Global · AI models
AI model rankings: how far Europe's best model is from the frontier
Models from European developers took 0.26% of the tokens routed through OpenRouter in the 30 days to 11 Oct 2026, against 58.8% for Chinese and 29.2% for US developers. On Arena's text leaderboard (8 Oct 2026), Europe's best model, mistral-large-4, ranks #115, 96 points behind #1 gemini-4-argon-high. Sources: OpenRouter, Arena, both CC BY 4.0.
Updated 12 October 2026 · data build 2026-10-10 · how we know · corrections
Europe's best model ranks #115 on Arena, 96 points behind the leader
Arena text leaderboard with style control, overall, 8 Oct 2026. 414 models ranked by human preference votes. Measured
| Rank | Model | Developer | Bloc | Score | 95% interval | Votes | Licence |
|---|---|---|---|---|---|---|---|
| 1 | gemini-4-argon-high | US | 1,525 | 1,517–1,534 | 4,892 | Proprietary | |
| 2 | claude-opus-5.5-high | anthropic | US | 1,507 | 1,499–1,515 | 6,272 | Proprietary |
| 3 | claude-opus-4-6-high | anthropic | US | 1,504 | 1,501–1,508 | 79,415 | Proprietary |
| 4 | claude-fable-5-high | anthropic | US | 1,504 | 1,499–1,508 | 40,026 | Proprietary |
| 5 | claude-opus-4-7-high | anthropic | US | 1,501 | 1,497–1,505 | 65,192 | Proprietary |
| 6 | claude-fable-5.1-max | anthropic | US | 1,501 | 1,494–1,507 | 13,063 | Proprietary |
| 7 | claude-opus-4-6 | anthropic | US | 1,498 | 1,494–1,501 | 84,095 | Proprietary |
| 8 | gemini-3.8-flash-high | US | 1,497 | 1,492–1,501 | 31,437 | Proprietary | |
| 9 | muse-spark-1.3-max | meta | US | 1,494 | 1,488–1,500 | 14,694 | Proprietary |
| 10 | claude-opus-4-7 | anthropic | US | 1,494 | 1,490–1,498 | 66,315 | Proprietary |
| 11 | muse-spark-1.2 (xHigh) | meta | US | 1,492 | 1,484–1,500 | 6,118 | Proprietary |
| 12 | muse-spark-1.1 | meta | US | 1,491 | 1,486–1,495 | 39,847 | Proprietary |
| 13 | claude-opus-5-high | anthropic | US | 1,490 | 1,486–1,494 | 62,934 | Proprietary |
| 14 | claude-opus-5-max | anthropic | US | 1,489 | 1,485–1,494 | 30,569 | Proprietary |
| 15 | muse-spark | meta | US | 1,489 | 1,483–1,495 | 14,168 | Proprietary |
Top 15 of 414. Ranks are Arena's own; overlapping intervals are statistical ties. In the top 10: US 10. In the top 50: US 40, China 10.
Download CSV ↓Source: Arena leaderboard dataset (CC BY 4.0) · arena.ai leaderboard · as of 2026-10-08 · Arena data by Arena, CC BY 4.0 · Compiled by StackPath
Europe's best rank slid from #16 to #115 since 2024-Q3; China's best is #16
Arena points between the #1 model and the best European (Mistral) and best Chinese model in the same snapshot. Lower is closer. Scores are not comparable across snapshots, gaps and ranks within a snapshot are. Measured
| Quarter | #1 model | Best European | Rank | Gap | Best Chinese | Rank | Gap |
|---|---|---|---|---|---|---|---|
| 2024-Q3 | o1-preview (openai) | mistral-large-2407 | 16 | 81 | qwen2.5-72b-instruct | 18 | 89 |
| 2024-Q4 | o1-2024-12-17 (openai) | mistral-large-2407 | 24 | 97 | deepseek-v3 | 4 | 54 |
| 2025-Q1 | chatgpt-4o-latest-20250326 (openai) | mistral-large-2407 | 46 | 127 | deepseek-v3-0324 | 5 | 40 |
| 2025-Q2 | gemini-2.5-pro (google) | mistral-medium-2505 | 19 | 86 | deepseek-r1-0528 | 7 | 54 |
| 2025-Q3 | gemini-2.5-pro (google) | mistral-medium-2508 | 27 | 45 | qwen3-max-preview | 8 | 20 |
| 2025-Q4 | gemini-3-pro (google) | mistral-large-3 | 41 | 75 | ernie-5.0-preview-1203 | 11 | 43 |
| 2026-Q1 | claude-opus-4-6-thinking (anthropic) | mistral-large-3 | 69 | 88 | qwen3.5-max-preview | 14 | 37 |
| 2026-Q2 | claude-fable-5 (anthropic) | mistral-medium-3.5 | 78 | 81 | qwen3.7-max-preview | 16 | 33 |
| 2026-Q3 | gemini-4-argon-high (google) | mistral-medium-3.5 | 112 | 98 | kimi-k3-max | 17 | 37 |
| 2026-Q4 | gemini-4-argon-high (google) | mistral-large-4 | 115 | 96 | kimi-k3-max | 16 | 37 |
The gap stays in a band (45–127 points) while the rank falls, probably because more models now crowd the scores just below the top (our reading).
