DeepSeek’s workhorse model now moves more tokens through OpenRouter than the leading models from OpenAI and Anthropic put together. In the week to November 7, DeepSeek’s V4.1 Flash 1017 processed 56.2 trillion tokens on the routing platform, about 4.1 times the combined volume of OpenAI’s GPT-6 Luna and Anthropic’s Claude Opus 5.5. DeepSeek’s models took 56.4 percent of all tokens handled by OpenRouter’s ten most-used models, up from 52 percent the week before, while the two American labs’ combined share slipped. The numbers cover one routing platform, not the whole market, but they show where developers are actually sending their work.
The numbers behind the lead
The headline figure belongs to the 1017 snapshot. DeepSeek V4.1 Flash 1017 processed 56.2 trillion tokens in the week to November 7, up 23 percent on the week before, and it leads the rankings by a wide margin on OpenRouter’s public leaderboard.
Its stablemate ranked second. DeepSeek V4.1 Flash processed 17.2 trillion tokens, up 13 percent, and together the two DeepSeek models handled more than half of all tokens in the top ten. The ten most-used models processed 130 trillion tokens in the week, up from 124.2 trillion the week before.
The gap to the American frontier models is wide. OpenAI’s GPT-6 Luna ranked fourth with 7.22 trillion tokens, down 34 percent week over week, while Anthropic’s Claude Opus 5.5 ranked seventh with 6.59 trillion, down 10 percent. The two labs’ combined share of top-ten tokens slipped from 14.7 percent to 13.7 percent in a single week.
Chinese labs hold most of the rest. Z.ai’s GLM 5.3 Flash ranked third with 14.1 trillion tokens, down 13 percent, taking Z.ai’s share from 16.7 percent to 10.8 percent. Tencent’s newly released Hy4 model debuted in fifth place with 7.05 trillion tokens, lifting the combined share of Tencent, StepFun and MiniMax from 16.6 percent to 19 percent.
Why the cheap model keeps winning
The 1017 model is not a new launch. It is the October 17 snapshot of DeepSeek’s V4.1 Flash, and DeepSeek’s own documentation says the deepseek-flash API name is now served by it. Developers who were already calling the Flash model were switched to the newer snapshot without changing their code, which helps explain why its volume jumped so fast.
Price is the other half of the story. Flash is DeepSeek’s budget tier, built for high-volume work, and it costs a fraction of what frontier models charge per token. Agent workloads, which fire off many small calls, favour exactly that combination: capable reasoning at a price where heavy use stays affordable.
DeepSeek’s push into agent tooling points the same way. Its Harness platform, now in developer preview, gives agent builders a managed way to run tools on DeepSeek models. The company has also positioned its autonomous coding agent as a flagship product, and both moves assume developers will route large volumes through cheap, fast models.
The rankings still need reading with care. They count tokens routed through OpenRouter only, leaving out direct API traffic to OpenAI, Anthropic, Google and DeepSeek itself. Token counts are not revenue: a cheap model can process far more tokens while earning far less.
OpenRouter’s own methodology notes add a second caution. Models differ in verbosity and tokenisation, so a higher token total shows how much a model is used, not how good it is. The rankings also exclude requests that users keep private, and they group model variants separately.
Our Read
This is a distribution story, not a capability story. Nothing in the rankings says Flash 1017 reasons better than GPT-6 Luna or Claude Opus 5.5; it says developers route far more work through it. That distinction matters, because the prize here is the default model for everyday agent work.
Defaults are set by price, speed and availability as much as by benchmark scores, and DeepSeek wins on all three for high-volume use. The open question is whether the lead survives contact with enterprise buyers, who weigh safety and compliance alongside cost. A model can dominate a router and still be kept out of regulated work.
The result also sharpens a tension inside DeepSeek’s own week. Independent testing keeps finding rough edges: a recent evaluation scored DeepSeek V4.1 Flash at 16 percent on prompts about politically sensitive topics, far above Western models. Heavy use and heavy scrutiny are arriving together, and the second will shape whether enterprises trust the first.
Why It Matters
Routing platforms have become the closest thing the industry has to a public scoreboard for real usage, and DeepSeek now dominates it. If developers keep choosing the cheapest capable model for agent work, frontier labs face pressure to cut prices or prove their premium is worth paying.
The commercial backdrop is moving just as fast. DeepSeek closed a $16 billion second funding round in late October, giving it the capital to keep prices low while it scales. The policy backdrop is narrower than the headlines suggest: Washington’s order removes DeepSeek from federal civilian systems but leaves private-sector and research use untouched.
Background
DeepSeek V4.1 Flash first appeared on OpenRouter on September 10, and the 1017 snapshot followed on October 17. The same family powers DeepSeek’s autonomous coding agent, which the company has positioned as its flagship agentic product. OpenRouter’s rankings are published under a CC BY 4.0 licence, and its data covers tokens processed through its API only.
Source: OpenRouter rankings · FourWeekMBA analysis
0 Comments