DeepSeek open-sources the harness, then raises the price of the model

DeepSeek published Harness v0.1, an MIT-licensed runtime for building coding agents, on August 13th, the same day its changelog confirms DeepSeek-V4-Pro reached the app, web interface and API ^[1][2][3]. Three days later, at 16:00 UTC on August 16th, the company replaces flat international API rates with peak and off-peak pricing that is an increase at every level for V4-Pro ^[4][5][6].

The part that changes an operator's options is not the model. Harness, also called dsh, sits on the Cordis plugin framework under one design rule: practically every component can be replaced, including models, tools, skills, sessions, sandboxes, storage, agent loops, orchestration and user interfaces ^[7]. The model is itself a plugin, and the provider documentation covers Anthropic, OpenAI, Codex, Amazon Bedrock, Google Vertex AI, Microsoft Azure and custom endpoints alongside DeepSeek's own API ^[8]. Standard agent mode reads and edits files, runs shell commands, searches locally and on the web, keeps a plan, calls reusable skills, delegates to subagents and applies approval policies; there is a Python SDK and a local web interface via `npx @deepseek-ai/dsh web` ^[9][10].

That is not the same product as Claude Code or Codex, which Anthropic and OpenAI have built out across terminals, development environments, cloud tasks and repository workflows ^[11]. What DeepSeek is shipping is the layer underneath: something you can inspect, fork and rebuild ^[11]. DeepSeek says Harness is a developer preview and will take compatibility-breaking changes, so the MIT license removes the legal barrier while the unstable interfaces keep the engineering cost high ^[12]. The roughly 94,000 stars and 8,600 forks GitHub showed on August 14th measure launch attention, not installations ^[13].

There is a second reason to want the runtime published. In its July 31st V4-Flash note, DeepSeek said its public code-agent evaluations ran on the then-unreleased Harness in minimal mode, configured with only a persistent shell and a file-editing tool ^[14]. Tool definitions, prompts and execution loops move agent scores materially, so publishing the harness is the difference between reading a vendor number and reproducing it ^[15]. DeepSeek reports 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 60.0 on Humanity's Last Exam with tools for the GA build, all vendor-reported, with several comparisons drawing on internal test sets ^[16].

The pricing side is where the harness earns its keep. The current card lists $0.003625 per million cached input tokens, $0.435 uncached and $0.87 output, against a 1 million-token context and 384,000-token maximum output ^[17]. Cached input is therefore about 120 times cheaper than uncached today ^[18]. After August 16th, DeepSeek says even the discounted rate raises cached input by about 507% and uncached input by 52%, which puts off-peak uncached near $0.66 and cached near $0.022, narrowing the caching advantage to roughly 30x ^[5][19]. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, seven hours a day, with everything else off-peak ^[4][20].

Watch three things. Whether the 0813 weights follow: Simon Willison notes weights exist for April's V4-Pro and July's V4-Flash-0731 but could not confirm plans for this build, and had to link OpenRouter because DeepSeek published no obvious announcement page ^[21][22]. Whether the rate card settles, since the pricing page still warns of a significant overall increase without a final number ^[23]. And whether anyone outside DeepSeek reproduces 87.9 on Terminal Bench using the published harness ^[16][15]. Meanwhile, `deepseek-v4-pro` is a moving alias with no published change log for 0813, so log the returned model version and rerun your own regressions ^[24].