AI news
Chinese Models Now Take 7 of OpenRouter's Top 10 Traffic Slots — and a Fraction of the Spend
Short answer
- We pulled OpenRouter's weekly token leaderboard today (Nov 6, 2026). Seven of the top ten slots belong to Chinese labs: DeepSeek V4.1 Flash 1017 (#1), DeepSeek V4.1 Flash (#2), GLM 5.3 Flash (#3), Step 5 Preview (#5), Hy4 preview (#6), MiniMax M3.1 Flash (#8), Hy4 (#10). The three outsiders: GPT-6 Luna (#4), Claude Opus 5.5 (#7), GPT-5.6 Luna (#9).
- Token traffic is not money. Per Goldman Sachs' analysis of OpenRouter data, open-weight models carried 72% of platform traffic but only 27% of spend as of September; CICC research counts Chinese models at 93.54T weekly tokens — roughly half of the platform's 188T total for the Oct 19–25 week, and 26 consecutive weeks at #1.
- What this means if you're buying: the commodity layer of the API market (cheap, high-volume tokens) is now decisively Chinese, while premium spend still concentrates on Anthropic and OpenAI. If your workload is price-sensitive, you're almost certainly going to end up on one of the models above — and we track their prices.
The leaderboard is the bluntest instrument OpenRouter publishes: pure token volume over the most recent complete week, no quality weighting, no price filtering. That's exactly why it's worth staring at. Here is the full top ten as we pulled it on November 6:
Who's actually on the list
DeepSeek holding ranks 1 and 2 simultaneously is the detail that tells you how this market works: V4.1 Flash and its 1017 point-release are two entries of essentially the same model family, and together they out-token everything else on the platform. Add GLM 5.3 Flash at #3 — Zhipu's budget flagship, which CICC's research note counts at 15T weekly tokens on its own — and the top three slots are all sub-dollar Chinese models.
Tencent's Hy4 appears twice too (#6 preview, #10 stable), which tracks with what we found when the stable release landed: the API is priced at roughly half the preview tier, and the traffic followed the discount. Step 5 Preview (#5) and MiniMax M3.1 Flash (#8) round out the Chinese presence.
The three US entries are instructive: GPT-6 Luna and GPT-5.6 Luna are OpenAI's budget tiers, and Claude Opus 5.5 — Anthropic's most expensive model — sits at #7 on sheer volume. Premium models can still chart on traffic; they just can't take the top of it anymore.
Traffic is not revenue
The second half of the story is the one buyers should internalize. Goldman Sachs' analysis of OpenRouter data (September) put open-weight models at 72% of platform traffic against 27% of platform spend. The dollar flow still points at closed premium models, even as the token flow points the other way. CICC's numbers (Oct 19–25 week) show Chinese models specifically at 93.54T of the platform's 188T weekly tokens — 26 straight weeks as the largest country group.
Read those together and the picture is a two-layer market: a Chinese-dominated commodity layer where tokens are nearly free and volume is enormous, and a Western premium layer where per-token prices are 10–50× higher and the actual money lives. Neither layer is "winning" — they're different products.
What we'd do with this
If you're routing volume work — classification, extraction, summarization, bulk chat — the leaderboard is your shortlist, and the pricing differences between these models are larger than the quality differences for most such tasks. If you're paying premium rates for work that a #1-ranked sub-dollar model does adequately, you're donating margin. Flip side: the premium tier still exists because for hard, high-stakes work, Claude Opus 5.5 and GPT-6 Luna earn their price. Know which layer your workload lives in before you shop.
We verify prices on these models continuously — recent catches: Hy4's stable tier at exactly half the preview price, and the K3.1 Pro weight release that put a 46.5-score model into open weights.
Verified first-hand on 2026-11-06: the complete top-10 leaderboard ("Top LLMs by weekly token usage on OpenRouter") scraped from openrouter.ai/rankings, and all model availability via the OpenRouter models API. The 72%/27% traffic-spend split is Goldman Sachs' analysis of OpenRouter data as reported by Chinese tech media; the 93.54T/188T token counts and 26-week streak are CICC research figures — neither reproduced by us, and both predate this week's leaderboard pull.
Sources: OpenRouter Rankings (scraped 2026-11-06), OpenRouter models API (checked 2026-11-06), Goldman Sachs analysis and CICC research as reported by moneymoneyhome — GLM-5.3 Flash 15T weekly tokens, Chinese models 26 weeks #1