大模型网摘OpenRouter token share, November snapshot: DeepSeek at 49.2%, open-weight cohort at 70.4%
According to the public OpenRouter ledger compiled by Dirac (data complete through November 5), the platform's daily token volume has reached 21.13T, with the open-weight cohort at 70.4% and proprietary models at 29.6%. A Hacker News thread on November 4 pushed the topic onto the front page.
The leaderboard: single-day share, November 5
Daily API token share by lab: DeepSeek 49.2% (10.39T), OpenAI 12.8% (2.69T), Zhipu z-ai 11.6% (2.44T), Anthropic 9.9% (2.08T), Google 4.7% (1.00T), MiniMax 3.6% (0.76T), Moonshot 3.6% (0.75T), Xiaomi 2.1% (0.44T), Meta 2.0% (0.42T), Mistral 0.4% (0.08T). The top four together account for nearly 85%, and the gap between the head and tail of the list spans more than two orders of magnitude.
Trend: open weights from 40% to 70%
The series starts on March 19: open weights held 40.67% that day, on a total of 2.184T daily tokens. By mid-June the open-weight share had climbed to around 60%; on November 5 it stood at 70.4%. Over roughly seven and a half months, daily volume grew about tenfold. Both cohorts grew in absolute terms; the share gain comes from a faster growth rate, not from one side shrinking.
The 53% headline versus the ledger's 49.2%
The Hacker News headline said DeepSeek alone accounted for 53%; the ledger's latest complete day (November 5) shows 49.2%. The two numbers are not necessarily in conflict: different statistics days and snapshot points introduce drift. When citing data like this, the ledger's complete-day basis is the safer reference.
Caveats
Three things to keep in mind. First, this is a single routing platform; it aggregates requests that flow through OpenRouter and does not represent the whole market. Second, shares are grouped by lab, so multiple models from one company are merged: Zhipu z-ai corresponds to the GLM family, Moonshot to Kimi. Third, token volume reflects usage, not revenue or compute: two providers with a tenfold difference in unit cost can post identical token shares while meaning very different things commercially.
What it means for model selection
For developers, the leaderboard works as a crowdsourced signal of what others are actually routing: open-weight models now carry about 70% of real traffic, and the Chinese labs together exceed 60%. Pair that signal with pricing and your own workload. See our DeepSeek Flash-1017 vs Kimi K3.1 Pro comparison, the current model lineup overview, and, for cost-sensitive batch jobs, the off-peak pricing guide.