.3-Flash·high · GLM-5.3-Flash·max
Conceptual Reasoning Index · DTBench
conceptualreasoning.ai
Conceptual Reasoning Index DTBench sub-score (CRI-rescaled accuracy, x100), cost = average USD per response (site cost chart); effort from the csv effort column; blank = default; read 2026-10-15 ✓xmodelGLM-5.3·max
Conceptual Reasoning Index · LMCA
conceptualreasoning.ai
Conceptual Reasoning Index LMCA sub-score (Pearson-r based, x100), cost = average USD per response (site cost chart); effort from the csv effort column (max rows carry the base name); blank = default; read 2026-10-15 ✓xmodelGLM-5.3·max
Convex evals (no guidelines, web research)
convex.dev
Convex coding evals, 112 evals in 8 categories, experiment 'no guidelines, web research' · total score in % with its error bar, billed cost per full run (USD, OpenRouter); effort medium from the runner's OpenRouter request (HEAD config, assumed for the runs); read 2026-10-01 ✓xmodelGLM-5.2·high
Convex evals (no guidelines)
convex.dev
Convex coding evals, 112 evals in 8 categories, experiment 'no guidelines' · total score in % with its error bar, billed cost per full run (USD, OpenRouter); effort medium from the runner's OpenRouter request (HEAD config, assumed for the runs); read 2026-10-01 ✓xmodelGLM-5.2·high
Convex evals (with guidelines)
convex.dev
Convex coding evals, 112 evals in 8 categories, experiment 'with guidelines' · total score in % with its error bar, billed cost per full run (USD, OpenRouter); effort medium from the runner's OpenRouter request (HEAD config, assumed for the runs); read 2026-10-01 ✓xmodelGLM-5.2·high
Convex LLM leaderboard v92a60e0b · no guidelines
convex.dev
Convex evals (112), benchmark version 92a60e0b (guidelines compressed 7010→4146 tokens), OpenRouter, reasoning effort medium, no guidelines; read 2026-10-15 ✓xmodelGLM-5.2·high
Convex LLM leaderboard v92a60e0b · no guidelines + web
convex.dev
Convex evals (112), benchmark version 92a60e0b (guidelines compressed 7010→4146 tokens), OpenRouter, reasoning effort medium, no guidelines + web; read 2026-10-15 ✓xmodelGLM-5.2·high
Convex LLM leaderboard v92a60e0b · with guidelines
convex.dev
Convex evals (112), benchmark version 92a60e0b (guidelines compressed 7010→4146 tokens), OpenRouter, reasoning effort medium, with guidelines; read 2026-10-15 ✓xmodelGLM-5.2·high
ctala
github-ctala
Claude Code CLI · reasoning not configured, temp 0.7 ✓xmodelGLM-5.3-Flash·max · GLM-5.2·high
DataLLM Lab: 9 executed Python tasks, Qwen3.8 Max family (Prime, Max, Max-0902), pass and cost per 1,000 tasks
datallmlab
9 executed Python tasks (python3 -I subprocess, 12 s timeout), temperature 0, max_tokens 4000, one scored attempt, via OpenRouter; three runs on different days (Prime 2026-10-02, Max 2026-08-20, Max-0902 2026-09-16); cost = measured tokens x list price on the priced-on date (source-computed, 'not a billing statement'); read 2026-10-20 ✓xmodelGLM-5.1·think · GLM-5·think · GLM-5.3·max · GLM-5.3-FlashX·max
datallmlab-coding9
datallmlab
—xmodelGLM-5.2·high
decisionpivot-core-pair
arxiv-2610.28522
—xmodelGLM-5.3·low
decisionpivot-subset1296
arxiv-2610.28522
—xmodelGLM-5.3·low · GLM-5.3·high
DeepSeek V4 preview card: apex
deepseek
apex: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: apexshort
deepseek
apexshort: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: browsecomp
deepseek
browsecomp: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: csimpleqa
deepseek
csimpleqa: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: gdpval
deepseek
gdpval: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: gpqa
deepseek
gpqa: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: hle
deepseek
hle: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: hle-tools
deepseek
hle-tools: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: hmmt26
deepseek
hmmt26: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: imo
deepseek
imo: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: mcpatlas
deepseek
mcpatlas: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: mmlupro
deepseek
mmlupro: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: simpleqav
deepseek
simpleqav: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: swemulti
deepseek
swemulti: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: swepro
deepseek
swepro: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: tb20
deepseek
tb20: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
DeepSeek V4 preview card: toolathlon
deepseek
toolathlon: DeepSeek-V4 preview model card (HF README, 2026-04-22/24); n, attempts, scaffold, judge not stated in the card (no footnote on sampling); DeepSeek rows by reasoning mode (Non-think/Think High/Think Max), other makers per column header; Table 1 (frontier) and Table 2 (modes) merged, the V4-Pro Max cell is identical in both; read 2026-10-19 ✓xmodelGLM-5.1·think
deepseek-v41-techreport: ale
deepseek
Agents' Last Exam (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness official-scaffold(ALE-CLI); all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: autobench
deepseek
AutomationBench v1.0.6 public set (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness official-scaffold; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: cybergym
deepseek
CyberGym (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness unstated; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: exploitgym
deepseek
ExploitGym (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness unstated; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: hle-tools
deepseek
HLE with tools (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness unstated; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: nl2repo
deepseek
NL2Repo-Bench (Score); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness DeepSeek-Harness-minimal; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: programbench
deepseek
ProgramBench (Almost@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness DeepSeek-Harness-minimal; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: tb30
deepseek
Terminal-Bench 3.0 (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness DeepSeek-Harness-minimal; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
deepseek-v41-techreport: tb40
deepseek
Terminal-Bench 4.0 (Pass@1); DeepSeek V4.1-Flash tech report / model card; sampling stated by the card: temperature=1.0, top_p=0.95 (report 5.3.1: temperature and top-p 1.0 for reasoning benchmarks); N=8 samples per task for DeepSWE v1.1, N=3 for Terminal-Bench 2.1 (Table 4 note); judge not stated; harness DeepSeek-Harness-minimal; all columns at max effort for report/card E; 1017 card effort not stated; read 2026-10-19 ✓xmodelGLM-5.3·max
Design Creativity Bench (UI generation; overall T-score, median cost per design)
arxiv-2610.11539
80 UI design tasks / 160 prompts, one generation per design, highest effort (Opus 5.5 at xhigh); Tables 2 and 9 of arXiv 2610.11539; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
dubesor-benchtable
dubesor
—xmodelGLM-5·think · GLM-4.7·think
E-Commerce Bench (arXiv 2608.30730)
arxiv-2608.30730
E-Commerce Bench, simulated e-commerce store, 365-day task, 18 tools, 4000-turn cap, 5 independent episodes per model in one fixed world, harness = the paper's agent loop; score = mean end-of-year total assets (kilo-yuan) from a 100k stake, bankrupt episodes kept in the mean; no cost reported; arXiv v1 2026-08-31; read 2026-10-20 ✓xmodelGLM-5.2·high · GLM-5.1·think
econ-eval leaderboard A (20 tasks, Opus-5 judge)
github-deluair-econ-eval
