Deepseek has released version 4-Pro-0813, open-sourced its agent software, and increased API costs effective August 16 at 4:00 p.m. UTC.
Model performance and availability
The new build moves the flagship model out of testing. It retains the same parameter count and one-million-token context window as previous versions. Existing integrations will continue to function without modification. Users can access the model via the “Expert Mode” interface on the app and web platforms.
The update adds native support for the OpenAI Responses API alongside Codex integration. Developers can adjust reasoning effort across three levels: “low,” “high,” and “max.” Deepseek recommends the middle setting for routine agent tasks.
Internal benchmarks show significant gains. Terminal Bench 2.1 scores rose from 72.1 to 87.9. DeepSWE scores jumped from 12.8 to 62.7. On several agent benchmarks, the model outperformed Claude Opus 4.8.
External rankings and context
Artificial Analysis confirms the improvement but places it in perspective. The V4-Pro model climbs from 45 to 53 on the Intelligence Index, matching GLM-5.2. This score remains behind Muse Spark at 57, Qwen 3.8 Max at 58, and Kimi K3 at 60. Claude Opus 5 leads the field with 63 points.
Deepseek has not yet published the weights for this new build. The April preview version remains available on Hugging Face.
The update addresses competition from the smaller V4 Flash model. At the end of July, Deepseek shipped update 0731 for V4 Flash. That version nearly matched the Pro Preview on the Intelligence Index while costing a fraction of the price.
Open-source agent software
Deepseek Harness v0.1 launches as a Developer Preview under the MIT license. The tool acts as an alternative to OpenAI’s Codex and Claude. It relies on the Cordis plugin system, allowing users to swap features like tools, sandboxes, sessions, and the user interface.
A continuous session log records every prompt, tool call, and result. Users can resume runs, create branches, or replay sessions. Minimal mode reduces the environment to a shell and file editor. Deepseek uses this setup for its own benchmark runs.
The software launches via npx through a local web interface. Deepseek notes potential compatibility issues. Cui Tianyi leads the project. He joined Deepseek from quantitative trading firm Jane Street in March 2026. A call for beta testers in early August attracted 712 projects within three days.
Price changes and impact
New rates begin on August 16 at 4:00 p.m. UTC. Deepseek announced the switch to peak and off-peak pricing in late June without providing specific figures or dates at that time. Time-based rates existed since February 2025, when the company offered discounts on V3 and R1 during nighttime hours.
Off-peak usage costs half as much as peak usage. Peak hours run from 1 a.m. to 4 a.m. and 6 a.m. to 10 a.m. UTC, aligning with the Chinese workday. For users in Europe, nearly the entire afternoon falls under the lower rate.
During off-peak hours, V4-Pro input costs rise from $0.435 to $0.66 per million tokens. Output costs jump from $0.87 to $1.98. During peak hours, those rates double to $1.32 and $3.96.
Cache hits see the steepest increase. Costs go from $0.003625 to $0.022 off-peak and $0.044 at peak. This shrinks the cache discount from about one-hundred-twentieth to one-thirtieth of the regular input price. For agents that repeatedly read the same files, this is the most expensive part of the change.
The new pricing partially reverses the price cut Deepseek rolled out in May. Cache hits will actually cost more than they did before that reduction. The price hike comes as the company raises new capital and prepares for an initial public offering.
