DeepSeek V4-Pro goes GA with a reasoning-effort dial
Moved to general availability on 16 August with agent upgrades, selectable reasoning effort at low, high or max, and native support for the OpenAI Responses API. The dial is the part that changes how you budget.
V4-Pro-0813 reached general availability on 16 August. Three things came with it: agent upgrades, a reasoning-effort setting with low, high and max positions, and native support for the OpenAI Responses API for Codex-style setups.
The effort dial is the consequential one. Until recently the amount of thinking a model did was the model's business, decided at training time and exposed at best as a separate SKU. Making it a request parameter hands the cost and latency trade to the caller, per call.
That is a different product shape than a tiered model family. With tiers you pick a model and live with its economics. With a dial you can spend heavily on the one query that matters and nothing on the ninety-nine that do not, inside the same integration. It rewards teams that know which of their queries are hard, and punishes ones that do not.
The Responses API support is the quieter change with the wider blast radius. A competitor implementing a rival's interface natively is a statement that the interface has stopped being a differentiator.
DeepSeek has had an eventful month on price in both directions. The movements have not formed a trend. This release adds a third axis to the same question: not what a token costs, but how much thinking you asked for.
Digital Applied — AI Model Releases: August 2026 Tracker and Dated Ledger → · LLM Stats — AI Updates Today (August 2026) — Latest AI Model Releases →