deluair/econ-eval README at commit 77f9f3e (16 Oct 2026), leaderboard of 2026-09-10: 20 economics tasks (quantitative, reasoning, coding, writing) x 5 samples = 100 completions/model, numeric/exact/code-exec graders + rubric tasks judged by Claude Opus 5; score = mean over tasks of task means; cost = total for the 100 completions from token counts x list price (prices read from the OpenRouter catalog 2026-07-31 / vendor pages); per-track tables are the same completions (not extra groups); ten models ran 2026-08-01, deepseek-flash added 2026-09-10 same tasks/samples/judge; read 2026-10-20 ✓xmodelGLM-5.2·max
EMB
vals.ai
Vals AI · bash+execute, 10 Claude models at max (payload compute_effort) — snapshot 27 Sep, re-priced ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Epoch · chess puzzles
epoch.ai
Epoch AI's own evaluation (chess_puzzles), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.2·high · GLM-5.3·max · GLM-5.3-Flash·max
Epoch · frontiermath tier 4 v2
epoch.ai
Epoch AI's own evaluation (frontiermath_tier_4_v2), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.3-Flash·max · GLM-5.3·max
Epoch · frontiermath tiers 1 3 v2
epoch.ai
Epoch AI's own evaluation (frontiermath_tiers_1_3_v2), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.2·high · GLM-5.3-Flash·max · GLM-5.3·max
Epoch · gpqa diamond
epoch.ai
Epoch AI's own evaluation (gpqa_diamond), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.3-Flash·max · GLM-5.3·max · GLM-5.2·high
Epoch · mystery game puzzles
epoch.ai
Epoch AI's own evaluation (mystery_game_puzzles), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max · GLM-5.2·high
Epoch · otis mock aime 2024 2025
epoch.ai
Epoch AI's own evaluation (otis_mock_aime_2024_2025), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.3-Flash·max · GLM-5.3·max · GLM-5.2·high
Epoch · simpleqa verified
epoch.ai
Epoch AI's own evaluation (simpleqa_verified), benchmarking hub archive read 30 Sep 2026; effort from the model-version suffix, mean score ✓xmodelGLM-5.3·max
epoch-ext-surface-evolver-bench
epoch.ai
—xmodelGLM-5.2·high · GLM-5.3-Flash·max
fcode20score
cognition
—xmodelGLM-5.3·low · GLM-5.3·high · GLM-5.3·max · GLM-5.3-Flash·low · GLM-5.3-Flash·high · GLM-5.3-Flash·max
Financial Touchstone v1.2 accuracy (annual-report QA, shared top-5 retrieval)
arxiv-2608.08634
arXiv 2608.08634 (9 Aug 2026) Financial Touchstone v1.2: 2,967 question-context-answer items on 495 annual reports, same top-5 retrieved context for every model, output cap 5,000 tokens, LLM judge Claude Opus 4.6 (16 models) / Claude Opus 4.8 (Kimi K2.6, GLM-4.7, Qwen3-Max, ERNIE 5.1); score = accuracy excluding retriever errors (Fig. 1); hallucination rate = same runs, other metric (not a group); read 2026-10-20 ✓xmodelGLM-5·think · GLM-4.7·think
FinancialAuditBench (2609.32835) — 45 audit workpaper tasks, bash-only agent, 8 runs/task
arxiv-2609.32835
Table 8/Table 9 main row, strict pass % and mean cost $/trajectory, 360 trajectories/model; verified against saved evidence in /work/.salvo14/s18/a1/src/, 2026-10-01; read 2026-10-01 ✓xmodelGLM-5.2·high
FinIndices Single-Index with formula hint, dev
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Single-Index queries; with formula hint; split dev; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Single-Index with formula hint, test
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Single-Index queries; with formula hint; split test; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Single-Index with formula hint, train
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Single-Index queries; with formula hint; split train; end-to-end accuracy %; single run on the training split; read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Single-Index without hint, dev
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Single-Index queries; without hint; split dev; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Single-Index without hint, test
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Single-Index queries; without hint; split test; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Single-Index without hint, train
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Single-Index queries; without hint; split train; end-to-end accuracy %; single run on the training split; read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Table-Index with formula hint, dev
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Table-Index queries; with formula hint; split dev; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Table-Index with formula hint, test
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Table-Index queries; with formula hint; split test; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Table-Index with formula hint, train
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Table-Index queries; with formula hint; split train; end-to-end accuracy %; single run on the training split; read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Table-Index without hint, dev
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Table-Index queries; without hint; split dev; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Table-Index without hint, test
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Table-Index queries; without hint; split test; end-to-end accuracy %; mean of 3 runs ± sd (Dev/Test); read 2026-10-19 ✓xmodelGLM-5·think
FinIndices Table-Index without hint, train
arxiv-2607.28661
FinIndices (Chinese listed-company financial statements, long context ~16k tokens); Table-Index queries; without hint; split train; end-to-end accuracy %; single run on the training split; read 2026-10-19 ✓xmodelGLM-5·think
FlavourBench Score (mean of 3 families x 178 tasks)
flavourbench-2608.20574
FlavourBench Table 3 leaderboard, 534 tasks, FB Score = equal-family mean, 2 collection panels pooled, OpenRouter fixed provider (GLM 5.3 direct); arXiv 2608.20574 (20 Aug 2026); family columns Table 10 are same runs (rule 8); read 2026-10-20 ✓xmodelGLM-5.3·low
FrontierSWE dominance score (GLM-5.3 card, as of 2026/08/14)
zhipu
FrontierSWE, evaluated by Proximal, 1M context, 128K output, dominance score as of 2026/08/14; held group frontierswe is mean-reward@5 (glm-5.3 30.18 there vs 78.1 here: different metric); read 2026-10-19 ✓xmodelGLM-5.3·max
FrontierSWE v2
frontierswe
Proximal frontierswe.com + Sonnet 5.5 card · 34 ultra-long tasks, max effort, proximus harness, mean reward@5; cost basis undocumented ✓xmodelGLM-5.3·max
gdpx-overall
semianalysis
—xmodelGLM-5.3·max
hehee9 Maze Bench Tier 1 (30 mazes)
hehee9maze-bcd609c
Maze image (2048x2048) -> single-line move string, per-maze score = 100 x progress x efficiency, 30 mazes (6 sizes x 5, 4x4..18x18), one run per model-maze, single call, image input, no tools, medium effort / 8K budget when supported; metric official mean score 0-100; cost = dashboard calculateCost = (input tokens x input rate + output tokens x output rate)/1e6 from rates stored per model in the JSON (before cache discounts); live board snapshot of 20 Oct 2026 (github.com/hehee9/maze-bench commit bcd609c, status completed); read 2026-10-20 ✓xmodelGLM-5.3-Flash·high
Hex DataBench
hex.tech
Hex DataBench, 100 analytics tasks in the Hex workspace, rubric-judged · score in %, average cost per task (USD), tokens per task (one total); effort from the row label; Opus 5, Fable 5.1, Fable 5, Opus 4.8, Sonnet 5 at every listed effort (Opus 5.5 in its own group); read 2026-10-01 ✓xmodelGLM-5.3·max · GLM-5.3·high · GLM-5.2·high · GLM-5.3·low
hf-kimi-k3.1-pro-aa-briefcase
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-ale-all
