What DeepSeek V4 Pro actually shipped

DeepSeek took V4 Pro out of preview on August 13, 2026. The served build is DeepSeek-V4-Pro-0813. The API name stays deepseek-v4-pro. On the app and the website the same model sits behind Expert Mode.

This is not a new architecture drop. The Hugging Face card for 0813 says it keeps the preview model structure and adds a DSpark speculative-decoding module. The April preview card still lists the size: 1.6 trillion total parameters, 49 billion active, one million tokens of context.

How you call it

DeepSeek’s pricing page lists both V4 Flash and V4 Pro with a 1M context window and a 384K maximum output. Thinking mode is on by default. You can also run a non-thinking mode.

The August 13 note adds three reasoning-effort levels for both Pro and Flash: low, high, and max. DeepSeek’s own guidance is low for simple work, high for ordinary agent jobs, and max for harder ones.

The same note says the API now speaks the OpenAI Responses API format and ships a one-click Codex setup.

What DeepSeek says improved

The company’s headline is agent work, not a new chatbot. The change log publishes its own scores for 0813 against the April Pro preview. A few of those rows:

Those are DeepSeek’s numbers, not an independent lab. Two of the coding-agent sets on that table, DSBench-FullStack and DSBench-Hard, are internal.

An independent index

Artificial Analysis scores the reasoning, max-effort 0813 build at 53 on its Intelligence Index. It scores V4 Flash at 40. Reuters repeated those two figures on August 13. The index is a blend of coding, tool use, scientific questions, and related tests. It is not a vote on whether the model “beats” any one closed system.

What it costs on the first-party API

New list prices took effect at 16:00 UTC on August 16, 2026. DeepSeek now bills a peak rate and an off-peak rate. Off-peak is half of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC.

Per 1 million tokens, official USD list prices:

At those peak cache-miss and output rows, Pro is three times Flash, not a different order of magnitude. Concurrent request caps differ too: 500 for Pro, 2,500 for Flash.

Flash versus Pro

Flash is the cheaper, higher-concurrency model. Pro is the one DeepSeek is selling as the agent flagship. Both share the 1M window and the same three effort knobs. If the job is a short completion, Flash is the price that matches that job. If the job is a long tool loop, Pro is the SKU DeepSeek published the new agent table for.

Open weights

The 0813 weights are on Hugging Face under an MIT license, at deepseek-ai/DeepSeek-V4-Pro-0813. That is the part that still separates this release from a closed API-only model. Serving 1.6T MoE weights is a hardware problem, not a license problem.

The mechanism to watch is not a brand war. It is the split DeepSeek is now explicit about: a cheap Flash SKU, a dearer Pro SKU, time-of-day API prices, and a published agent table for the expensive one.

Source cutoff: DeepSeek GA note, change log, and pricing page; Hugging Face 0813 and preview cards; Reuters and Artificial Analysis, as retrieved August 17, 2026.