Chinese AI firm DeepSeek ends the V4-Pro preview and ships production-ready agent tools today, just as steeper API rates take effect.
DeepSeek moved its flagship V4-Pro model to full general availability this week with the V4-Pro-0813 checkpoint. The update lands alongside an open-source agent framework and a sharp rise in API pricing that begins at 16:00 UTC on August 16.
Stronger agent performance
The new build focuses on real-world agent work: tool use, multi-step coding, and long-horizon tasks. DeepSeek reports big jumps over the April preview. Terminal Bench 2.1 rose from 72.1 to 87.9. DeepSWE climbed from 12.8 to 62.7. Other agent scores, including NL2Repo and Cybergym, also improved markedly.
Developers can now dial reasoning effort to low, high, or max. The company recommends high for everyday agent jobs. The model keeps its 1-million-token context window and supports native OpenAI Responses API plus one-click Codex setup. It remains available in the app and web under Expert Mode, with the same API model name.
We’re launching DeepSeek-V4-Pro today! 🚀
— DeepSeek (@deepseek_ai) August 13, 2026
🔷 Major Agent upgrades with strong production gains!
🔷 Flexible reasoning effort for V4-Pro & V4-Flash: low for simple tasks, high for daily Agent workflows, max for complex tasks.
🔷 Native OpenAI Responses API support, optimized for… pic.twitter.com/ZsGbUpiu1T
Open-source agent harness
On the same day, DeepSeek released DeepSeek Harness v0.1 under the MIT license. The developer-preview framework treats every component as a plugin: models, tools, sessions, sandboxes, and UI. Built on the Cordis system, it lets teams swap pieces freely and aims to serve as an alternative to closed agent runtimes such as OpenAI’s Codex.
Users can launch it with a single command (npx @deepseek-ai/dsh web) or clone the GitHub repo. DeepSeek positions the combination of model plus harness as the practical path to production agents.
🧩 DeepSeek Harness v0.1 is now available in Developer Preview!
— DeepSeek (@deepseek_ai) August 13, 2026
🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license.
🔹 Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one…
Price hikes and peak pricing
The performance gains come with higher costs. From 16:00 UTC August 16, DeepSeek switches to peak and off-peak rates. Peak hours (roughly aligned with Chinese business hours) charge roughly double the off-peak price.
For V4-Pro, off-peak rates move to about $0.66 per million cache-miss input tokens and $1.98 for output. Peak rates hit $1.32 and $3.96. Cache-hit prices also rise sharply. The old flat rates of $0.435 / $0.87 are gone. Similar increases apply to V4-Flash.
DeepSeek says the change helps allocate capacity more efficiently. Even at the new peak levels the model remains far cheaper than many closed competitors, yet the jump is large enough that teams running heavy agent loops will feel it.
What it means for developers
The release marks DeepSeek’s clearest push into production agent infrastructure. Better scores, selectable reasoning depth, open tooling, and native API compatibility lower the barrier to building autonomous coding and workflow systems. The higher prices, however, signal that sustained high usage will no longer come at the previous bargain rates.
Developers who can shift workloads to off-peak hours or lean on caching will limit the impact. Those building long-running agents should re-evaluate cost projections immediately.
The combination of stronger agent features and open harness code gives the open-weight community a serious new option. Whether the higher prices slow adoption or simply fund the next round of progress will become clear in the coming weeks.
AI Disclosure: This article was created with the assistance of artificial intelligence tools and was reviewed and edited by the Glowls News editorial team before publication.