Download CSV ↓Source: Arena leaderboard dataset (CC BY 4.0) · as of 2026-10-08 · last snapshot in each quarter · Compiled by StackPath
On usage, European models are 0.26% of routed tokens
Share of tokens (prompt plus completion) processed through OpenRouter, by the bloc of each model's developer. Measured
646 trillion tokens across 695 models, data to 11 Oct 2026. Undisclosed = stealth models listed without a developer; Unclassified = small developers we have not mapped.
Counting Poolside as European (it has a Paris office; Bloomberg gives its HQ as San Francisco) would raise Europe's 30-day share to 1.1%.
Source: OpenRouter rankings (CC BY 4.0) · as of 2026-10-12 · OpenRouter data licensed under CC BY 4.0; bloc mapping ours · Compiled by StackPath
Chinese models took over the weekly top 10 in the past year
Share of all OpenRouter tokens each week, by the bloc of the models in that week's top 10; the rest are outside the top 10. Measured
Source: OpenRouter rankings (CC BY 4.0) · as of 2026-10-12 · weekly top-10 chart · Compiled by StackPath
Most-used models, last 30 days
Top 20 models on OpenRouter by tokens; European models listed below with their rank. Measured
| # | Model | Developer | Bloc | Tokens (trillion) |
|---|---|---|---|---|
| 1 | deepseek-v4.1-flash-20260910 | deepseek | China | 104.65 |
| 2 | space-bunny-alpha | stealth | Undisclosed | 56.72 |
| 3 | glm-5.3-flash-20260826 | z-ai | China | 55.11 |
| 4 | hy4-preview-20260827 | tencent | China | 39.82 |
| 5 | gpt-5.6-luna-20260709 | openai | US | 32.26 |
| 6 | deepseek-v4-flash-20260731 | deepseek | China | 30.96 |
| 7 | mimo-v2.6-flash-20260921 | xiaomi | China | 26.59 |
| 8 | nemotron-3-ultra-550b-a55b-20260604 | nvidia | US | 23.75 |
| 9 | gpt-6-luna-20260922 | openai | US | 16.12 |
| 10 | deepseek-v4-flash-20260423 | deepseek | China | 14.45 |
| 11 | mimo-v2.5-20260422 | xiaomi | China | 13.48 |
| 12 | glm-5.3-20260816 | z-ai | China | 13.04 |
| 13 | hy3-20260706 | tencent | China | 12.87 |
| 14 | jev-1.13-20260917 | typesafe | Unclassified | 9.5 |
| 15 | gemini-3.8-flash-20260902 | US | 9.01 | |
| 16 | step-5-preview-20261008 | stepfun | China | 8.84 |
| 17 | claude-opus-5.5-20260921 | anthropic | US | 8.56 |
| 18 | muse-spark-1.3-contributor-20260902 | meta | US | 8.08 |
| 19 | gpt-5.6-sol-20260709 | openai | US | 6.96 |
| 20 | kimi-k3-20260715 | moonshotai | China | 6.95 |
European models in the 30-day list: mistral-nemo (#61, 1.17T), mistral-small-3.2-24b-instruct-2506 (#145, 0.11T), mistral-large-4-0-20261006 (#148, 0.1T), mistral-small-24b-instruct-2501 (#184, 0.05T), mistral-small-2603 (#194, 0.04T), mistral-embed-2312 (#205, 0.03T), ministral-8b-2512 (#212, 0.02T), mistral-medium-3.5-20260430 (#215, 0.02T).
