Deepseek Ships Improved V4 Pro, Open-Sources Agent Software, and Raises API Prices

On August 13, 2026, DeepSeek announced that its flagship model V4-Pro has moved from beta to general availability. The endpoint deepseek-v4-pro now serves build V4-Pro-0813. The model name, parameter count, and 1M token context window remain unchanged, and DeepSeek states existing integrations will continue to work without any adjustments. In the app and web interface, the model is offered as "Expert Mode" and now includes native support for the OpenAI Responses API, integrating Codex. Reasoning effort can be set to "low," "high," or "maximum," with DeepSeek recommending the middle setting for everyday agents.

According to DeepSeek's own comparison table, V4-Pro's score on Terminal Bench 2.1 jumps from 72.1 to 87.9, and DeepSWE from 12.8 to 62.7. On several agent benchmarks, the model surpasses Claude Opus 4.8. However, data from Artificial Analysis shows V4-Pro's Intelligence Index rises from 45 to 53, tying with GLM-5.2, but still trailing Muse Spark (57), Qwen 3.8 Max (58), Kimi K3 (60), and Claude Opus 5 (63). DeepSeek has not yet released weights for the new build; Hugging Face still hosts the April preview.

Alongside the model update, DeepSeek released Harness v0.1 as a developer preview under the MIT license. This open-source agent software is positioned as an alternative to OpenAI Codex and Claude, built on the newly released Cordis plugin system, where all features—from tools and sandboxes to sessions and UI—are replaceable plugins. It provides continuous session logs recording every prompt, tool call, and result, with support for resume, branching, and replay.

Harness's minimal mode retains only the shell and file editor, which DeepSeek used for its own benchmarks. The software is launched via npx and offers a local web interface, though DeepSeek warns of compatibility issues. The project is led by Cui Tianyi, who joined DeepSeek in March 2026 from quantitative trading firm Jane Street. When the team called for testers in early August, 712 projects signed up within three days.

New API pricing takes effect on August 16 at 16:00 UTC. DeepSeek announced the shift to peak/off-peak pricing in late June but did not disclose specific numbers and dates at that time. Off-peak usage costs are halved. Peak hours are UTC 1:00-4:00 and 6:00-10:00, corresponding to Chinese workdays. For European users, nearly the entire afternoon falls under the lower rate.

During off-peak hours, V4-Pro input price rises from $0.435 to $0.66 per million tokens, and output from $0.87 to $1.98. During peak hours, these rates double to $1.32 and $3.96. Cache hits see the largest increase, from $0.003625 to $0.022 off-peak and $0.044 peak, with the cache discount shrinking from about 1/120 to 1/30. For agents repeatedly reading the same files, this is the most expensive part. The new pricing partially offsets the May price cut, with cache hit costs even higher than before the cut. This increase comes as the company raises new capital and prepares for an IPO.