DeepSeek Harness reaches 1.5 million users and 92T tokens a week

While the industry debates which model is smarter on the latest benchmark, users are voting with their tokens. And judging by the latest figures, DeepSeek Harness, a framework for working with DeepSeek models, has gained impressive traction even before its official release.

What we know

According to figures shared by the project's official account, DeepSeek Harness has more than 1.5 million active users in beta. Together, they consume over 92T DeepSeek tokens a week. The author of an analysis on the Telegram channel "Russkiy IT Biznes" compares this with OpenRouter's traffic: it works out to roughly half the platform's total volume across all models. The comparison is approximate, but it gives a sense of the scale.

The product is formally still a Release Candidate. An important caveat: the statistics include only users of DeepSeek's official API. If other ways of accessing the models and subscriptions are included, the author estimates that the user base could be 1.5–2 times as large. This is an estimate, not a confirmed figure.

Plugins as the key signal

The most interesting metric here is not even token volume. Nearly 60% of users create their own plugins. For a product in beta, that is a very high share: people are doing more than running a ready-made workflow; they are adapting the tool to their work.

Imagine a typical situation. A team uses the model for code reviews, but also needs the results to flow automatically into its issue tracker and follow internal naming conventions. Previously, this meant writing custom wrapper scripts around the API. If the framework provides a built-in plugin mechanism, those scripts become portable modules. Judging by the share of users building them, that is precisely what draws people in.

For the market, this means value is shifting from the model itself to the environment around it: how tools connect, how context is stored, and how work can be shared. The model becomes an interchangeable component, while established habits and accumulated plugins remain.

How to interpret these figures cautiously

First: the figures come from the project itself, and the post provides no independent verification. Beta-stage marketing statistics tend to favor convenient metrics, and "active user" can mean anything from a developer using the product every day to someone who ran it once.

Second: 92T tokens a week indicates intensity of use, but not quality. Agent loops can burn through enormous numbers of tokens on retries and long context. A high token count does not necessarily mean more value per unit of spending.

Third: the OpenRouter comparison has limitations. These are different access channels, and their respective shares need to be calculated carefully. The post's author also openly notes that they use the tool almost every day, so this is the perspective of an invested practitioner rather than a neutral analyst.

What to do in practice

If you are building an LLM-based product, take a closer look at how you handle extensibility. The question is whether your users or internal teams can add behavior themselves, beyond simply choosing which model to connect. The high share of plugin authors in another product is a sign that this capability is in real demand.

If you are a developer, it makes sense to try Harness in your daily workflow while it is still a release candidate and assess token costs on real tasks in advance. If you are a product manager, track not only request counts but also the share of users who customize the tool. It is one of the clearest indicators that people are using a product seriously.

The next step is to wait for the release and independent data. For now, the picture is this: DeepSeek has developed an ecosystem around its API, and it is growing faster than many expected.