Deepseek is making three moves at once: it has promoted an updated V4-Pro build, opened its agent software to developers, and set higher API prices. The combination matters because it ties model performance, agent tooling, and operating costs into one package for teams already building on Deepseek.
The new build is V4-Pro-0813, delivered through the existing deepseek-v4-pro endpoint. Deepseek says current integrations can continue without changes, while users in the app and on the web can access the model through "Expert Mode."
What Changed In V4-Pro
The model name, parameter count, and one-million-token context window remain the same. That continuity is important for developers because the update is presented as an upgrade to the existing product rather than a new integration target.
Deepseek has added native support for the OpenAI Responses API with Codex integration. The model also exposes three reasoning effort settings: "low," "high," and "max." For everyday agent use, Deepseek recommends the middle setting.
The biggest reported gains are in agent-related benchmarks. According to Deepseek's own comparison table, Terminal Bench 2.1 rose from 72.1 to 87.9, while DeepSWE moved from 12.8 to 62.7. The company also says V4-Pro beat Claude Opus 4.8 on several agent benchmarks.
Those results point to a clearer product direction. Deepseek is not only competing on general model quality; it is also trying to make V4-Pro more useful in workflows where a model reads, edits, calls tools, and continues a task over multiple steps.
How It Ranks Against Other Models
Artificial Analysis also shows improvement, but its ranking places the update in a more crowded field. V4-Pro moved from 45 to 53 on the Intelligence Index, putting it level with GLM-5.2.
That still leaves several models ahead in the same comparison. Muse Spark is listed at 57, Qwen 3.8 Max at 58, and Kimi K3 at 60. Claude Opus 5 remains above them with 63 points.
Deepseek has not released the weights for the new build yet. The April preview version is still available on Hugging Face, so developers looking for open weights do not yet have the new V4-Pro-0813 build in that form.
The update also helps separate V4-Pro from V4 Flash. At the end of July, Deepseek shipped update 0731 for V4 Flash, and that smaller model nearly matched the Pro Preview on the Artificial Analysis Intelligence Index while costing much less.
Deepseek Harness Becomes Open Source
Alongside the model update, Deepseek is releasing Deepseek Harness v0.1 as a Developer Preview under the MIT license. The company presents it as open-source agent software and as an alternative to OpenAI's Codex and Claude.
The software is built on the newly released Cordis plugin system. In that design, major parts of the agent environment are swappable plugins, including tools, sandboxes, sessions, and the UI.
A continuous session log records each prompt, tool call, and result. Runs can be resumed, branched, and replayed, which gives developers a way to inspect and repeat agent behavior rather than treating each run as a black box.
Deepseek also includes a minimal mode that reduces the setup to the shell and file editor. The company uses that configuration for its own benchmark runs. Harness launches through npx in a local web interface, though Deepseek warns about compatibility issues.
The project is led by Cui Tianyi, who joined Deepseek from quantitative trading firm Jane Street in March 2026. When Deepseek asked for beta testers in early August, 712 projects signed up within three days.
API Pricing Is Moving Higher
The new API rates take effect on August 16 at 4:00 p.m. UTC. Deepseek had announced a move to peak and off-peak pricing at the end of June, but it had not provided the exact prices or start date then.
Time-based pricing is not entirely new for the company. Since February 2025, Deepseek has offered a discount on V3 and R1 during nighttime hours.
Under the new structure, off-peak usage costs half as much as peak usage. Peak hours are 1 a.m. to 4 a.m. and 6 a.m. to 10 a.m. UTC, matching the Chinese workday. For users in Europe, almost the full afternoon falls into the lower-priced period.
The V4-Pro price increases are material:
- Off-peak input rises from $0.435 to $0.66 per million tokens.
- Off-peak output rises from $0.87 to $1.98 per million tokens.
- Peak input is $1.32 per million tokens.
- Peak output is $3.96 per million tokens.
- Cache hits rise from $0.003625 to $0.022 off-peak and $0.044 at peak.
The cache change is especially important for agent workloads. The cache discount moves from about one-hundred-twentieth of the regular input price to one-thirtieth. For agents that repeatedly retrieve the same files, the source article identifies this as the most expensive part of the pricing shift.
Why The Package Matters
Deepseek is improving V4-Pro for agent use at the same time it is making agent tooling more available. That gives developers a clearer path: use the upgraded model, run it through open-source harness software, and tune integrations around the Responses API and Codex support.
But the pricing change complicates the economics. The new rates partially reverse the price cut Deepseek rolled out in May, and cache hits will cost more than they did before that reduction.
The timing is also notable because the source reports that Deepseek is raising new capital and preparing for an initial public offering. For customers, the practical takeaway is straightforward: V4-Pro appears stronger for agent workflows, Deepseek Harness is now available to inspect and adapt, and API budgets need to be recalculated before August 16 at 4:00 p.m. UTC.