DeepSeek Harness Hits 1.54 Million Daily Users and 92 Trillion Tokens a Day

DeepSeek said on X on Sunday, 8 November 2026, that its open-source agent app Harness had reached 1.54 million daily active users on 4 November, less than three months after version 0.1 launched on 13 August. Those users are consuming about 92 trillion tokens a day through DeepSeek's official service, the company said, in figures reported the same day by Bloomberg. The announcement is one day old at the time of writing, and it is the clearest public signal yet that the race in AI has moved from chatbots to agents.

A harness is the software layer that turns a language model into an agent that can run code, call tools and carry out multi-step tasks with limited human supervision, rather than simply answering a question. DeepSeek Harness is that layer, built as an open-source app, and its growth in the last five weeks has been steep. At the end of September the figure was about 271,000 daily users. By 4 November it was 1.54 million, a rise of about 5.7 times in just over a month.

What DeepSeek announced

The numbers were first shared by the Harness team in an update on 5 November, reported by Chinese outlet Yicai, and DeepSeek repeated them on X on Sunday, 8 November. Bloomberg, the South China Morning Post and the Business Times covered the milestone over the weekend and on Monday. The headline figures are simple, but each one carries a detail that matters.

First, the user count. Harness reached 1.54 million daily active users on 4 November, less than three months after its debut. The desktop app arrived on 29 September, and the growth since then has been rapid. The seven-day average was 271,000 on 30 September and 429,000 on 7 October, so the curve bent sharply upward in the second half of October.

Second, the token count. Harness users are consuming about 92 trillion tokens a day, which works out at about 60 million tokens per active user per day, and about 1.1 billion tokens per second across the whole service. That is far more than a typical chatbot conversation uses, because agents keep working for hours on long-running tasks, reading files, running code and checking their own work. Bloomberg described the figure as among the first public evidence of how steeply agents escalate demand for processing power.

Third, the plugin ecosystem. The team says about 35,000 third-party plugins have been built for Harness, and about 60 percent of active users have built or installed at least one plugin. DeepSeek plans an official plugin marketplace and a support programme for plugin authors, but it has announced no launch date for either.

A laptop showing a code editor on a wooden desk by a window, with a potted succulent plant and a keyboard in the foreground

Illustration: A laptop showing a code editor on a wooden desk by a window, with a potted succulent plant and a keyboard in the foreground

Why the growth matters

DeepSeek, the Hangzhou-based startup that jolted Silicon Valley with its low-cost reasoning model R1 last year, is staking a claim to the next frontier of the industry: infrastructure for AI agents, software that lets AI act on a user's behalf. Harness is the vehicle for that claim. It handles tasks such as coding, file analysis and research with limited human supervision, and it is designed to work with external software through plugins.

The competition is direct. OpenAI's Codex and Anthropic's Claude Code are the best-known agent tools in the American market, and Harness is positioned as the open, developer-led alternative. That positioning matters because the agent layer is where users build habits. A developer who writes plugins for one harness, and who trusts it with real tasks, is less likely to switch to another. DeepSeek's bet is that an open ecosystem with tens of thousands of community plugins will grow faster than a closed product, and the last five weeks of numbers support that bet.

The growth in numbers

DateDaily active usersNotes
13 August 2026Not disclosedHarness version 0.1 launches
29 September 2026About 271,000Desktop app released
30 September 2026271,000Seven-day average
7 October 2026429,000Seven-day average
4 November 20261,540,000Daily active users, 92 trillion tokens a day
A modern glass office building with people walking in the plaza in front of it

Illustration: A modern glass office building with people walking in the plaza in front of it


What the numbers do not show

These are DeepSeek's own figures, and they have not been independently verified. No outside audit has been cited in any of the reporting, and DeepSeek has not published a methodology for how it counts a daily active user. That does not make the numbers false, but it means they should be read as a company disclosure, not as an independently established fact.

The count is also partial by design. DeepSeek's figure covers only people using Harness through its official application programming interface, which is to say people connecting to DeepSeek's own servers. It does not count anyone who runs Harness with other models, or who self-hosts the open-source code on their own machines. The real number of Harness users could therefore be higher than 1.54 million, or the official service could be growing while self-hosted use stays flat. The published figure cannot tell us which is true.

The plugin numbers carry a similar caveat. DeepSeek says about 35,000 third-party plugins have been built, and that 60 percent of users have created at least one, but the complexity of those plugins varies widely. A plugin that reformats a file and a plugin that runs a multi-step research workflow are both counted as one plugin. The count shows a busy community, not a measure of quality.

No new capabilities, pricing or licence terms were announced with the milestone. Harness is the same open-source app it was in August, with a desktop version added in September and a plugin marketplace still to come. Readers should not treat this announcement as a product launch. It is a growth update, and it should be read as one.

How Harness compares

The agent race is now a contest between different ideas about who controls the tools. Anthropic's approach is a polished product with deep integration into its own models, and our report on the launch of Claude Duet shows how that strategy works in practice. DeepSeek's approach is the opposite: an open app, a community plugin ecosystem, and a model that other developers can run themselves.

The Chinese model landscape is moving fast around Harness. Our coverage of Kimi K3.1 Pro showed a 28 trillion parameter model built for long, demanding tasks, and our report on Google Gemini 4 Carbon showed the American response at the frontier. Harness sits between those two stories. It is not a model, but the software that lets a model do useful work, and its growth suggests that the agent layer may matter as much as the model underneath it.

A data center corridor lined with rows of server racks stretching into the distance

Illustration: A data center corridor lined with rows of server racks stretching into the distance


What to watch next

Three things will show whether this growth lasts. The first is the plugin marketplace. If DeepSeek ships it, and if developers can distribute and earn from plugins through it, the ecosystem could grow faster still. If it stays a promise, the 35,000 plugins will remain a community effort without a commercial engine behind it.

Key Takeaways

  • DeepSeek said on 8 November 2026 that Harness reached 1.54 million daily active users on 4 November, less than three months after version 0.1 launched on 13 August.
  • Those users consume about 92 trillion tokens a day through DeepSeek's official service, or about 60 million tokens per user per day.
  • Daily users rose from about 271,000 at the end of September to 1.54 million on 4 November, a rise of about 5.7 times in just over a month.
  • About 35,000 third-party plugins have been built, and about 60 percent of users have built or installed at least one. A plugin marketplace is planned, with no launch date announced.
  • The figures are DeepSeek's own disclosures and have not been independently verified, and they cover only users of the official service, not self-hosted open-source runs.
  • No new capabilities, pricing or licence terms were announced with the milestone.

Conclusion

Harness has grown faster than almost any AI tool of the last year, and the token numbers show that agents use computing power on a scale chatbots never did. That is a real signal about where the industry is heading. But the figures are self-reported, partial and unaudited, so the right response is interest, not certainty. Watch the plugin marketplace, watch retention, and watch for verification. If the growth holds, Harness will be one of the defining products of the agent era. If it does not, this will be remembered as a strong first quarter, not a turning point.

Sources

 

Comments

Post a Comment