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-ale-cli
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-apex-agents
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-automationbench-aa
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-biomystery
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-business-operation
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-deck-arena
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-deepswe
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-design-bench
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-exploitbench
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-exploitgym
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-frontier-finance
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-gdppdf
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-gdpval-aa
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-gpqa-diamond
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-hallu-hard-tools
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-hallu-notools
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-handbook
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-hle-diamond
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-hle-full
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-hle-tools
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-kcb2
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-kcb3
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-officeqa-pro
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-officeqa-pro-v2
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-posttrainbench
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-programbench
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-swemarathon
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-tb4
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-tbs
moonshot
—xmodelGLM-5.3·max
hf-kimi-k3.1-pro-xdev-hard
moonshot
—xmodelGLM-5.3·max
hlb-deep-swe-claude-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-codex
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-hermes
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-kimi-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-mini-swe-agent
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-openclaw
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-opencode
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-pi
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-deep-swe-qwen-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-claude-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-codex
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-hermes
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-kimi-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-mini-swe-agent
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-openclaw
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-opencode
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-pi
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-featurebench-qwen-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-claude-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-codex
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-hermes
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-kimi-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-mini-swe-agent
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-openclaw
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-opencode
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-pi
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
hlb-terminal-bench-2.1-qwen-code
joelniklaus-harness
—xmodelGLM-5.3-Flash·high · GLM-5.3-Flash·low
ibmartian-full-p-other
ironbee
—xmodelGLM-5.3·high · GLM-5.3·low
IdentityBench v4.1 v41_lite
github-identitybench
IdentityBench v4.1 v41_lite: 60 enterprise IGA scenarios x 6 governance categories = 360 category tasks, 1,628 required actions, 12-turn cap (T0 single monolith), seed 20261019, 1 run per model, README of 2026-10-20 (commit 56591d0), published 2026-10-19; metric tasks solved of 360; cost = run cost of the whole suite (billed or estimated); read 2026-10-20 ✓xmodelGLM-5.3-Flash·max
ITBench-AA (SRE)
artificialanalysis
Artificial Analysis ITBench-AA (SRE); score only from the model-page payload; effort from payload effort.slug; read 2026-10-15 ✓xmodelGLM-5.1·think · GLM-5.3·max · GLM-5.3-Flash·max
lb2-browser-analysis-driven-action
livebench
—xmodelGLM-5.3-Flash·max
lb2-browser-constraint-adherence
livebench
—xmodelGLM-5.3-Flash·max
lb2-browser-data-aggregation
livebench
—xmodelGLM-5.3-Flash·max
lb2-browser-sequential-planning
livebench
—xmodelGLM-5.3-Flash·max
lb2-browser-structural-integrity
livebench
—xmodelGLM-5.3-Flash·max
lb2-coding-agent-loop
livebench
—xmodelGLM-5.3-Flash·max
lb2-coding-cloud-infrastructure
livebench
—xmodelGLM-5.3-Flash·max
lb2-coding-hard-fixes
livebench
—xmodelGLM-5.3-Flash·max
lb2-coding-llm-api-integration
livebench
—xmodelGLM-5.3-Flash·max
lb2-coding-memory-tools
livebench
—xmodelGLM-5.3-Flash·max
lb2-coding-product-ux
livebench
—xmodelGLM-5.3-Flash·max
lb2-dataanalysis-consecutive-events
livebench
—xmodelGLM-5.3-Flash·max
lb2-dataanalysis-table-join
livebench
—xmodelGLM-5.3-Flash·max
lb2-dataanalysis-table-reformat
livebench
—xmodelGLM-5.3-Flash·max
lb2-design-documents-docx
livebench
—xmodelGLM-5.3-Flash·max
lb2-design-pdf
livebench
—xmodelGLM-5.3-Flash·max
lb2-design-presentations-ppt
livebench
—xmodelGLM-5.3-Flash·max
lb2-longhorizon-delegated-operations
livebench
—xmodelGLM-5.3-Flash·max
lb2-longhorizon-finance-compliance-apps
livebench
—xmodelGLM-5.3-Flash·max
lb2-longhorizon-long-run-operations
livebench
—xmodelGLM-5.3-Flash·max
lb2-longhorizon-operations-commerce-apps
livebench
—xmodelGLM-5.3-Flash·max
lb2-longhorizon-performance-optimisation
livebench
—xmodelGLM-5.3-Flash·max
lb2-longhorizon-systems-engineering
livebench
—xmodelGLM-5.3-Flash·max
lb2-reasoning-logic-with-navigation
livebench
—xmodelGLM-5.3-Flash·max
lb2-reasoning-spatial
livebench
—xmodelGLM-5.3-Flash·max
lb2-reasoning-theory-of-mind
livebench
—xmodelGLM-5.3-Flash·max
lb2-reasoning-zebra-puzzles
livebench
—xmodelGLM-5.3-Flash·max
lb2-research-legal
livebench
—xmodelGLM-5.3-Flash·max
lb2-research-regulatory-compliance
livebench
—xmodelGLM-5.3-Flash·max
lb2-research-science-evidence
livebench
—xmodelGLM-5.3-Flash·max
lb2-research-strategy
livebench
—xmodelGLM-5.3-Flash·max
lb2-research-tax-finance
livebench
—xmodelGLM-5.3-Flash·max
lechmazur Adversarial LLM Debate Benchmark — Bradley-Terry rating
lechmazur-debate
lechmazur/debate, side-swapped adversarial debates, multi-model judge panel (incl. a Claude judge), Bradley-Terry rating with match counts; read 2026-10-05 ✓xmodelGLM-5.3·high
lechmazur Creative Writing Benchmark — comparison score
lechmazur-writing
lechmazur/writing, matched creative-brief stories judged pairwise by rotating LLM evaluators, Thurstone/BT-style comparison score centered at 0, bootstrap 95% CI; read 2026-10-05 ✓xmodelGLM-5.3·max
lechmazur NYT Connections Extended Benchmark — full 940-puzzle board
lechmazur-nyt
lechmazur/nyt-connections, 940 puzzles, quadratic-v1 scoring (g/4)^2 per puzzle, objective (not model-judged); read 2026-10-05 ✓xmodelGLM-5.2·high · GLM-5.3·high
lechmazur NYT Connections Extended Benchmark — harder 100-puzzle subset
lechmazur-nyt
lechmazur/nyt-connections, harder 100-puzzle subset, same quadratic-v1 scoring, objective; read 2026-10-05 ✓xmodelGLM-5.2·high · GLM-5.3·high
LiveBench 2.0 · Browser Use suite
livebench
