DeepSeek V4 Pro

What it is

DeepSeek-V4-Pro-0813 is the general-availability release of DeepSeek’s flagship, rolled out on 13 August 2026 across the app, the web and the API. It supersedes the V4 Pro preview that launched and was open-sourced on 24 April 2026, and the company’s model card is direct about where the work went: “greatly enhanced agentic capabilities and performance improvements”, built on the same V4 base with a new speculative-decoding module the company calls DSpark.

The open publishing habits carry over. The weights are on Hugging Face under an MIT licence, 66 checkpoint shards released in FP8, with the technical report on arXiv. The listing shows 1.7 trillion total parameters, up from the preview’s stated 1.6 trillion, though DeepSeek’s own text does not restate a parameter count for the GA build, so treat the 1.7 figure as Hugging Face’s reading of the uploaded files rather than a company claim. Context is one million tokens, with output up to 384,000, and the API adds three reasoning-effort levels, low, high and max, plus native support for the OpenAI Responses format alongside the Anthropic one.

The benchmark claims, and whose harness ran them

DeepSeek’s own table shows the GA build far ahead of the April preview on agentic coding: Terminal Bench 2.1 at 87.9 against the preview’s 72.1, DeepSWE at 62.7 against 12.8, NL2Repo at 61.5 against 38.5. The company also prints its methodology, and it deserves quoting: the code-agent scores were “evaluated with the minimal mode of DeepSeek Harness as the agent framework, using the max reasoning effort level”. Its own harness, its own settings. The jumps are plausible, and they are unaudited.

The independent reading is calmer. Artificial Analysis scores V4-Pro-0813 at 53 on its Intelligence Index, third of the 106 models in its class and well above the open-weight median of 27, but seven points behind the other open-weight flagship of the moment, Kimi K3, at 60. The same comparison has DeepSeek answering at twice K3’s output speed and roughly a third of its blended price, which is the trade the model has always offered: nearly the intelligence, for a fraction of the cost.

The price rise is the story

From the April preview onward, V4 Pro’s API price was a flat $0.87 per million output tokens at any hour. At 16:00 UTC on 16 August 2026 that ended. DeepSeek moved to peak and off-peak billing, and those are the rates in force: output at $1.98 per million off-peak and $3.96 at peak, with peak defined as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, and off-peak set at half peak. Weekends bill at the off-peak rate throughout. Input on a cache miss is $0.66 off-peak and $1.32 at peak. Cached input, the number that makes long agent sessions cheap, went from $0.003625 per million to $0.022 off-peak and $0.044 at peak, a rise of roughly twelve times at the top of the clock.

Even at peak, $3.96 per million output undercuts most frontier APIs. But the era in which DeepSeek’s price was a constant you could ignore in capacity planning is over, and the change landed three days after the applause for the GA release. Our own V4 page called this model line “the open-weight price disruptor”; the disruptor now has peak hours.

Where it runs

The API serves the model at api.deepseek.com in both OpenAI and Anthropic formats, the chat product is free in the browser, and OpenRouter lists deepseek/deepseek-v4-pro-0813 with the full million-token window and all three reasoning efforts. Anyone with the hardware can also simply take the MIT weights.

For the April preview this build replaces, and the mixture-of-experts design underneath both, see DeepSeek V4.