DeepSeek

On 13 August 2026 «DeepSeek» released two things on the same day: «DeepSeek Harness», an agent harness published open source under the MIT license, and the general availability of «DeepSeek-V4-Pro». Three days later, on 16 August at 16:00 UTC, a new API price grid took effect, splitting rates into peak and off-peak hours with off-peak set 50 percent below peak. The harness is given away and the inference is billed, which inverts the packaging Anthropic uses around «Claude Code».

The Chinese lab communicates through unsigned institutional product pages: no technical blog, no named author, no engineer to address a question to. The Harness launch page ran roughly four hundred words and contained not a single figure. Its strongest technical claim, that anything reaching a model request must be reconstructable from the session log and that a runtime invariant asserts it, appears in `docs/architecture.md` rather than in the announcement.

This is the same posture visible a year earlier, when the October 2025 Batch issue covered DeepSeek's cuts to inference costs: cost and access as the argument, not benchmark theatre.

The competitive picture is contradictory. «DeepSeek» competes directly with «modèles américains de pointe», and its share on OpenRouter had doubled by the time of the WAIC 2026 opening in Shanghai. Its web presence, over the same period, measures 40 percent below its February 2025 peak. Developer channel up, consumer channel down.