Research  /  AI model adoption
Live Refreshed 27 Aug 2026, 19:50 CET · data as of 26 August 2026

AI model adoption

Data as of 26 August 2026. Measured over 2026-08-20 to 2026-08-26. Source: OpenRouter (https://openrouter.ai/rankings).

Regional share of routed tokens, this week
Where the top 30 models were built
Model origin, not user location.

Base image: NASA Earth Observatory, Earth at Night 2012. Public domain. Arcs are illustrative, not measured routes.

Share of routed tokens

Regional share of routed tokens, by where the model was built.

  • Asia: 43.95% of routed tokens (-12.65pp week on week), 13 of the top 30 models. Led by DeepSeek V4 Flash, ranked 2 of 60 models measured.
  • Europe: 0.14% of routed tokens (-0.12pp week on week), no model in the published top 30. Led by Mistral Nemo, ranked 50 of 60 models measured.
  • US: 25.96% of routed tokens (-10.49pp week on week), 16 of the top 30 models. Led by NVIDIA Nemotron 3 Ultra 550b A55b (free), ranked 6 of 60 models measured.
  • Other: 29.95% of routed tokens (+23.26pp week on week), 1 of the top 30 models. Led by Stealth Ox Alpha, ranked 1 of 60 models measured.
Most-routed models

Most-routed models, 26 August 2026.

  1. 1. Stealth Ox Alpha (stealth, Other) - 24.58% of routed tokens.
  2. 2. DeepSeek V4 Flash (31 Jul 2026) (DeepSeek, Asia) - 11.14% of routed tokens.
  3. 3. Xiaomi Mimo V2.5 (Xiaomi, Asia) - 9.3% of routed tokens.
  4. 4. Tencent Hy3 (Tencent, Asia) - 6.3% of routed tokens.
  5. 5. DeepSeek V4 Flash (23 Apr 2026) (DeepSeek, Asia) - 5.03% of routed tokens.
  6. 6. NVIDIA Nemotron 3 Ultra 550b A55b (free) (NVIDIA, US) - 4.91% of routed tokens.
  7. 7. OpenAI GPT 5.6 Luna (OpenAI, US) - 3.7% of routed tokens.
  8. 8. Google Gemini 3.7 Flash (Google, US) - 3.14% of routed tokens.
  9. 9. Z.ai GLM 5.2 (Z.ai, Asia) - 3.02% of routed tokens.
  10. 10. DeepSeek V4 Pro (23 Apr 2026) (DeepSeek, Asia) - 1.65% of routed tokens.
  11. 11. OpenAI GPT 5.6 Sol (OpenAI, US) - 1.4% of routed tokens.
  12. 12. MiniMax M3 (MiniMax, Asia) - 1.35% of routed tokens.
  13. 13. Anthropic Claude Opus 5 (Anthropic, US) - 1.34% of routed tokens.
  14. 14. Moonshot Kimi K3 (Moonshot, Asia) - 1.31% of routed tokens.
  15. 15. Poolside Laguna S 2.1 (free) (Poolside, US) - 1.31% of routed tokens.
  16. 16. Anthropic Claude Sonnet 5 (Anthropic, US) - 1.02% of routed tokens.
  17. 17. NVIDIA Nemotron 3.5 Lightning (free) (NVIDIA, US) - 0.98% of routed tokens.
  18. 18. DeepSeek V4 Pro (13 Aug 2026) (DeepSeek, Asia) - 0.92% of routed tokens.
  19. 19. Z.ai GLM 5.3 (Z.ai, Asia) - 0.84% of routed tokens.
  20. 20. Google Gemini 3 Flash Preview (Google, US) - 0.72% of routed tokens.
  21. 21. Google Gemini 2.5 Flash Lite (Google, US) - 0.69% of routed tokens.
  22. 22. Anthropic Claude 4.6 Sonnet (Anthropic, US) - 0.63% of routed tokens.
  23. 23. OpenAI GPT 5.6 Terra (OpenAI, US) - 0.56% of routed tokens.
  24. 24. Upstage Solar Pro4 (upstage, Asia) - 0.52% of routed tokens.
  25. 25. Google Gemini 3.6 Flash (Google, US) - 0.51% of routed tokens.
  26. 26. OpenAI GPT OSS 120b (OpenAI, US) - 0.47% of routed tokens.
  27. 27. Xiaomi Mimo V2.5 Pro (Xiaomi, Asia) - 0.47% of routed tokens.
  28. 28. xAI Grok 4.6 (xAI, US) - 0.44% of routed tokens.
  29. 29. Google Gemini 3.1 Flash Lite (Google, US) - 0.41% of routed tokens.
  30. 30. DeepSeek V3.2 (DeepSeek, Asia) - 0.4% of routed tokens.
Record book

Media and press may use the data and graphics with clear attribution to yellow3.io.

AI model adoption in 30 seconds
What it is

A live read on where AI demand actually flows - by where the model was built (Asia, the US, Europe) - measured in real routed usage, not announcements or benchmarks.

Why it matters

It is the clearest public signal of who is winning the AI race, and exactly how far behind Europe sits. The lead moves week to week, and most leaders are reading last year's map.

So what

If you are deciding what to build on, who to depend on, or how exposed you are to one region's models, this is the map - not the press release.

This instrument tracks regional share of routed AI traffic and grades its own weekly call in public. Updated live.

yellow3 on the record
This week's call
Last week's call, graded

Methodology & sources

We take positions only where the evidence supports them - and grade them in public.

How we build research - the method →