BULLETIN / BENCHMARKS
Benchmarks, from across the internet.
14 models, 12 independent evaluations, 44 scores. Every number links to the page it was read from.
Scores were read from the named evaluations on 2026-09-30. Higher is better unless the column says otherwise. A blank cell means the evaluation had not published a score for that model, or the review only had a vendor-reported figure. Vendor-reported numbers are not on this page; each review discusses them separately.
| Model | AA IndexArtificial Analysis | Arena textArena (arena.ai) | Vals IndexVals AI | Epoch ECIEpoch AI | FrontierMath T1-3Epoch AI | GPQA DiamondEpoch AI | ARC-AGI-2ARC Prize | Arena WebDevArena (arena.ai) | Arena visionArena (arena.ai) | Terminal-Bench 4.0Artificial Analysis | HLE (AA)Artificial Analysis | HLE (Scale)Scale Labs | Price inUSD/M tokens | Price outUSD/M tokens | Contexttokens |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5Anthropic · closed | 58max effort | 1509 | 66.97% | · | 91.2% | 90.6% | 93.3% | 1820 | · | 59.6% | 61.4% | · | 4 | 20 | 1M tokens |
| GPT-6.1 SolOpenAI · closed | 52max effort | · | · | · | · | · | · | · | · | · | · | · | 2 | 10 | 1,050,000 tokens |
| Gemini 3.8 FlashGoogle · closed | 41high effort | 1492 | 54.83% | 157.1value | · | 95.4% | 89.2% | · | · | · | · | 44.52 | 0.75 | 3.75 | 1,048,576 input tokens |
| Grok 4.7xAI · closed | 46high and xhigh effort both scored 46 | · | 54.95% | · | · | · | · | 1636xhigh | · | · | · | · | 2 | 6 | 500,000 tokens |
| Muse Spark 1.3Meta · closed | 48max effort | · | 58.16%max effort | 156.9 | 74.4%xhigh effort | · | · | 1655 | 1290max effort | · | · | · | 1.25 | 4.25 | 1M tokens |
| Mistral Medium 3.5Mistral AI · open weights | 14 | · | · | · | · | · | · | · | · | · | · | · | 1.5 | 7.5 | 256k tokens |
| DeepSeek-V4.1-FlashDeepSeek · open weights | 39max effort | · | 51.32% | · | · | · | · | · | · | · | · | · | 0.30 | 1.20 | 1M tokens |
| Qwen3.8-MaxAlibaba Cloud (Qwen) · closed | 450902 snapshot | · | 48.27% | · | · | · | · | 1671 | 1301 | · | · | · | 2 | 6 | 1,000,000 tokens |
| Kimi K3Moonshot AI · open weights | 44 | 1488 | 50.70% | · | · | · | 60.4% | 1659 | · | · | · | · | 3.00 | 15.00 | 1,048,576 tokens |
| Gemma 4 (E2B, E4B, 12B Unified, 26B A4B, 31B)Google · open weights | 1931B model | · | · | 142.731B model | · | · | · | · | · | · | · | · | · | · | 128K tokens (E2B, E4B); 256K tokens (12B, 26B A4B, 31B) |
| gpt-oss-120bOpenAI · open weights | 12high effort | · | · | · | · | · | · | · | · | · | · | · | · | · | 131,072 tokens |
| Granite 4.2 (3B, 8B, 30B)IBM · open weights | · | · | · | · | · | · | · | · | · | · | · | · | · | · | 128K tokens native, extendable to 512K (8B model card) |
| Sarvam 30B and Sarvam 105BSarvam AI · open weights | · | · | · | · | · | · | · | · | · | · | · | · | · | · | 65,536 tokens (Sarvam 30B) |
| MiMo-V2.6-ProXiaomi · open weights | 46 | 1480 | 55.20% | · | · | · | · | · | · | · | · | · | · | · | 1M tokens |
Where the numbers come from
- Artificial Analysis Intelligence Index, published by Artificial Analysis.
- Arena text score, published by Arena (arena.ai).
- Vals Index, published by Vals AI.
- Epoch Capabilities Index, published by Epoch AI.
- FrontierMath Tiers 1 to 3, published by Epoch AI.
- GPQA Diamond (Epoch), published by Epoch AI.
- ARC-AGI-2, published by ARC Prize.
- Arena web development score, published by Arena (arena.ai).
- Arena vision score, published by Arena (arena.ai).
- Terminal-Bench 4.0 (measured by Artificial Analysis), published by Artificial Analysis.
- Humanity's Last Exam (measured by Artificial Analysis), published by Artificial Analysis.
- Humanity's Last Exam (Scale Labs leaderboard), published by Scale Labs.
Prices and context windows come from each vendor's own documentation, as cited in the reviews. Scores change as evaluations add runs; the date at the top is the date every cell was read.