DeepSeek, a start-up in Hangzhou, said users of its Harness agent platform were consuming about 92 trillion tokens a day through its official API as of November 4, the South China Morning Post reported. DeepSeek puts that at about 1.1 billion tokens per second. Tokens are small chunks of text a model reads or writes. Harness lets an AI agent run code and use other software for a user. It went live on August 13, and a desktop app followed on September 29.
DeepSeek also said there are about 35,000 third-party plug-ins, and that around 60 per cent of active users have built or installed at least one. Cui Tianyi, who joined the Harness team in March, said on Sunday that the goal is for Harness to act as "an operating system for the agent era". The Paper reported that DeepSeek set up its first product division last month, led by Cui. The source gives no revenue, pricing or user counts.
These are DeepSeek's own figures. A token count shows how much the system is used, not how many tasks were finished, so buyers can ask for tokens per finished task. Harness agents can run code and use outside software, and the plug-ins come from third parties. That makes it worth asking what each plug-in can reach and who reviews it.