DeepSeek ships V4 Pro out of testing, open-sources agent software, raises API prices

DeepSeek has promoted its flagship model V4-Pro out of preview and into production with the release of build V4-Pro-0813. The model retains its one-million-token context window and parameter count, and existing integrations will continue to work without modification. In the app and on the web, users can access the model via "Expert Mode." A significant addition is native support for the OpenAI Responses API with Codex integration, and developers can now set reasoning effort to one of three levels—"low," "high," and "max"—with DeepSeek recommending the middle setting for everyday agentic workflows.

Performance metrics show substantial gains. On DeepSeek's own comparison table, Terminal Bench 2.1 scores jumped from 72.1 to 87.9, and DeepSWE scores climbed from 12.8 to 62.7. On several agent benchmarks, V4-Pro beat Claude Opus 4.8. However, according to Artificial Analysis's Intelligence Index, V4-Pro now scores 53 (up from 45 in the preview), tying GLM-5.2 but still trailing Muse Spark at 57, Qwen 3.8 Max at 58, Kimi K3 at 60, and Claude Opus 5 at the top with 63 points. The timing of the release reflects competitive pressure: at the end of July, DeepSeek had shipped update 0731 for its smaller V4 Flash model, which nearly matched the Pro Preview on the Artificial Analysis Intelligence Index while costing far less.

Alongside the model release, DeepSeek is open-sourcing Deepseek Harness v0.1 under the MIT license as a Developer Preview. The agent orchestration software is positioned as an alternative to OpenAI's Codex and Claude-based frameworks. It is built on the Cordis plugin system, where all features—tools, sandboxes, sessions, and the UI—are swappable plugins. Every prompt, tool call, and result is logged in a continuous session, and runs can be resumed, branched, and replayed. A minimal mode strips the interface to a shell and file editor, which DeepSeek uses for its own benchmark runs. The software launches via npx through a local web interface, though compatibility warnings apply. The project is led by Cui Tianyi, who joined DeepSeek from quantitative trading firm Jane Street in March 2026. When DeepSeek put out a call for beta testers in early August, 712 projects signed up within three days.

DeepSeek announced the new API pricing structure on August 16 at 4:00 p.m. UTC, introducing peak and off-peak rates that align with Chinese business hours. Peak hours run from 1 a.m. to 4 a.m. and 6 a.m. to 10 a.m. UTC; off-peak hours apply the rest of the day. For V4-Pro, off-peak input prices rise from $0.435 to $0.66 per million tokens, and output jumps from $0.87 to $1.98. During peak hours, those rates double to $1.32 and $3.96 for input and output respectively. Cache hits see the steepest increase, going from $0.003625 off-peak to $0.022 off-peak and $0.044 at peak, shrinking the cache discount from about one one-hundredth to one thirtieth of the regular input price. For agentic applications that repeatedly read the same files, this represents the most expensive component of the change. The new pricing partially undoes the price cut DeepSeek rolled out in May, and cache hits will actually cost more than they did before that reduction. DeepSeek has announced it is raising new capital and preparing for an initial public offering.