Published: 14 August 2026
DeepSeek released DeepSeek-V4-Pro-0813 on 13 August 2026, moving its flagship V4-Pro model out of preview, and published its agent software Deepseek Harness v0.1 as open source under the MIT licence. On DeepSeek’s own harness the new build scores 87.9 on Terminal-Bench 2.1, against 72.1 for the April V4-Pro preview, and the company announced higher API prices that take effect on 16 August 2026. The model keeps its one million token context window and existing integrations continue to work without changes.
What changed in DeepSeek-V4-Pro-0813?
The deepseek-v4-pro endpoint now serves the V4-Pro-0813 build. DeepSeek says the model name, parameter count and one million token context window are unchanged, so callers were moved onto the new checkpoint without a code change. In the DeepSeek app and on the web the model sits behind an Expert Mode toggle.
The headline changes are operational rather than architectural. DeepSeek-V4-Pro-0813 and DeepSeek-V4-Flash-0731 both gained a selectable reasoning effort with three levels, low, high and max, replacing the single fixed thinking budget of the earlier builds. DeepSeek recommends the high setting for everyday agent work. The release also adds native support for the OpenAI Responses API with one click Codex setup. DeepSeek has not published weights for this build; the April V4-Pro preview remains the most recent version on Hugging Face.
DeepSeek-V4-Pro-0813 benchmarks and technical specs
The figures DeepSeek published come from its own evaluation harness, run in minimal mode, and the company’s numbers for rival models do not always match those labs’ published figures, so the comparison reads best within the V4 family.
- On that harness DeepSeek-V4-Pro-0813 scores 87.9 on Terminal-Bench 2.1, against 82.7 for DeepSeek-V4-Flash-0731 and 72.1 for the April V4-Pro preview build,
- 62.7 on DeepSWE against 12.8 for the preview,
- It reports 83.3 on CyberGym,
- 74.1 on Toolathlon-Verified,
- 31.8 on the public AutomationBench split,
- 42.7 percent on Humanity’s Last Exam without tools rising to 60.0 percent with tools.
Independent scoring puts the gain in context. Artificial Analysis moves V4-Pro from 45 to 53 on its Intelligence Index, level with GLM-5.2 from Z.ai, and behind Meta’s Muse Spark at 57, Alibaba’s Qwen3.8-Max at 58, Moonshot’s Kimi K3 at 60 and Anthropic’s Claude Opus 5 at 63.
Secondary coverage describes the model as a mixture of experts system with 1.6 trillion total parameters and 49 billion active per token; DeepSeek did not restate the parameter count in the launch post, so treat that figure as reported rather than confirmed at this release.
What is Deepseek Harness v0.1?
Alongside the model, DeepSeek published Deepseek Harness v0.1 as a Developer Preview under the MIT licence. It is the agent software the lab uses internally, pitched as an open alternative to OpenAI Codex and Claude Code. Harness is built on the Cordis plugin system, in which every feature is a swappable plugin, from tools and sandboxes through to sessions and the interface.
A continuous session log records every prompt, tool call and result, and runs can be resumed, branched and replayed. A minimal mode strips the setup down to a shell and a file editor, and that is the configuration DeepSeek used for its own benchmark runs. The software launches through npx into a local web interface, and DeepSeek warns of compatibility issues at this stage. The project is led by Cui Tianyi, who joined DeepSeek from the quantitative trading firm Jane Street in March 2026.
DeepSeek-V4-Pro-0813 availability and pricing
DeepSeek-V4-Pro-0813 is available now through the DeepSeek API under the existing deepseek-v4-pro model name, in the DeepSeek app and on the web under Expert Mode, and through third party routers including OpenRouter. The launch post lists app, web and API availability only, with no accompanying weight release.
New API rates take effect on 16 August 2026 at 16:00 UTC, and split into peak and off peak pricing. Off peak, V4-Pro input rises from 0.435 dollars to 0.66 dollars per million tokens and output from 0.87 dollars to 1.98 dollars. Peak rates are double that, at 1.32 dollars and 3.96 dollars. Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, matching the Chinese working day, which means most of the European afternoon falls under the lower rate. Cache hits see the steepest increase, from 0.003625 dollars to 0.022 dollars off peak and 0.044 dollars at peak, shrinking the cache discount from roughly one hundred and twentieth of the regular input price to one thirtieth. For agents that repeatedly read the same files, that is the most expensive part of the change.
For more information on pricing and API you can visit DeepSeek’s API documentation.
