ଡିପସିକ୍ V4 ପ୍ରୋ 0813

The flagship of the DeepSeek V4 series: a Mixture-of-Experts model with 1.6T total and 49B activated parameters over a one-million-token context, using the same hybrid attention design as V4 Flash. It is trained with the Muon optimiser for stability and served in mixed FP4 and FP8 precision. DeepSeek positions its maximum thinking configuration as the best open-source model available for coding, reasoning and agentic tasks.

The 0813 build is the general-availability V4-Pro, shipped to the app, the web experience and the API on the same day and reachable in the chat products through Expert Mode. DeepSeek leads on agent capability rather than on raw exam scores, and says the improvement is most visible in production rather than on a leaderboard - the argument being that a model which finishes long tool-driven jobs matters more than one that peaks on a single-shot question.

The numbers the lab published for this checkpoint are its highest across the V4 line: 42.7 on Humanity's Last Exam without tools and 60.0 with them, 87.9 on Terminal Bench 2.1, 61.5 on NL2Repo, 83.3 on Cybergym, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified, 25.7 on Agents' Last Exam, 31.8 on the public AutomationBench split, and 71.1 and 67.2 on its internal DSBench full-stack and hard-problem sets. Read against the preview weights the gap is stark on the agentic evaluations in particular - the lab's own table puts V4-Pro-Preview at 72.1 on Terminal Bench 2.1, 38.5 on NL2Repo and 12.8 on DeepSWE.

Two controls arrived with the release and apply to both V4 models. Thinking effort became a three-position dial - low, high and max - and DeepSeek gives concrete guidance rather than leaving it to taste: low for simple requests, high for everyday agent work, max for genuinely complex scenarios. The API also gained native OpenAI Responses API support with a one-command Codex configuration, on top of the existing OpenAI and Anthropic compatible surfaces.

Pricing changed shape at the same time. DeepSeek introduced peak and off-peak rates, with off-peak set at half the peak price, explicitly to push batchable work into quieter hours, effective 16:00 UTC on 16 August 2026. The model name stayed `deepseek-v4-pro`, so callers moved to the GA weights without touching their code. Where this model stops is cost rather than capability: it is several times the price of V4-Flash, and DeepSeek's own framing sends everyday high-volume traffic to Flash and reserves Pro for the hard end.