On Thursday (13), DeepSeek released the V4-Pro-0813 version of its language model, made the DeepSeek Harness agent software available as open source, and announced new API prices with rates varying by time of day.
New version of V4 Pro
The deepseek-v4-pro endpoint now delivers the V4-Pro-0813 build. The company says the model name, parameter count, and 1 million-token context window have not changed, and existing integrations will keep working without adjustments. In the app and on the web, the model appears in 'Expert Mode', with native support for the OpenAI Responses API and integration with Codex. The reasoning effort level can be low, high, or maximum, with the intermediate level recommended for everyday agent use.
In DeepSeek's tests, V4-Pro-0813 rose from 72.1 to 87.9 on Terminal Bench 2.1 and from 12.8 to 62.7 on DeepSWE, surpassing Claude Opus 4.8 in several agent benchmarks. The Artificial Analysis Intelligence Index, meanwhile, puts the model at 53 points, versus 45 for the previous version, tied with GLM-5.2. It still trails Muse Spark (57), Qwen 3.8 Max (58), Kimi K3 (60), and Claude Opus 5 (63). DeepSeek has not published the weights of the new build; the April preview remains on Hugging Face.
The update comes after the smaller V4 Flash model, with its July 0731 update, practically matched the Pro Preview's performance on the Artificial Analysis index at a much lower cost.
DeepSeek Harness
DeepSeek also launched DeepSeek Harness v0.1, a developer preview under the MIT license. The tool is presented as an alternative to OpenAI's Codex and Claude, and uses the Cordis plugin system, in which all functions are swappable components. A continuous session log records prompts, tool calls, and results, allowing executions to be resumed, branched, and reproduced. In minimal mode, only shell and file editor remain, a configuration the company uses in its own benchmarks. The project is led by Cui Tianyi, who joined DeepSeek in March 2026 from Jane Street. A call for testers in early August attracted 712 projects in three days.
API price adjustment
The new rates take effect on August 16 at 16 UTC. DeepSeek had announced the shift to peak and off-peak pricing in late June, without detailing amounts or a date. Off-peak usage costs half the price. Peak hours run from 1 to 4 and from 6 to 10 UTC, aligned with the Chinese workday; for Europe, almost the whole afternoon falls under the lower rate.
Off-peak, V4-Pro's input price rises from US$0.435 to US$0.66 per million tokens, and the output price from US$0.87 to US$1.98. At peak, the prices double to US$1.32 and US$3.96. Cache hits go from US$0.003625 to US$0.022 off-peak and US$0.044 at peak, cutting the discount from about 1/120 to 1/30 of the normal input price. For agents that read the same files repeatedly, this is the item with the biggest impact.
The adjustment partially reverses the price cut made in May, and cache hits now cost more than they did before the reduction. The increase comes as DeepSeek raises funds and prepares to go public.