LiveBench 2.0 Browser Use suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3-Flash·max
LiveBench 2.0 · Coding suite
livebench
LiveBench 2.0 Coding suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
LiveBench 2.0 · Data Analysis suite
livebench
LiveBench 2.0 Data Analysis suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
LiveBench 2.0 · Design suite
livebench
LiveBench 2.0 Design suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
LiveBench 2.0 · Long-Horizon suite
livebench
LiveBench 2.0 Long-Horizon suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
LiveBench 2.0 · Reasoning suite
livebench
LiveBench 2.0 Reasoning suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
LiveBench 2.0 · Research suite
livebench
LiveBench 2.0 Research suite, single run, every model at its highest effort (max), cost = suite total at list price; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
llm_chess — aggregated Elo vs Random + Dragon engine (all Claude models)
llm_chess
maxim-saplin/llm_chess, LLM plays Black vs RANDOM_PLAYER and Dragon engine, Elo anchored to Dragon ladder, 95% CI, average_game_cost as cost per game; read 2026-10-05 ✓xmodelGLM-5.3-Flash·high
llm2014-logic-2026-10-median
llm2014-logic
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
LMArena Agent
lmarena.ai
Agent Mode real sessions (Code/Chat/Work) · cost=median $/task (14d window), score=Net Improvement composite signal, effort read from model-variant label; live board, read 27 Sep 2026 ✓xmodelGLM-5.3·max
LMArena Text
lmarena.ai
Text arena (Overall) · blind pairwise votes, Elo rating, no cost; effort read from the arena's model key (e.g. claude-opus-5.5-high-vertex); live board, read 30 Sep 2026 ✓xmodelGLM-5.3·max
LMArena WebDev
lmarena.ai
Code arena, WebDev board (Overall) · blind pairwise votes, Elo rating, no cost; effort read from the arena's model key (e.g. claude-sonnet-5-5-high); live board, read 30 Sep 2026 ✓xmodelGLM-5.3·max
LMArena WebDev — HTML sub-category
lmarena.ai
lmarena-ai/leaderboard-dataset, configs/webdev, category=webdev-html, latest split, publish 2026-10-02; read 2026-10-05 ✓xmodelGLM-5.3·max
LMArena WebDev — React sub-category
lmarena.ai
lmarena-ai/leaderboard-dataset, configs/webdev, category=webdev-react, latest split, publish 2026-10-02; read 2026-10-05 ✓xmodelGLM-5.3·max
M-GATE grammar error detection, MCC mean over 30 languages (temperature 0)
mgate-2608.03803
M-GATE (RWS TrainAI) July 2026 reporting snapshot, Appendix J J.1; 53 models / 82 configurations March-July 2026; linguist-crafted adversarial grammar items, 30 languages, temperature 0, one run; list prices only (no run spend: no cost); arXiv 2608.03803 (4 Aug 2026); default-temperature 3-run and per-language columns are other summaries of the same task (rule 8); read 2026-10-20 ✓xmodelGLM-5.2·high
M-GATE round-trip translation, hard subset, normalised score mean over 29 targets
mgate-2608.03803
M-GATE (RWS TrainAI) July 2026 reporting snapshot, Appendix J J.2; 53 models / 82 configurations March-July 2026; 100 English sources round-tripped into 29 languages, hard subset of 50 sentences, judges GPT-5.4 Mini + Claude Haiku 4.5 + Gemini 3.1 Flash Lite at temp 0; list prices only (no run spend: no cost); arXiv 2608.03803 (4 Aug 2026); default-temperature 3-run and per-language columns are other summaries of the same task (rule 8); read 2026-10-20 ✓xmodelGLM-5.2·high
MathArena · aime 2025
matharena.ai
MathArena, aime--aime_2025: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think
MathArena · aime 2026
matharena.ai
MathArena, aime--aime_2026: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think · GLM-5.2·max
MathArena · apex 2025
matharena.ai
MathArena, apex--apex_2025: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max · GLM-5·think
MathArena · april
matharena.ai
MathArena, arxiv--april: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · april
matharena.ai
MathArena, arxiv_false--april: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · brumo 2025
matharena.ai
MathArena, brumo--brumo_2025: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think
MathArena · february
matharena.ai
MathArena, arxiv--february: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max · GLM-5·think
MathArena · february
matharena.ai
MathArena, arxiv_false--february: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think · GLM-5.2·max
MathArena · hmmt feb 2025
matharena.ai
MathArena, hmmt--hmmt_feb_2025: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think
MathArena · hmmt feb 2026
matharena.ai
MathArena, hmmt--hmmt_feb_2026: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max · GLM-5·think
MathArena · january
matharena.ai
MathArena, arxiv--january: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think
MathArena · june
matharena.ai
MathArena, arxiv--june: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · june
matharena.ai
MathArena, arxiv_false--june: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · june
matharena.ai
MathArena, arxivlean--june: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · march
matharena.ai
MathArena, arxiv--march: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max · GLM-5·think
MathArena · march
matharena.ai
MathArena, arxiv_false--march: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think · GLM-5.2·max
MathArena · may
matharena.ai
MathArena, arxiv--may: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · may
matharena.ai
MathArena, arxiv_false--may: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5.2·max
MathArena · shortlist 2025
matharena.ai
MathArena, apex--shortlist_2025: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think · GLM-5.2·max
MathArena · smt 2025
matharena.ai
MathArena, smt--smt_2025: each model run 4 times on each problem · accuracy in % with its 95 % CI in the flags, cost = the average cost of one run on one problem (USD), mean input/output tokens; effort from the model name; ⚠️ (model released after the competition) kept as a flag; read 2026-09-30 ✓xmodelGLM-5·think
MazeBench ASCII with Python tools
mazebench
MazeBench AI leaderboard (mazebench.com/leaderboard/ai?mode=ascii), ASCII observations, Python tools enabled; score = gems collected /100 (the board's default bar), 1 run per (model, harness), snapshot of updated_at 2026-09-14 / generated 2026-10-08; cost = recorded API cost per run (basis per row); read 2026-10-20 ✓xmodelGLM-5.3·high
MCP-Atlas public (GLM-5 card)
zhipu
MCP-Atlas 500-task public set think mode, 10 min timeout, judge Gemini 3 Pro; GLM-5 card; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
MCP-Atlas public set (GLM-5.2 card)
zhipu
MCP-Atlas 500-task public set, think mode, 10 min timeout, judge Gemini-3.0-Pro; table value 76.8 (chart shows 77.0); read 2026-10-19 ✓xmodelGLM-5.1·think
Mercor · FrontierSWE v2
mercor.com
Mercor, FrontierSWE v2: the open-source benchmark independently evaluated by Mercor · score in %, ± kept in the flags; the card shows its top 5 models only; effort from the board's per-model field; no cost; read 2026-09-30 ✓xmodelGLM-5.3·max
Mercor · Long Context Reasoning (AA-LCR) Extended
mercor.com
Mercor, Long Context Reasoning (AA-LCR) Extended: the public benchmark extended with new expert-built Mercor tasks (89 Mercor tasks) · score in %, ± kept in the flags; the card shows its top 5 models only; effort from the board's per-model field; no cost; read 2026-09-30 ✓xmodelGLM-5.3·max
Mercor APEX-Accounting
mercor.com
Mercor's own board, APEX-Accounting, harness loop_agent · pass@1 in %, Mercor's ± half-width kept in the flags, mean score beside it; effort from the board's per-model field; no cost; read 2026-09-30 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Mercor APEX-Agents
