From Benchmark Parity to Usage Dominance: Chinese Models Took 4 of Every 5 Tokens on OpenRouter Last Week — and Have Led US Models for 24 Straight Weeks
Original (Oct 5, 2026 — Bloomberg Intelligence analyst Robert Lea): the capability gap that anchors US export-control policy has collapsed to a rounding error — DeepSeek V4.1 Flash matches Anthropic's frontier on LiveBench, a ~3% gap.
Update (Oct 12, 2026): the deployment metric is no longer a race — it's a rout. Citing OpenRouter, the global LLM aggregation platform, the state-run Global Times reports that for the week of October 5–11, "AI models worldwide processed 172 trillion tokens... Chinese models accounted for 86.74 trillion tokens, compared with 18.97 trillion for US AI models" — roughly 4.6x the US figure — with Chinese usage up 50.96% week-on-week against 17.1% for US models. "It was the 24th consecutive week that Chinese models led their US counterparts in token usage."
The top five models by weekly tokens were all Chinese:
- DeepSeek V4.1 Flash — 41T tokens, +60% WoW (and this after DeepSeek "significantly raised API prices for its latest flagship models in mid-August," per Fast Think Institute president Tian Feng — growth despite price hikes)
- Z.ai GLM 5.3 Flash — 12.2T, +26%
- Xiaomi MiMo-V2.6-Flash — 11.5T, +19%
- StepFun Step 5 Preview — 8.84T
- Tencent Hy4 preview — 8.28T, +27%
Together those five models processed 47.9% of all tokens on the platform. The base effect is the real headline: per CNBC (cited in the same report), Chinese models accounted for just 6–13% of OpenRouter usage in February — the shift to near-total dominance of the open-routing market happened inside eight months.
Caveats worth keeping in mind: the source is Global Times (Chinese state media) citing OpenRouter data, and OpenRouter skews toward price-sensitive, API-driven developers — it is not the whole market. But the direction is unambiguous, and it converts From Benchmark Parity to Usage Dominance: Chinese Models Took 4 of Every 5 Tokens on OpenRouter Last Week — and Have Led US Models for 24 Straight Weeks from a benchmark story into a market-share story: the 3% capability gap is now compounded by a cost-of-inference gap that shows up as 86.74T vs 18.97T tokens. This is also the demand backdrop against which the Nvidia–Reflection talks (Reflection AI Unveils Beam: The "DeepSeek of the West" Bets on Efficiency Over Raw Capability — and Nvidia Is Now in Talks to Buy It) make sense — Washington's answer to cheap Chinese open weights is a US open-weight champion, and even Intel's CEO was quoted this week calling Chinese frontier-model adoption "very fast" (Reuters, Oct 12 — seen in headlines only).