DeepSeek shipped its flagship with no announcement — same 1.6 trillion parameters, five times the agent score
Late on Wednesday night in Beijing, DeepSeek's API documentation quietly switched from the V4-Pro preview to a build named DeepSeek-V4-Pro-0813. There was no blog post and no launch event. It is the first finished version of the flagship that had been in preview since 24 April, and the architecture did not change: 1.6 trillion parameters in a mixture of experts, about 49 billion active per token, a million tokens of context, MIT licence. Everything moved in post-training, and it moved most where agents work. On DeepSeek's own scorecard DeepSWE goes from 12.8 to 62.7, Terminal Bench from 72.1 to 87.9, Cybergym from 52.7 to 83.3 and NL2Repo from 38.5 to 61.5. The API now also speaks the Responses and Anthropic protocols, so it drops into Codex and Claude Code without a wrapper. Pricing holds at ¥3 per million input tokens and ¥6 per million output — with a note from the company that it will rise significantly before long.
Related stories
Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21ChatGPT can now read your iMessages and send them — and the setting that lets it skip asking is the one OpenAI warns about
2026-08-21