mercor.com
Mercor's own board, APEX-Agents, harness loop_truncated_tools_agent · pass@1 in %, Mercor's ± half-width kept in the flags, mean score beside it; effort from the board's per-model field; no cost; read 2026-09-30 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Mercor APEX-v1 (The AI Productivity Index)
mercor.com
Mercor's own board, APEX-v1 (The AI Productivity Index), harness singleshot_agent · pass@1 in %, Mercor's ± half-width kept in the flags, mean score beside it; effort from the board's per-model field; no cost; read 2026-09-30 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
mercor-apexacct-mean
mercor.com
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
mercor-apexv1-mean
mercor.com
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
Mine Odyssey: Minecraft spatial navigation (success rate, 180 tasks)
arxiv-2610.11328
180 navigation tasks (125 outdoor/55 indoor), 500 decisions or 6 h; Opus 5 at 4 efforts plus Opus 5.5 max; Table 1 of arXiv 2610.11328; read 2026-10-15 ✓xmodelGLM-5.3-Flash·max
MineBench voxel arena
minebench
minebench.ai live API · voxel builds, 15 prompts, Elo arena read 29 Sep, measured cost for 8 of 12 Claude models ✓xmodelGLM-5.3-Flash·max · GLM-5.3·max
NonoBench
nonobench.com
NonoBench, 30 nonogram puzzles · puzzles solved per effort level (the board's effort-ladder chart); total cost of the 30 puzzles for each model's best configuration (leaderboard rows); Sonnet 4.5 reasoning off/on as nothink/default; read 2026-09-30 ✓xmodelGLM-5.3-Flash·low · GLM-5.3-Flash·high · GLM-5.3-Flash·max · GLM-5.3·low · GLM-5.3·high · GLM-5.3·max · GLM-4.7·think
NYT Connections, one submission per puzzle, 20 puzzles (matsonj/eval-connections)
matsonj/eval-connections
results.md (data behind matsonj.github.io/eval-connections); pts out of max 100 (20 puzzles x up to 5 pts, wins/traps/groups also recorded); cost is total USD for the 20-game run; models ran on different dates (see confound); single run each; read 2026-10-01 ✓xmodelGLM-5.3-Flash·low · GLM-5.3·low
ockbench-top200
ockbench
—xmodelGLM-5.2·high
OpenRouter VGI-Bench (video general intelligence, 439 multiple-choice questions on long videos): accuracy and cost per question
openrouter-vgi-bench
VGI-Bench public split 439 questions (by Seldon), run through OpenRouter default routing, one turn, full video + question, temperature 0, graded letter after 'Answer:'; listed only with >=395 answered; scores aggregate successful runs of the newest 90 days; board last run 2026-10-18 12:18 UTC; cost = average cost per question as printed; read 2026-10-20 ✓xmodelGLM-5.3-Flash·max
reactbench
reactbench
—xmodelGLM-5.3-Flash·max · GLM-5.3·max · GLM-5.2·high
reactbench-fix
reactbench
—xmodelGLM-5.3·max · GLM-5.3-Flash·max · GLM-5.2·high
reactbench-write
reactbench
—xmodelGLM-5.3-Flash·max · GLM-5.3·max · GLM-5.2·high
roblox-leetcode-easy
hf-roblox-llm-leaderboard
—xmodelGLM-5.2·high · GLM-5.3·high · GLM-5.3-Flash·high
roblox-leetcode-hard
hf-roblox-llm-leaderboard
—xmodelGLM-5.2·high · GLM-5.3·high · GLM-5.3-Flash·high
roblox-leetcode-medium
hf-roblox-llm-leaderboard
—xmodelGLM-5.2·high · GLM-5.3·high · GLM-5.3-Flash·high
roblox-qa-freeform
hf-roblox-llm-leaderboard
—xmodelGLM-5.2·high · GLM-5.3·high · GLM-5.3-Flash·high
roblox-qa-multiple-choice
hf-roblox-llm-leaderboard
—xmodelGLM-5.2·high · GLM-5.3·high · GLM-5.3-Flash·high
Rubric transfer: expert score, ordinary rubric
arxiv-2610.10809
Rubric-transfer study on HealthBench-style typical split (n=105, 15 per theme over 7 themes): responses from a panel of response models are scored by judge Gemini 3.8 Flash (reasoning effort medium, majority of 3) against the expert rubric ('ordinary') or a foreign rubric; expert score 0..1 with 95% paired bootstrap CI; sampling temperature = each provider's default; ordinary expert rubric; read 2026-10-19 ✓xmodelGLM-5.3-Flash·low
rubric-expert-share-lost
arxiv-2610.10809
—xmodelGLM-5.3-Flash·low
SenseBench (WSD) · prompt p001
sense-bench.com
SenseBench lexEN v1, word-sense disambiguation, prompt set p001 · accuracy in % (± in the flags), cost in USD per million items as the board prices it; effort from the model name; read 2026-09-30 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·low · GLM-5.3·low
SpireBench: Slay the Spire Ascension 20 Ironclad, mean final floor reached
arxiv-2610.12596
10 runs per model on the same held-out seeds, themed reskin with note-taking; Table 1 and Table 26 of arXiv 2610.12596; read 2026-10-15 ✓xmodelGLM-5.3·low · GLM-5.3·max · GLM-5.3-Flash·low · GLM-5.3-Flash·max
SuperCLUE 2025年11月 代码生成
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年11月, column 代码生成, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-4.7·think
SuperCLUE 2025年11月 幻觉控制
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年11月, column 幻觉控制, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-4.7·think
SuperCLUE 2025年11月 数学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年11月, column 数学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-4.7·think
SuperCLUE 2025年11月 科学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年11月, column 科学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-4.7·think
SuperCLUE 2025年11月 精确指令遵循
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年11月, column 精确指令遵循, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-4.7·think
SuperCLUE 2025年度测评 代码生成
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年度测评, column 代码生成, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
SuperCLUE 2025年度测评 幻觉控制
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年度测评, column 幻觉控制, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
SuperCLUE 2025年度测评 数学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年度测评, column 数学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
SuperCLUE 2025年度测评 智能体(任务规划)
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年度测评, column 智能体(任务规划), closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
SuperCLUE 2025年度测评 科学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年度测评, column 科学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
SuperCLUE 2025年度测评 精确指令遵循
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2025年度测评, column 精确指令遵循, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think · GLM-4.7·think
SuperCLUE 2026年3月 代码生成
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年3月, column 代码生成, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think
SuperCLUE 2026年3月 幻觉控制
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年3月, column 幻觉控制, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think
SuperCLUE 2026年3月 数学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年3月, column 数学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think
SuperCLUE 2026年3月 智能体(任务规划)
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年3月, column 智能体(任务规划), closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think
SuperCLUE 2026年3月 科学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年3月, column 科学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think
SuperCLUE 2026年3月 精确指令遵循
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年3月, column 精确指令遵循, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5·think
SuperCLUE 2026年5月 代码生成
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年5月, column 代码生成, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5.1·think
SuperCLUE 2026年5月 幻觉控制
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年5月, column 幻觉控制, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5.1·think
