AI model ranking: intelligence, benchmarks and price
Each dot is a model: the higher, the better it ranks; the further left, the cheaper it is. Click a model to see it alone.
| Model | Developer | Open weights | License | Score | Input / output price | Blended price | GPQA Diamond |
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5Pareto | 🇺🇸 Anthropic | no | Proprietary | 167.3 | €3.57 / €17.85 | €7.14 | 90.6 % |
| GPT-6 Astra | 🇺🇸 OpenAI | no | Proprietary | 166.4 | €8.92 / €44.62 | €17.85 | 95.8 % |
| Claude Sonnet 5.5Pareto | 🇺🇸 Anthropic | no | Proprietary | 165.0 | €1.79 / €8.92 | €3.57 | 95.6 % |
| Claude Fable 5.1 | 🇺🇸 Anthropic | no | Proprietary | 164.7 | €8.92 / €44.62 | €17.85 | — |
| GPT-5.6 Sol | 🇺🇸 OpenAI | no | Proprietary | 161.7 | €3.57 / €17.85 | €7.14 | 93.5 % |
| GPT-5.6 Terra | 🇺🇸 OpenAI | no | Proprietary | 159.6 | not published | — | 93.3 % |
| Kimi K3 | 🇨🇳 Moonshot AI | yes | Kimi K3 license | 157.4 | not published | — | 93.1 % |
| Gemini 3.7 FlashPareto | no | Proprietary | 157.3 | €0.67 / €3.35 | €1.34 | 94.8 % | |
| Muse Spark 1.3 | 🇺🇸 Meta | no | Proprietary | 156.8 | not published | — | — |
| Gemini 3.8 Flash | no | Proprietary | 156.7 | €0.67 / €3.35 | €1.34 | 95.4 % | |
| GPT-5.6 Luna | 🇺🇸 OpenAI | no | Proprietary | 156.4 | not published | — | 91.6 % |
| Qwen3.8 Max | 🇨🇳 Alibaba | no | Proprietary | 156.4 | not published | — | 92.7 % |
| Grok 4.6 | 🇺🇸 xAI | no | Proprietary | 156.4 | €1.79 / €5.35 | €2.68 | 94.0 % |
| GLM-5.3 | 🇨🇳 Z.ai | yes | MIT | 155.6 | €1.25 / €3.93 | €1.92 | 90.9 % |
| DeepSeek V4 Pro | 🇨🇳 DeepSeek | yes | MIT | 155.3 | €1.18 / €3.53 | €1.77 | 91.7 % |
| DeepSeek V4.1 FlashPareto | 🇨🇳 DeepSeek | yes | MIT | 154.9 | €0.27 / €1.07 | €0.47 | — |
| Grok 4.5 | 🇺🇸 xAI | no | Proprietary | 153.9 | €1.79 / €5.35 | €2.68 | 93.4 % |
| GLM-5.3 FlashPareto | 🇨🇳 Z.ai | yes | MIT | 151.9 | €0.13 / €0.45 | €0.21 | 90.2 % |
| Inkling Small | 🇺🇸 Thinking Machines | yes | Apache 2.0 | 150.2 | not published | — | 88.5 % |
| Kimi K2.7 Code | 🇨🇳 Moonshot AI | yes | — | 150.0 | not published | — | 87.9 % |
| Qwen3.8 27B | 🇨🇳 Alibaba | yes | Apache 2.0 | 149.4 | not published | — | — |
| Inkling | 🇺🇸 Thinking Machines | yes | Apache 2.0 | 148.5 | not published | — | 88.3 % |
| MiniMax M3 | 🇨🇳 MiniMax | yes | MiniMax Community License | 146.9 | not published | — | 90.9 % |
| Qwen3.5 397B | 🇨🇳 Alibaba | yes | Apache 2.0 | 146.7 | not published | — | 86.4 % |
| Nemotron 3 Ultra | 🇺🇸 Nvidia | yes | OpenMDW-1.1 | 146.2 | not published | — | 85.4 % |
| Gemini 3.5 Flash-Lite | no | Proprietary | 145.1 | €0.27 / €2.23 | €0.76 | 83.3 % | |
| Qwen3.7 Flash | 🇨🇳 Alibaba | no | Proprietary | 144.6 | not published | — | 82.3 % |
| Qwen3.6 35B-A3B | 🇨🇳 Alibaba | yes | — | 143.9 | not published | — | 84.8 % |
| Gemma 4 31B | yes | Apache 2.0 | 142.7 | not published | — | 75.8 % | |
| Claude Haiku 4.5 | 🇺🇸 Anthropic | no | Proprietary | 142.4 | €0.89 / €4.46 | €1.78 | 71.2 % |
| Gemma 4 26B A4B | yes | Apache 2.0 | 141.8 | not published | — | 73.2 % | |
| Mistral Medium 3.5 | 🇫🇷 Mistral AI | yes | Modified MIT | 141.3 | €1.34 / €6.69 | €2.68 | — |
| gpt-oss-120b | 🇺🇸 OpenAI | yes | Apache 2.0 | 139.9 | not published | — | 75.8 % |
| Qwen3.5 9B | 🇨🇳 Alibaba | yes | — | 139.5 | not published | — | 79.0 % |
| gpt-oss-20b | 🇺🇸 OpenAI | yes | Apache 2.0 | 137.8 | not published | — | 60.8 % |
| Llama 4 Maverick | 🇺🇸 Meta | yes | Llama 4 | 132.2 | not published | — | 67.0 % |
| Mistral Small 3.2 | 🇫🇷 Mistral AI | yes | Apache 2.0 | 131.7 | not published | — | 49.1 % |
Method
Score: Epoch AI's capability index, summarizing many benchmarks run by Epoch (GPQA, SWE-bench, FrontierMath…). Price: developers' official rates, converted to euros at the ECB rate; the blended price weights input 3/4 and output 1/4. Models without an official price are in the table but not on the chart.
Sources: Scores: Epoch AI, Capabilities & benchmarking (CC BY 4.0), up to 2026-09-29 · Prices: developers' pages, checked 2026-10-03 · ECB rate of 2026-10-09: €1 = $1.1206
FAQ
What does “open weights” mean?
Weights are the billions of numbers that hold what a model has learned. An open-weight model can be downloaded and run on your own machine; a closed model is only available through its developer's API. Open weights does not mean open source: the license may restrict use.
How is model intelligence measured?
With Epoch AI's Capabilities Index (ECI), which combines the results of many benchmarks: scientific reasoning (GPQA Diamond), coding (SWE-bench Verified), mathematics (FrontierMath), factual accuracy (SimpleQA Verified).
Where do the prices come from?
From the developers' official pricing pages, in dollars per million tokens, converted to euros at the European Central Bank reference rate. The blended price counts 3/4 of the input price and 1/4 of the output price.
What is the Pareto line?
It links the models that no other model beats on both intelligence and price: the best trade-offs right now.