Fig 1 · Weekly token share, top-9 vendors. Remaining vendors roll into OpenRouter's “Others” bucket and are shown when they enter top-9.
OpenRouter's rankings page ships a full year of weekly token data per vendor. We analysed it. While benchmark leaderboards argue about Claude, GPT-5 and Gemini, the actual inference market has shifted somewhere else entirely — Chinese open-weight labs went from a rounding error to more than half of the tracked flow in twelve months, and the total market is up NaN×.
Stacked share of total OpenRouter top-9 token volume. The “how many tokens” view — it undercounts premium vendors whose tokens cost 10–30× more. The dollar view is § 02.
Same weeks, same vendors — each vendor's token count is multiplied by its current average blended $/M, computed from the live model catalog weighted by real usage. The “how much money” view. Anthropic's token share is 0.0% but its dollar share is 0.0% — that gap is the whole premium-lane thesis.
Last 4 weeks. Positive gap = premium pricing, below-weight on tokens. Negative gap = volume play, below-weight on dollars.
| Vendor | Tokens | Dollars | Gap (pts) | Avg $/M |
|---|
Combined share of .
First 4 weeks of the series vs the last 4. Growth ratio is absolute tokens, not share.
| Vendor | Year-ago | Now | Δ pts | Past tok | Now tok | Growth |
|---|
The short answer is no, and the longer answer has three parts.
1 · The market is stratifying, not choosing. Claude Opus 4.6 is the #1 model on our inverted leaderboard by dollar spend — $25M/month across 24 apps — because it is priced at $5/$25 per million. But it is only ~4th by token volume (2.4T), and Anthropic as a vendor is ~12% of total tokens, down from ~25% last year. Both things are true at once: the premium lane still pays Anthropic real money, and the commodity lane has moved decisively elsewhere.
2 · It is following price, not quality. The vendors gaining share (Qwen, Xiaomi, MiniMax, DeepSeek, Z.ai, StepFun) share one property: aggressive open-weight pricing, often sub-$1/M blended. On the inverted model leaderboard, Qwen3.6 Plus is in 27 of 30 apps and MiMo-V2-Pro handles 5.5T tokens at ~$1.50/M blended — a tiny fraction of what a comparable Claude Opus run would cost.
3 · Benchmark-to-spend correlation is weak, and inverted at the top. Among models with both token volume and a published benchmark score, the relationship between benchmark rank and market share is closer to anti-correlated. Premium models capture most of the dollar spend because they're priced 10–30× higher; they do not capture tokens. Agents — when given a free choice through a router — route most tokens to “good enough and 20× cheaper” rather than “best and 20× more expensive”.
This cuts two ways. It says benchmark leaderboards overstate real-world adoption for frontier models. It also says the market may be under-weighting quality in agentic workflows where a wrong answer is expensive — we do not yet have the data to distinguish “efficient routers” from “cheap routers paid for by downstream users”.
openrouter.ai/rankings RSC stream. 53 weeks of vendor-level tokens, 77 weeks of model-level. Captured 2026-06-23.Seeing a model, vendor, or app shift that isn't reflected here? Tell us — we reply within 48 hours and update the analysis.