SuperCLUE 2026年5月 数学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年5月, column 数学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5.1·think
SuperCLUE 2026年5月 智能体(任务规划)
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年5月, column 智能体(任务规划), closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5.1·think
SuperCLUE 2026年5月 科学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年5月, column 科学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5.1·think
SuperCLUE 2026年5月 精确指令遵循
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年5月, column 精确指令遵循, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; read 2026-10-19 ✓xmodelGLM-5.1·think
SuperCLUE 2026年7月 幻觉控制
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年7月, column 幻觉控制, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; cost = SuperCLUE 2026年7月_成本 sheet (same 18 models/effort strings/builds as this board), CNY, tokens x list unit price; read 2026-10-19 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
SuperCLUE 2026年7月 数学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年7月, column 数学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; cost = SuperCLUE 2026年7月_成本 sheet (same 18 models/effort strings/builds as this board), CNY, tokens x list unit price; read 2026-10-19 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
SuperCLUE 2026年7月 智能体任务规划
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年7月, column 智能体任务规划, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; cost = SuperCLUE 2026年7月_成本 sheet (same 18 models/effort strings/builds as this board), CNY, tokens x list unit price; read 2026-10-19 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
SuperCLUE 2026年7月 智能体编程
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年7月, column 智能体编程, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; cost = SuperCLUE 2026年7月_成本 sheet (same 18 models/effort strings/builds as this board), CNY, tokens x list unit price; read 2026-10-19 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
SuperCLUE 2026年7月 科学推理
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年7月, column 科学推理, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; cost = SuperCLUE 2026年7月_成本 sheet (same 18 models/effort strings/builds as this board), CNY, tokens x list unit price; read 2026-10-19 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
SuperCLUE 2026年7月 精确指令遵循
superclue-general
SuperCLUE general leaderboard (CLUE), edition 2026年7月, column 精确指令遵循, closed-book private questions via API, n/attempts/judge not stated on the sheet; sheet 总排行榜 only; cost = SuperCLUE 2026年7月_成本 sheet (same 18 models/effort strings/builds as this board), CNY, tokens x list unit price; read 2026-10-19 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Surface Evolver bench
yhenon-surface-evolver
yhenon/surface-evolver-llm-eval aggregates.json · 16 runs/model, OpenRouter recorded cost, Opus 4.8 none/high, Sonnet 5 medium, Fable 5 high ✓xmodelGLM-5.2·high · GLM-5.3-Flash·max
surfevolver-pass
yhenon-surface-evolver
—xmodelGLM-5.2·high · GLM-5.3-Flash·max
Surge AI · Antidote: Everyday Edition
surgehq.ai
Surge AI Antidote: Everyday Edition leaderboard (surgehq.ai/benchmarks/antidote); Elo from human preference comparisons (95% CI in flags); effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max
Surge AI · Chartography
surgehq.ai
Surge AI Chartography leaderboard (surgehq.ai/benchmarks/chartography); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3-Flash·max
Surge AI · ComplexConstraints
surgehq.ai
Surge AI ComplexConstraints leaderboard (surgehq.ai/benchmarks/complex-constraints); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Surge AI · DAYJOB: Finance
surgehq.ai
Surge AI DAYJOB: Finance leaderboard (surgehq.ai/benchmarks/dayjob-finance); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Surge AI · DAYJOB: Healthcare
surgehq.ai
Surge AI DAYJOB: Healthcare leaderboard (surgehq.ai/benchmarks/dayjob-healthcare); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Surge AI · EnterpriseBench: CoreCraft
surgehq.ai
Surge AI EnterpriseBench: CoreCraft leaderboard (surgehq.ai/benchmarks/enterprisebench-corecraft); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Surge AI · GDP.pdf
surgehq.ai
Surge AI GDP.pdf leaderboard (surgehq.ai/benchmarks/gdp-pdf); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3-Flash·max
Surge AI · GDP.xlsx
surgehq.ai
Surge AI GDP.xlsx leaderboard (surgehq.ai/benchmarks/gdp-xlsx); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max
Surge AI · HANDBOOK.md Agents
surgehq.ai
Surge AI HANDBOOK.md Agents leaderboard (surgehq.ai/benchmarks/handbook); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3-Flash·max · GLM-5.3·max
Surge AI · Hemingway-bench
surgehq.ai
Surge AI Hemingway-bench leaderboard (surgehq.ai/benchmarks/hemingway-bench); Elo from human preference comparisons (95% CI in flags); effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Surge AI · Riemann-bench
surgehq.ai
Surge AI Riemann-bench leaderboard (surgehq.ai/benchmarks/riemann-bench); score in %; effort from the model label; vendor-run; no cost; read 2026-10-15 ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
swe-rebench 07/26
swe-rebench
ReAct minimal · Opus5/Fable5/Sonnet5 all at high ✓xmodelGLM-5.2·high
TasteVal compute multiplier vs expert baseline (Researcher models, 8 ML-research tasks)
arxiv-2610.06824
8 uncontaminated AI R&D tasks, 6 seeds/model/task, 40 H100-hours; Researcher model + fixed Coder (Opus 4.8); Table 2 of arXiv 2610.06824; score only (cost in round 2); read 2026-10-15 ✓xmodelGLM-5.3·max
TasteVal performance multiplier vs expert baseline
arxiv-2610.06824
8 uncontaminated AI R&D tasks, 6 seeds/model/task, 40 H100-hours; Researcher model + fixed Coder (Opus 4.8); Table 2 of arXiv 2610.06824; score only (cost in round 2); read 2026-10-15 ✓xmodelGLM-5.3·max
tau-bench 2 airline (text) leaderboard
github-sierra-tau2-bench
tau2-bench v1.0.1 airline domain, text modality, pass^1 (%) over 4 trials, user simulator gpt-5.2 low reasoning, standard tau-bench scaffold, submitted by Sierra Research (Feb-May 2026), board head 989d328 of 2026-10-20; read 2026-10-20 ✓xmodelGLM-5·think
tau-bench 2 retail (text) leaderboard
github-sierra-tau2-bench
tau2-bench v1.0.1 retail domain, text modality, pass^1 (%) over 4 trials, user simulator gpt-5.2 low reasoning, standard tau-bench scaffold, submitted by Sierra Research (Feb-May 2026), board head 989d328 of 2026-10-20; read 2026-10-20 ✓xmodelGLM-5·think
tau-bench 2 telecom (text) leaderboard
github-sierra-tau2-bench
tau2-bench v1.0.1 telecom domain, text modality, pass^1 (%) over 4 trials, user simulator gpt-5.2 low reasoning, standard tau-bench scaffold, submitted by Sierra Research (Feb-May 2026), board head 989d328 of 2026-10-20; read 2026-10-20 ✓xmodelGLM-5·think
Tinybird SQL benchmark
llm-benchmark.tinybird.live
Tinybird LLM SQL benchmark, 50 natural-language questions answered in SQL over 200M GH Archive rows, via OpenRouter · composite score (efficiency, exactness); no effort set (default); no cost; read 2026-10-01 ✓xmodelGLM-5.3-Flash·max · GLM-5.2·high · GLM-5.3-FlashX·max
TLA-Verus tlaverus_algo 40 tasks: p@3
tla-verus-2610.14923
TLA-Verus Table 12: tlaverus_algo 40 tasks, Verus proof generation with verifier feedback (<=10 attempts), 3 samples per task, temperature 1.0, all via OpenRouter, completion limit 32,768 (verusbench) / 128,000 (tlaverus); p@3 = tasks solved by >=1 of 3 samples (p@1, turns, tokens = same runs); cost = provider billed figure summed over the whole run; arXiv 2610.14923 (Oct 2026); read 2026-10-20 ✓xmodelGLM-5.2·high
