DeepSeek said its Harness AI-agent framework had reached 1.54 million daily active users less than three months after launch, according to a November 9 report by the South China Morning Post. The company also reported around 92 trillion tokens per day through Harness users of its official API as of November 4. The two figures measure different kinds of activity.
The user count and API volume answer different questions
The 1.54 million figure is DeepSeek’s reported count of daily active Harness users. The API figure concerns tokens consumed per day by Harness users accessing DeepSeek’s official API. The South China Morning Post equated that November 4 volume to about 1.1 billion tokens per second.
Token volume is not a count of people or an average per user. The figures therefore describe separate dimensions of reported Harness use, rather than two ways to express the same metric.
Treat the 1.54 million figure as a reported daily-user count and the 92 trillion figure as API token volume: they describe different kinds of activity.
A growth graphic combines different measurement windows
A graphic attributed to Hesamation shows daily-user values rising from 271,000 on September 30 to 1.54 million on November 4. It labels the earlier points as seven-day averages, while the November 4 point represents a single day. The graphic also carries a “+545% since desktop release” annotation.
Because the plotted values use different measurement windows, they cannot be read as a like-for-like start and end for that growth percentage. The displayed figures do not establish the +545% claim as growth between those endpoints.
What DeepSeek Harness does
The South China Morning Post describes a harness as software infrastructure that lets a language model act as an agent and execute tasks, rather than only respond to prompts. That context helps explain why a user count and API token volume can capture different aspects of activity around the framework.
For product-development context, NeoTeo’s earlier coverage of Harness 0.2.2-rc.1 describes its experimental delegation to local coding CLIs and separate Git worktrees.