DeepSeek

On 13 August 2026 DeepSeek released its coding agent harness, DeepSeek Harness, under an MIT license and declared DeepSeek-V4-Pro generally available, both the same day. Three days later, at 16:00 UTC on 16 August, a new API price schedule took effect with peak and off-peak rates, off-peak set 50% below peak. Harness given away, inference billed.

The lab speaks through unsigned institutional product pages: no technical post, no named author. That reticence matters, because the sharpest claims attached to its name are made by others. Anthropic's threat intelligence report of 10 September 2026 names seven Chinese laboratories on distillation and places DeepSeek beside Moonshot AI: both serve Claude responses to their own customers without telling them, and both convert the thinking signature into a complete reasoning trace by cross-session replay. The traffic recorded against DeepSeek is relay of coding harness users toward Opus, 12.1 million exchanges in fourteen days.

Its commercial position points two ways at once. In October 2025 the lab was cutting inference costs. By mid-July 2026 its share on OpenRouter had doubled, in the same frame as Xi Jinping's 17 July WAIC address in Shanghai arguing for open weights and low inference costs against the American approach. On the web, usage sits 40% below the February 2025 peak.

So the lab gives away the tool, raises the price of what the tool consumes, grows where developers route API calls, and shrinks where consumers click. It competes with modèles américains de pointe partly by serving their output.