DeepSeek Open-Sources Its Agent Harness — and Raises API Prices on the Same Day

DeepSeek shipped three things simultaneously on August 13th: a significantly better V4 Pro model, an MIT-licensed agent harness, and API price increases of up to 4×. The combination tells you something about where DeepSeek’s strategy has landed — and the tradeoffs are real for every team currently building on its API.

What Happened

The model: V4-Pro-0813 moves out of preview and into general availability with substantial benchmark gains. On Terminal Bench 2.1, the score jumped from 72.1 to 87.9. On DeepSWE — the agentic software engineering benchmark that now matters more than most — it went from 12.8 to 62.7. The Artificial Analysis Intelligence Index climbed from 45 to 53, placing it behind Claude Opus 5 (63 points) but competitive with the field of frontier-class models. The release also adds native OpenAI Responses API support and adjustable reasoning effort (low, high, and max settings).

The agent software: DeepSeek Harness v0.1 is now open under MIT license. Built on the Cordis plugin system, it treats every component — tools, sandboxes, session handlers, and UI — as a swappable plugin. The harness tracks prompts, tool calls, and results, and critically supports resuming, branching, and replaying runs. 712 projects signed up for beta testing within three days of the initial callout.

The prices: Effective August 16, DeepSeek is introducing peak/off-peak pricing tied to Chinese business hours. V4-Pro input goes from $0.435 to $0.66 (off-peak) or $1.32 (peak) per million tokens. Output goes from $0.87 to $1.98 or $3.96. Cache hits take the steepest hit — the discount drops from one-hundredth to one-thirtieth of input rates, which meaningfully reduces the economics of prompt caching for high-frequency workloads.

Why It Matters

The open-source Harness release is the most strategically significant of the three moves. The agentic coding space is consolidating around a small number of tool frameworks, and the team that ships a credible, composable harness under a permissive license owns developer mindshare in a way that benchmark rankings alone do not.

The resume-branch-replay capability specifically deserves attention. Most agent frameworks treat runs as atomic: they either complete or fail. A harness that supports branching and replaying creates a workflow model much closer to how engineers actually debug agent failures — which is to say, iteratively, with access to intermediate state. That’s table stakes for any framework that wants to be used in production rather than demos.

The price increase simultaneously signals that DeepSeek has moved past the phase where subsidized pricing was the primary competitive lever. At $1.32/million input tokens at peak, it’s no longer dramatically cheaper than Claude Sonnet-class models for inference-heavy workloads. The historical DeepSeek price disruption narrative — which was always partly provisional — now has a clear expiration date.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.