TLA-Verus tlaverus_llm 40 tasks: p@3
tla-verus-2610.14923
TLA-Verus Table 13: tlaverus_llm 40 tasks, Verus proof generation with verifier feedback (<=10 attempts), 3 samples per task, temperature 1.0, all via OpenRouter, completion limit 32,768 (verusbench) / 128,000 (tlaverus); p@3 = tasks solved by >=1 of 3 samples (p@1, turns, tokens = same runs); cost = provider billed figure summed over the whole run; arXiv 2610.14923 (Oct 2026); read 2026-10-20 ✓xmodelGLM-5.2·high
TLA-Verus verusbench 150 tasks: p@3
tla-verus-2610.14923
TLA-Verus Table 11: verusbench 150 tasks, Verus proof generation with verifier feedback (<=10 attempts), 3 samples per task, temperature 1.0, all via OpenRouter, completion limit 32,768 (verusbench) / 128,000 (tlaverus); p@3 = tasks solved by >=1 of 3 samples (p@1, turns, tokens = same runs); cost = provider billed figure summed over the whole run; arXiv 2610.14923 (Oct 2026); read 2026-10-20 ✓xmodelGLM-5.2·high
Toolathlon-Verified leaderboard (current, 108 tasks)
toolathlon
Toolathlon-Verified (released 30 Jun 2026), 108 tasks, 32 MCP servers / 604 tools + 7 local toolkits, Pass@1 % (+-CI) over 3 runs; rows carry the 'Evaluated by us' badge unless noted; Agent 'Default'; page fetched 2026-10-20 (newest row 2026-08-30); cells identical to held cells (vendor cards/blogs copying the board) dropped; read 2026-10-20 ✓xmodelGLM-5.3-Flash·max
UGI Leaderboard — UGI composite (all Claude models × efforts)
ugi-leaderboard
DontPlanToEnd/UGI-Leaderboard, fixed prompt set, composite UGI score; test dates vary by model (CSV Test Date column); read 2026-10-05 ✓xmodelGLM-4.7·think
UGI Leaderboard — World Model factual recall (all Claude models × efforts)
ugi-leaderboard
DontPlanToEnd/UGI-Leaderboard, recipe/geoguessr/weight/music percent-error battery; test dates vary by model; read 2026-10-05 ✓xmodelGLM-4.7·think
UGI Leaderboard — Writing quality (all Claude models × efforts)
ugi-leaderboard
DontPlanToEnd/UGI-Leaderboard, fixed prompt set, LLM-judged Writing score; test dates vary by model (CSV Test Date column); read 2026-10-05 ✓xmodelGLM-4.7·think
Vals Code Migration
vals.ai
Vals AI · Code Migration, 10 Claude entries, compute_effort max (Haiku thinking null) — snapshot 27 Sep, re-priced ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals CorpFin v2
vals.ai
Vals AI · CorpFin v2, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.2·max
Vals CyberBench
vals.ai
Vals AI · OSS-Fuzz PoC+patch, mini-swe-agent, 5 Claude models at max — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals Finance Agent v2
vals.ai
Vals AI · Finance Agent v2, Claude models at their stated compute_effort, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.2·max
Vals GPQA
vals.ai
Vals AI · GPQA, Claude models at their stated compute_effort, measured cost/test, near-ceiling → cost only ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals HLAB
vals.ai
Vals AI · Harvey's Legal Agent Benchmark, all at max, several at floor — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals IOI
vals.ai
Vals AI · IOI, Claude models at their stated compute_effort, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals Legal Research
vals.ai
Vals AI · Legal Research, Claude models at their stated compute_effort, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals LegalBench
vals.ai
Vals AI · LegalBench, Claude models at their stated compute_effort, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals LiveCodeBench
vals.ai
Vals AI · LiveCodeBench, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals MedCode
vals.ai
Vals AI · ICD-10-CM coding, 2,755 records, all at max ✓xmodelGLM-5.3·max
Vals MedScribe
vals.ai
Vals AI · MedScribe, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals MMLU-Pro
vals.ai
Vals AI · MMLU-Pro, Claude models at their stated compute_effort, measured cost/test, near-ceiling → cost only ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals MMMU
vals.ai
Vals AI · MMMU, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.3-Flash·max
Vals MysteryMechanism
vals.ai
Vals AI · MysteryMechanism, Claude models at their stated compute_effort, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max
Vals ProgramBench
vals.ai
Vals AI · ProgramBench, max, mostly floor scores, 3 entries without cost — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals ProofBench
vals.ai
Vals AI · ProofBench v1.1, all at max (Opus 5 now max in payload), 4 saturated — snapshot 28 Sep, re-priced ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals Public Benefits v1.1
vals.ai
Vals AI · Public Benefits v1.1, Claude models at their stated compute_effort, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max
Vals SkillsBench
vals.ai
Vals AI · SkillsBench with skills, OpenHands, 5 Claude models at max — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max
Vals SRE-Bench
vals.ai
Vals AI · binary reverse engineering (despite the name), all at max — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max · GLM-5.3-Flash·max
Vals SWE-bench
vals.ai
Vals AI · SWE-bench, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.2·max · GLM-5.3·max · GLM-5.3-Flash·max
Vals Tax Agent
vals.ai
Vals AI · Tax Agent, 10 Claude entries at their stated compute_effort — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max
Vals TaxEval v2
vals.ai
Vals AI · TaxEval v2, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.3·max · GLM-5.2·max · GLM-5.3-Flash·max
Vals TB 2.1
vals.ai
Vals AI · Terminal-Bench 2.1, Claude models at their stated compute_effort (Sonnet 5.5, Opus 5.5 high) — snapshot 27 Sep, re-priced ✓xmodelGLM-5.2·max
Vals TB 4.0
vals.ai
Vals AI · Terminal-Bench 4.0, all Claude at max, measured cost/test — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max
Vals TB-Science
vals.ai
Vals AI · Terminal-Bench-Science, all at max — snapshot 27 Sep, re-priced ✓xmodelGLM-5.3·max
Vals Vibe Code Bench 1-100
vals.ai
Vals AI · Vibe Code Bench 1-100, Claude models at their stated compute_effort, measured cost/test ✓xmodelGLM-5.3·max
valscodemig11
vals.ai
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
valsemb11
vals.ai
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
valsfincoach
vals.ai
—xmodelGLM-5.3·max
valshlab-2026-10-21
vals.ai
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
valsimmigration
vals.ai
—xmodelGLM-5.3·max
valsindex4
vals.ai
—xmodelGLM-5.3·max
valslegalindex
vals.ai
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
valssage11
vals.ai
—xmodelGLM-5.3-Flash·max
valstaxwp
vals.ai
—xmodelGLM-5.3·max · GLM-5.3-Flash·max
vlm-exam counting
github-vlmexam
roboflow/vlm-exam · 74 samples, LLM-judge accuracy ✓xmodelGLM-5.3-Flash·low
vlm-exam detection
github-vlmexam
roboflow/vlm-exam · 250 images, mAP@50 ✓xmodelGLM-5.3-Flash·low
vlm-exam extraction
github-vlmexam
roboflow/vlm-exam · 97 samples, LLM-judge accuracy; same models as OCR ✓xmodelGLM-5.3-Flash·low
vlm-exam identification
github-vlmexam
roboflow/vlm-exam · 32 samples, near ceiling ✓xmodelGLM-5.3-Flash·low
vlm-exam OCR
github-vlmexam
roboflow/vlm-exam · 37 samples, similarity; Fable 5.1 & Opus 5.5 low/high ×3 runs, others low ×1 ✓xmodelGLM-5.3-Flash·low
vlmexam-text
github-vlmexam
—xmodelGLM-5.3-Flash·low · GLM-5.3-Flash·high
How the numbers are built — the method

Every value in this report is computed from the raw measurements by one procedure — nothing is hand-entered. The full methodology is in METHODOLOGY.md (from collection to the fit) and DISPLAY-METHODOLOGY.md (from the fit to the tier picks); the validation scripts are in model/validation.

