DeepSeek's open-source agent runtime, Harness, reported 1.54 million daily active users as of November 4—and about 92 trillion tokens consumed per day through the company's official API. That works out to roughly 60 million tokens per active user, a rare hard number for how much compute agent workloads burn compared with chatbots.
The figures come from DeepSeek Harness's own November 8 X article, "Growing Together with Our Plugin Community," which restates product metrics earlier covered in Chinese tech media. Bloomberg's English write-up, carried by outlets including Free Malaysia Today, framed the same disclosure as evidence that agents are becoming a daily-use product category and that DeepSeek is building an open plugin ecosystem against OpenAI and Anthropic. The numbers are company-reported and cover official-API users only.
What DeepSeek disclosed
According to the Harness team:
- Launch: Harness v0.1 went live August 13, 2026; the desktop app followed September 29.
- Adoption: By November 4—less than three months after launch—daily active users on the official API reached 1.54 million, up from about 271,000 around the end of September in secondary reporting of DeepSeek's earlier tallies.
- Compute: Those users were consuming roughly 92 trillion tokens per day.
- Plugins: About 35,000 third-party plugins exist; roughly 60% of active users have built or installed at least one.
- Roadmap: DeepSeek plans an official plugin marketplace plus a creator-support program with token grants, technical discussions, and promotional placement. No launch date was given.
Dividing 92 trillion by 1.54 million yields about 59.7 million tokens per daily active user—consistent with DeepSeek's and Bloomberg's "about 60 million" framing. The metric measures token throughput through DeepSeek's hosted API for Harness users, not completed tasks, revenue, or the full population of people running Harness against other model endpoints.
Why ~60 million tokens per user matters
Chatbots typically spend a few thousand to tens of thousands of tokens on a session. Agents that edit files, call tools, browse, spawn subagents, and loop for minutes or hours can burn orders of magnitude more. DeepSeek's aggregate—about 1.1 billion tokens per second across the counted user base, if the daily total is spread evenly—is among the clearest public signals yet that agent products scale inference demand differently from conversational assistants.
That appetite has direct cost consequences. DeepSeek raised API prices in August 2026 when it launched Harness alongside V4-Pro, introducing peak and off-peak rates that made many workloads substantially more expensive than the prior flat schedule. High per-user token counts amplify those price moves. They also stress China's constrained access to leading-edge GPUs: every multi-hour agent loop is an inference bill and a scarce-accelerator bill at once. Efficiency at the model and harness layer is not a marketing theme here; it is a survival constraint.
Open ecosystem versus closed agent products
Harness is not a chatbot skin. DeepSeek describes it as an MIT-licensed agent harness built on Cordis with an "everything is a plugin" architecture—models, tools, skills, sessions, sandboxes, and UI elements are meant to be swappable. The August launch positioned it as open infrastructure for agent workflows, overlapping territory occupied by Anthropic's Claude Code and OpenAI's Codex without being a feature-for-feature clone of either commercial product.
The November disclosure doubles down on ecosystem strategy. Thirty-five thousand third-party plugins and a planned official marketplace are DeepSeek's bid to compete on breadth of applications, not only on model scores or token price. The company explicitly says it wants Harness to serve as "the operating system of the agent era," with built-in features replaceable by community plugins. That is a different bet from a vertically integrated coding agent locked to one lab's model.
Readers should not confuse DeepSeek Harness with Harness.io, the unrelated software-delivery company. The product here is DeepSeek's agent runtime.
Caveats the headline can hide
Several limits belong next to the 1.54 million figure:
- Self-reported. DeepSeek published the metrics; no independent audit is public.
- Official-API scope. Users running Harness against non-DeepSeek models or self-hosted endpoints are outside the count—so the figure can understate total software adoption while overstating DeepSeek-hosted engagement if readers treat it as a global install base.
- Plugin quality varies. DeepSeek itself says community plugins "vary in maturity." A large plugin catalog is not the same as a large set of production-grade tools.
- Measurement detail. Secondary English coverage has described earlier September figures as multi-day averages in some accounts and the 1.54 million print as a November 4 snapshot in others. DeepSeek's X article states the November 4 daily-active number without publishing the full averaging methodology.
What to watch next
Three follow-ups will show whether Harness is a durable platform or a spike. First, whether the official marketplace launches with real discovery, security review, and payment or grant mechanics—or stays a promise. Second, whether token consumption per user stays elevated as workflows mature, or falls as caching, smaller models, and tighter tool loops improve. Third, whether Western labs answer with more open harnesses of their own or keep pushing closed, integrated agent products.
For now, the usable fact is simpler. On DeepSeek's own books, a young open agent app is already driving chatbot-dwarfing token volumes for more than a million daily official-API users. That is early evidence—not proof—that the industry's shift from chat to agents is a compute story as much as a product story.
Primary source: DeepSeek Harness, "Growing Together with Our Plugin Community" (X article, November 8, 2026). Corroborating coverage: IT Home, TechNode, Bloomberg via Free Malaysia Today, and VentureBeat's August 13 launch report.