Download all developers (CSV) ↓Source: OpenRouter rankings (CC BY 4.0) · as of 2026-10-12 · Compiled by StackPath
Artificial Analysis: Mistral Large 4 scores 38, the leader 57.6
Intelligence Index v4.3.2. Headline scores only, with links; we don't republish their table. Reported
| Model | Developer | Bloc | Score | Note |
|---|---|---|---|---|
| Mistral Large 4 (preview) | Mistral AI | Europe | 38 | Ranked 64th of 225 models compared; weights due end-October 2026, so listed as proprietary for now. source (6 Oct 2026) |
| Claude Opus 5.5 (Max) | Anthropic | US | 57.6 | Highest Intelligence Index score among the models shown on OpenRouter's benchmark view, which republishes Artificial Analysis scores; models not on OpenRouter are not in that view. source (11 Oct 2026) |
| Qwen3.8 Max | Alibaba (Qwen) | China | 53.4 | Highest-scoring Chinese model in the same OpenRouter benchmark view. source (11 Oct 2026) |
| Mistral Large 3 | Mistral AI | Europe | 9.3 | Mistral's previous flagship, same view, for scale of the jump. source (11 Oct 2026) |
Source: Artificial Analysis, Mistral Large 4 article · OpenRouter benchmark view (Artificial Analysis scores) · as of 2026-10-11 · Compiled by StackPath
25 models above 10^25 FLOP in Epoch's data, 0 from Europe
Models whose published training-compute estimate is at least 10^25 FLOP (the EU AI Act's systemic-risk threshold), by developer's country. Measured
| Notable models by year | US | China | Europe | Other |
|---|---|---|---|---|
| 2019 | 56 | 6 | 22 | 6 |
| 2020 | 37 | 8 | 13 | 5 |
| 2021 | 54 | 21 | 24 | 8 |
| 2022 | 68 | 12 | 17 | 12 |
| 2023 | 86 | 24 | 17 | 18 |
| 2024 | 57 | 32 | 8 | 9 |
| 2025 | 64 | 35 | 2 | 10 |
| 2026 | 49 | 29 | 0 | 6 |
A model counts once for each bloc of its developers (joint US–European work counts for both). The current year is partial.
Download CSV ↓Source: Epoch AI, AI models database (CC BY 4.0) · as of 2026-10-12 · Compiled by StackPath
Epoch has no compute estimate for many recent frontier models, Mistral's included, so this is a floor. Interface counted 36 models above 10^25 FLOP by Oct 2025, two of them Mistral's (Interface, 22 Oct 2025; not re-checked). The largest European model with a published estimate in Epoch's data: Apertus 70B (ETH Zurich,Ecole Polytechnique F´ed´erale de Lausanne (EPFL),Swiss National Supercomputing Centre (CSCS),Swisscom, 6.7×10^24 FLOP).
Our reading. Europe's position in models is narrow: one lab, Mistral, accounts for almost all European presence on these boards, and its open-weight releases are the main route to usage. Chinese developers now take most routed tokens, and Europe's most-used model on OpenRouter is an older, small open model (Mistral NeMo) rather than its flagship, which suggests price per token matters at least as much as rank. None of these measures covers first-party apps (ChatGPT, Gemini, Claude, Le Chat), where most consumer use happens.
OpenRouter is one router: it over-represents developers and price-sensitive API traffic, and excludes first-party apps (ChatGPT, Gemini, Claude.ai, Le Chat) and direct API use. Token share is not market share, revenue share or user share. Tokens count prompt plus completion. Cheap models used in agent loops pile up tokens fast, which favours low-price models. OpenRouter's public models view lists the models active in the period (about 600–700). Authors are mapped to a bloc by HQ of the developing company (parent for subsidiaries); stealth models with no named developer are shown as Undisclosed. Arena scores are human-preference votes (Bradley–Terry ratings with style control). Ranks with overlapping confidence intervals are statistical ties; ratings are not comparable across snapshots, so the history compares gaps within each snapshot. Epoch counts only models with a published training-compute estimate; many frontier models (including recent Gemini, Claude and Mistral flagships) have none, so the 10^25 FLOP count is a floor. Artificial Analysis scores are headline figures only (its redistribution terms are unclear).
Bloc mapping: by the HQ of the developer's parent company. Google DeepMind (London) counts as US, the same as Microsoft AI's London hub. Mapping table: developer → bloc (CSV) ↓
Questions people ask
- What is the best European AI model?
- On Arena's text leaderboard (8 Oct 2026) the highest-ranked European model is mistral-large-4 from Mistral AI, at #115. Artificial Analysis scored the Mistral Large 4 preview 38 on its Intelligence Index (6 Oct 2026), the highest for a model from outside the US and China.
- How much AI usage goes to European models?
- On OpenRouter, 0.26% of tokens in the last 30 days went to models from European developers, against 58.8% for Chinese and 29.2% for US developers. OpenRouter is one router and excludes first-party apps.
- Are Chinese AI models ahead of European ones?
- On these measures, yes. The best Chinese model on Arena ranks #16 against #115 for the best European, and Chinese developers took 58.8% of OpenRouter tokens in 30 days.
- Does StackPath run its own benchmarks?
- No. We combine open data from OpenRouter, Arena and Epoch AI (all CC BY 4.0) and cite Artificial Analysis headline scores with a link. The bloc mapping (which developer counts as European) is ours and downloadable.
- Why is Google DeepMind not counted as European?
- We map every developer by the HQ of its parent company. Google DeepMind is based in London but owned by Alphabet, so it counts as US, the same rule we apply to Microsoft AI's London hub and Meta's Paris lab.