In short : scores from different benchmarks do not share a scale — a benchmark near its ceiling compresses differences, a hard one inflates them — so the report never averages raw scores or score ratios. A latent-quality model estimates, for every benchmark, its own zero and its own gain, and for every (model, effort) couple one quality and one cost. Each couple stays atomic: no model×effort separability is assumed.

The procedure

  1. Every number traced to its source. Each source is recorded as facts: the value as the source writes it, the exact text or table cell it is read from, the URL and the date of the capture; agents read every value line by line, and a measurement whose facts lack their proof stays out. Registries of models (the names each was served under, its effort settings and defaults, dated), routes (makers' APIs, resellers, routers, harnesses), benchmarks and their versions, publishers and dated prices let one build decide, by a fixed decision tree, which model, configuration and price each measurement belongs to.
  2. Same task, same configuration. Measurements are compared only within a group — one source, one task, one harness, one setting, one scale, one version of a board; a group that mixes them is split. Every run is dated, and its model and effort are read as the maker and every route in between served them on that date — never assumed: a route the source does not name is never presumed to be the maker's API, and every route that could have served the run must give the same configuration. At the least doubt on the model, the effort or what was run, the measurement stays in the data file and out of the fit.
  3. Costs at today's prices. Every cost is put on the model's current price from the prices in force on its run date: the source's own figure at that price, the price change applied to the run's own token counts, or, when the token mix is unknown, a price band whose width is added to the cost's precision. A cost that cannot be brought to the current price stays out.
  4. Each metric on its natural scale. A bounded score (a percentage, a score out of N) goes through a logit, which removes floor and ceiling compression; money and points go through a log; an Elo or an unknown composite is left as is; cost goes through a log. The scale is read from the metric's label, never from the benchmark's identity; whether lower is better is read from the benchmark's registry, and such a score is turned in its own unit first (an attack success rate of 30 % reads 70 %); a score whose direction is not established stays out. On the logit scale a score near 0 % or 100 % is noisier, and weighs accordingly. Every value carries its reading precision: the rounding of a printed number, and for a value read off a chart, its resolution and the chart's own error (measured on runs drawn on two charts, about 2 % on a digitised cost).
  5. One model for all groups. Every group gets its own offset, gain and noise, so the unit and zero of a metric never matter. A couple's latent value is shared by all groups; on top of it, a publisher's systematic effect on a couple and on a model (shared by the model's effort levels) and a couple's deviation on a task type are estimated, so a publisher with many benchmarks, or a couple tested mostly on one kind of task, cannot tilt the result. The sizes of these effects are estimated for each model maker on its own, not tied to the other makers'; a maker's own charts about its own models carry an estimated home effect, also each maker's own, left out of every value. The fit takes the couples of eight makers measured in the same groups (Anthropic, OpenAI, Google, xAI, DeepSeek, Qwen, GLM, Kimi), on one scale; this page shows the GLM couples. Robust (Student-t) noise whose tails are estimated; early-access runs are down-weighted by an estimated factor. No couple serves as a reference in the fit: the latent values sum to zero. Fitted in Stan, published only if the standard convergence checks pass.
  6. Weights follow from the data. A group of k couples informs the spacing of couples in proportion to its discrimination² × (k − 2); a saturated or noisy benchmark weighs little; a publisher's weight saturates at its own systematic effect. No weight is set per source; the one weight set by rule is that of a single run: when a source gives k quality measurements of one run (several metrics, sub-scores or slices of the same executions), each weighs 1/k, so the run counts as one measurement.
  7. Nothing counts twice. One execution printed twice — a system card reprinting an earlier card's series, the same run drawn on two charts, a board copying another's numbers, one run scored with two metrics — is detected and counted once: its score and its cost must both agree, within their reading precisions and a rounding made the wrong way. Versions of one board that re-grade the same runs (recognised by identical costs) count those runs once. Before the fit, two groups that still share identical costs on three couples or more stop it until the copy is resolved or the coincidence proven from the sources. An index keeps, for a couple whose index components are already in the data, only what its missing components add, or is left out.
  8. Values and intervals. Quality = a couple's latent quality θ, shown as the score it predicts averaged over the benchmark panel; cost = its cost on a task of typical size, shown as a multiple of the cheapest couple. No couple serves as a reference, in the fit or on the page. Each couple carries its own 16–84 % interval, so two couples compare through their two intervals. A couple measured by a single publisher is not shown.
  9. From the fit to the picks. Every decision is taken on the latent scale (log cost, θ). The Pareto frontier keeps the couples nothing beats on both axes; they are the candidates of every pick. A price trend — what a given quality costs today, fitted on the latest model of each family — says how many times cheaper or dearer each couple is than the going rate. The four tiers spread their targets across the frontier and pick, around each target, the couple that brings the most quality for its cost, each tier paying for quality from a quarter of the market price (grunt work) to four times it (cutting-edge thinking); the crown goes to the couple furthest below the trend.

Key values

  • Set by hand, never tuned to the data: weakly informative priors (half-Student-t(3, 0, 2.5) on every spread, Gamma(2, 0.1) on the noise tails); a logit kept half a step inside 0 and 100 %; a printed value's precision at its rounding (last digit / √12); an effect seen in a single measurement left out; a couple shown only when two publishers measured it; a reprint recognised by values that agree within their precisions, on two values at least and on all they share but one.
  • Uncertainty. Every interval is 16–84 % (one standard deviation): the couple's own uncertainty.
  • Published only when the fit is sound: R̂ ≤ 1.01 and at least 400 effective draws on every parameter, no divergent transition. The Monte Carlo error of each value is checked against its last displayed digit; in this fit every figure on this page is stable to that digit.
  • Display choices: the quality axis is compressed below the weakest tier target; the frontier, the price trend and the four tiers use constants listed with their reasons in DISPLAY-METHODOLOGY.md.

Techniques and software

  • Bayesian hierarchical model sampled by Hamiltonian Monte Carlo with the No-U-Turn sampler (Hoffman & Gelman 2014), written in Stan and run with nutpie on both axes, six chains whose draws continue in the same run until the criteria are met (CmdStan only to resume a stopped fit from its stored draws, or to restart a chain left in a far region from a good chain's last draw); built and checked stage by stage following the Bayesian workflow.
  • Convergence checked with the rank-normalised R̂ and effective sample sizes of Vehtari et al. 2021, computed with ArviZ.
  • Intervals as quasi-standard errors (Firth & de Menezes 2004), so that any two couples compare through their own intervals.
  • Scales: the empirical logit of bounded scores, with the binomial shape of its noise.

Checks (on the fit of 29 September 2026, before the other makers' couples joined the data). Scores held out of the fit and predicted: median error 1.96 points, against 4.78 for averaging score ratios; the 16–84 % predictive interval covers 66 % of them. On synthetic data with known qualities, pairs of couples in the wrong order: 0.1–0.8 % (0.4–2.2 % for score ratios). Without the largest publisher, the model vendor itself, costs move by 2.4 % and qualities by 0.9 % at the median. Details and scripts: METHODOLOGY.md.

Alexandre Gensse · github.com/alexandregensse-blip

All figures are indicative and for informational purposes only — derived from public third-party measurements, not an official benchmark, and not affiliated with or endorsed by any model maker. Verify before relying on any value.

Data and text under CC BY 4.0 · code under MIT · download the data (CSV)