Reasoning and knowledge
- Gemini 4 Argon96.4%
- GPT-6.1 Sol95.4%
- Grok 4.792.7%
- Qwen3.8-Max92.3%
- DeepSeek V4 Pro (0813)91.7%
- Z.ai GLM-5.390.9%
- Claude Opus 5.590.6%
- Mistral Large 488.1%
Loading
Benchmarks
Published scores, with the source on the number.
Each board uses one published test. A model appears only when that test has a score. Vendor-reported figures stay behind their own filter.
Independent scores only, unless the vendor-reported filter is on. Reasoning and knowledge: Gemini 4 Argon at 96.4%. Maths: GPT-6.1 Sol at 93.7%. Coding: Claude Opus 5.5 at 96.3. Instruction following: Claude Opus 5.5 at 1521. Human preference: Gemini 4 Argon at 1517. Overall LiveBench: Claude Opus 5.5 at 86.5.
Epoch’s own GPQA Diamond run. A one-point gap here is inside the published error bar for several models. Ranks use independent scores only.
| Rank | Model | Score | Price in / out, $ / 1M |
|---|---|---|---|
| 1 |  Gemini 4 ArgonGoogle | 96.4% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.2Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = high; run started 2026-10-13Epoch AI Benchmarking Hub (internal run) | $2.00 / $10.00 |
| 2 |  GPT-6.1 SolOpenAI | 95.4% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.4Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = max; run started 2026-09-29Epoch AI Benchmarking Hub (internal run) | $2.00 / $10.00 |
| 3 |  Grok 4.7xAI (SpaceXAI) | 92.7% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.6Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = xhigh; run started 2026-09-22Epoch AI Benchmarking Hub (internal run) | $2.00 / $6.00 |
| 4 |  Qwen3.8-MaxAlibaba | 92.3% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.7Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = xhigh; run started 2026-09-02Epoch AI Benchmarking Hub (internal run) | $2.00 / $6.00 |
| 5 |  DeepSeek V4 Pro (0813)DeepSeek | 91.7% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.5Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = max; run started 2026-08-18Epoch AI Benchmarking Hub (internal run) | $1.32 / $3.96 |
| 6 |  Z.ai GLM-5.3Z.ai (Zhipu) | 90.9% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.6Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = max; run started 2026-08-24Epoch AI Benchmarking Hub (internal run) | $1.40 / $4.40 |
| 7 |  Claude Opus 5.5Anthropic | 90.6% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.9Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = max; run started 2026-09-22Epoch AI Benchmarking Hub (internal run) | $4.00 / $20.00 |
| 8 |  Mistral Large 4Mistral AI | 88.1% IndependentGPQA DiamondIndependent · accessed 2026-10-26Standard error 1.9Epoch Inspect harness, zero-shot CoT, API-default temperature, empty system prompt; mean of repeated runs (usually 16); reasoning effort = high; run started 2026-10-12Epoch AI Benchmarking Hub (internal run) | $1.36 / $4.18 |
Price is the published list rate per million tokens. An unverified price is named in the link title. Blended charts use 25% input and 75% output.
Reasoning and knowledge
Reasoning and knowledge against list price
Vertical axis is the score in this category. Horizontal axis is blended list price, 25% input and 75% output, on a log scale. An unverified price is still plotted and named in the label.
Sources and caveats
The file was read on 2026-10-26. 115 scores count (97 independent, 18 vendor-reported). A test is shown when at least three independent runs exist.
- Vendor-reported numbers are not comparable with each other or with independent runs.
- 8 scores for a different checkpoint are excluded from these rankings.
- LMArena ratings are human preference, not accuracy.
- Data: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from https://epoch.ai/benchmarks. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores are shown as percentages, rounded to one decimal.
- Human-preference ratings: LMArena Text Arena leaderboard (lmarena.ai), via the lmarena-ai/leaderboard-dataset on Hugging Face, licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Ratings rounded to whole numbers.
- LiveBench 2.0 results from https://livebench.ai (White et al., LiveBench). Benchmark code licensed Apache-2.0.
- Vendor-reported scores are taken from each developer’s published model card or launch post, linked per score. They were not independently verified by PowerfulBot.