Leaderboard
Every score here carries the health of the benchmark that produced it and who reported the number. A high score on a saturated benchmark is still a high score — it just tells you less than it appears to.
At a glance
Each card is one row from the leaderboard, picked by the rule in its heading. “Highest on a benchmark” is not “best”.
Highest on GPQA
GPT-6 Astra (max)OpenAI
95.8%ind.
nearing saturationsupersededabout GPQA
Highest with open weights on GPQA
DeepSeek V4 Pro (max)DeepSeek
91.7%ind.
nearing saturationsupersededabout GPQA
Among models recorded as releasing their weights. 258 of the 286 ranked rows have no recorded answer and are not counted either way.
Lowest price within 10% of the top on GPQA
DeepSeek V4.1 Flash (max)DeepSeek
89.8%ind.
$0.04 ~ in, $0.08 ~ out, per million tokens
nearing saturationsupersededabout GPQA
Ranked by input plus output price per million tokens. Models with no readable price are not counted.
Highest on Terminal-Bench
GPT-5.5OpenAI
84.7%ind.
activecontaminatedabout Terminal-Bench
A system score: the model and its harness measured together.
Highest Arena Elo, text to image
gpt-image-2.5-sunburstOpenAI
1425 ±6.8Arena
No health flags: no benchmark health record exists for an arena board.
A preference ranking from public pairwise votes, not a benchmark score.
Ranked on GPQAnearing saturationsupersededon every rowind.every number independently measured13 of 80 benchmarks have scoresfetched 2026-10-19
Why this ranking?
GPQA is the benchmark most models here are measured on. Scores on different benchmarks are not comparable, so only models measured on this one are ranked. It is not the healthiest benchmark here, and it is the basis anyway because it is the most measured: Humanity's Last Exam is healthier but covers 49 of the 286 model-variants GPQA does. Read the ranking as “best on GPQA”, not “best”.
13 of the 80 benchmarks tracked here have any score at all, so this ranking describes the part of the field that Epoch and Arena have measured, not the whole of it. The other 67 are catalogued with their health.
Every number in this view was measured independently of the model's maker. Every row carries superseded; the Health column shows only what a row adds to that. Shown once rather than repeated down a column that would read the same on every row.
- ind.
- independently measured
- vendor
- reported by the model's maker
- ~
- OpenRouter's routed price, not a list price
- ◇
- open weights
- (max)
- the setting a score was measured at
- —
- not reported
| # | Model | GPQA | Epoch index | Context | Price in / out | Released | Health |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra (max)OpenAI | 95.8% | 166.4 | 1.05M ~ | $10.00 ~ / $50.00 ~ | 2026-09-03 | no flags beyond those on every row |
| 2 | Claude Sonnet 5.5 (max)Anthropic | 95.6% | 164.4 | 1M ~ | $2.00 ~ / $10.00 ~ | 2026-09-28 | no flags beyond those on every row |
| 3 | GPT-6.1 Sol (max)OpenAI | 95.4% | 166.3 | 1.05M ~ | $2.00 ~ / $10.00 ~ | 2026-09-29 | no flags beyond those on every row |
| 4 | Gemini 3.8 Flash (high)Google DeepMind | 95.4% | 157.2 | not reported | not reported / not reported | 2026-09-02 | no flags beyond those on every row |
| 5 | Gemini 3.7 Flash (high)Google DeepMind | 94.8% | 157.2 | 1.05M ~ | $0.75 ~ / $3.75 ~ | 2026-08-13 | no flags beyond those on every row |
| 6 | GPT-5.4 Pro (xhigh)OpenAI | 94.6% | 159.0 | not reported | not reported / not reported | 2026-03-05 | no flags beyond those on every row |
| 7 | Gemini 3.1 ProGoogle DeepMind | 94.4% | 154.7 | not reported | not reported / not reported | 2026-02-19 | disputed |
| 8 | GPT-6 Sol (max)OpenAI | 94.3% | 162.6 | 1.05M ~ | $2.00 ~ / $10.00 ~ | 2026-09-22 | no flags beyond those on every row |
| 9 | Gemini 3.6 Flash (high)Google DeepMind · best of 3 configurations | 94.1% | 154.2 | 1.05M ~ | $0.75 ~ / $3.75 ~ | 2026-07-21 | no flags beyond those on every row |
| 10 | Grok 4.6 (high)xAI · best of 2 configurations | 94.0% | 156.3 | not reported | not reported / not reported | 2026-08-12 | no flags beyond those on every row |
| 11 | gpt-5.5-pre-release (xhigh)OpenAI | 94.0% | not reported | not reported | not reported / not reported | 2026-04-23 | no flags beyond those on every row |
| 12 | gpt-5.5-pro-pre-release (xhigh)OpenAI | 93.9% | not reported | not reported | not reported / not reported | 2026-04-23 | no flags beyond those on every row |
| 13 | Claude Opus 5 (max)Anthropic · best of 3 configurations | 93.9% | 162.6 | 1M ~ | $4.00 / $20.00 | 2026-07-24 | no flags beyond those on every row |
| 14 | GPT-5.6 Sol (max)OpenAI · best of 3 configurations | 93.5% | 161.7 | 1.05M ~ | $4.00 ~ / $20.00 ~ | 2026-07-09 | no flags beyond those on every row |
| 15 | Grok 4.5 (high)xAI | 93.4% | 153.9 | not reported | not reported / not reported | 2026-07-08 | no flags beyond those on every row |
| 16 | GPT-5.6 Terra (max)OpenAI · best of 3 configurations | 93.3% | 159.4 | 1.05M ~ | $2.00 ~ / $12.00 ~ | 2026-07-09 | no flags beyond those on every row |
| 17 | GPT-5.4 (xhigh)OpenAI · best of 5 configurations | 93.3% | 156.7 | not reported | not reported / not reported | 2026-03-05 | no flags beyond those on every row |
| 18 | Kimi K3 (max)Moonshot · best of 3 configurations | 93.1% | 157.4 | not reported | not reported / not reported | 2026-07-16 | no flags beyond those on every row |
| 19 | Gemini 3.5 Flash (high)Google DeepMind · best of 3 configurations | 92.8% | 154.4 | not reported | not reported / not reported | 2026-05-19 | no flags beyond those on every row |
| 20 | Grok 4.7 (xhigh)xAI | 92.7% | 154.4 | not reported | not reported / not reported | 2026-09-21 | no flags beyond those on every row |
| 21 | Qwen3.8-Max (xhigh)Alibaba | 92.7% | not reported | not reported | not reported / not reported | 2026-08-02 | no flags beyond those on every row |
| 22 | Gemini 3 ProGoogle DeepMind | 92.6% | 152.9 | not reported | not reported / not reported | 2025-11-18 | no flags beyond those on every row |
| 23 | Qwen3.8-Max-0902 (xhigh)Alibaba | 92.3% | 154.9 | not reported | not reported / not reported | 2026-09-01 | no flags beyond those on every row |
| 24 | GLM-5.2 (max)Z.ai (Zhipu AI) · best of 3 configurations | 91.9% | 151.9 | not reported | not reported / not reported | 2026-06-16 | no flags beyond those on every row |
| 25 | DeepSeek V4 Pro (max) ◇DeepSeek · best of 2 configurations | 91.7% | 149.2 | 1.05M ~ | $0.96 ~ / $1.91 ~ | 2026-04-24 | no flags beyond those on every row |
Showing 25 of 210 ranked rows.
Not ranked here
166 models are on the site but cannot be placed in the ranking above. That is information rather than an omission: a model released days ago with no independent scores yet is worth seeing.