DeepSeek V4 Pro Goes General Availability: The Agentic Flagship with a 1 Million Token Window

by Geth Author

Yesterday, DeepSeek officially moved its flagship model out of preview. DeepSeek-V4-Pro-0813 is now generally available across the company's API, web app, and mobile clients — and it arrives with a sharply focused pitch: agents.

The architecture is a mixture-of-experts model with 1.6 trillion total parameters, of which 49 billion are active on each forward pass. Its context window reaches one million tokens, and it can produce outputs up to 384,000 tokens long — both numbers that make it practical for processing large codebases, extended legal documents, or multi-step reasoning chains that require holding a lot of state at once. The model can run in either thinking or non-thinking mode, with thinking effort settable at low, high, or max.

What DeepSeek is emphasizing with V4 Pro is not raw benchmark performance but agent readiness. The API has been updated to be natively compatible with the OpenAI Responses API format, meaning existing tool-calling workflows written for GPT can migrate with minimal code changes. Built-in Codex integration is included out of the box. Alongside the launch, DeepSeek published benchmark results: 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo — metrics that measure multi-step shell task completion, software engineering capability, and natural-language-to-repository workflows respectively.

Pricing currently sits at $0.435 per million input tokens (cache miss) and $0.87 per million output tokens. A price increase takes effect at 16:00 UTC on August 16, introducing peak and off-peak billing, with off-peak rates at half the peak price. If you want to lock in current pricing, the window closes in roughly 48 hours.

Independent benchmarking is already underway — headline scores from model providers have a track record of diverging from community results, and Terminal Bench in particular varies significantly with prompting strategy. That caveat aside, V4 Pro's combination of a genuine one-million-token context window, mixture-of-experts efficiency, and direct OpenAI API compatibility makes it one of the more immediately deployable model drops of 2026. If you've been building agents on GPT-4.1 or Sonnet 5 and want a lower-cost alternative with more headroom on output length, this is worth evaluating before the price structure changes on Saturday.