Skip to content
Today
Sections

Build Independently confirmed4 publishers 3 min read Published Updated

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

Harness v0.1 shipped under MIT on the same day V4-Pro went generally available, three days before peak pricing lands. The lock-in it targets is the runtime, not the weights.

The Engineer · Build desk

Send a correction

What happened

  • DeepSeek released an open-source runtime for building coding agents on August 13th, pushing beyond model APIs into the developer tooling controlled by Anthropic's Claude Code and OpenAI's Codex.
  • DeepSeek's official changelog confirms that V4-Pro reached DeepSeek's app, web interface and API on August 13th.
  • VentureBeat reported that DeepSeek Harness v0.1 entered developer preview on August 13th under the MIT license.
  • DeepSeek will replace its flat international API rates with peak and off-peak pricing at 16:00 UTC on August 16th. Peak periods run 01:00 to 04:00 UTC and 06:00 to 10:00 UTC; all other hours use the off-peak rate.
  • For V4-Pro the pricing change is an increase at every level; even the discounted rate raises cached-input pricing by about 507% and uncached-input pricing by 52%.

Why it matters

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 [4][1][6]. 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 [2][7][10].

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 [9]. 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 [18]. 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` [11][13].

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 [19]. What DeepSeek is shipping is the layer underneath: something you can inspect, fork and rebuild [19]. 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 [20]. The roughly 94,000 stars and 8,600 forks GitHub showed on August 14th measure launch attention, not installations [21].

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 [5].

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 [16]. Cached input is therefore about 120 times cheaper than uncached today [22]. 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 [7][23]. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, seven hours a day, with everything else off-peak [2][17].

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 [8][24]. Whether the rate card settles, since the pricing page still warns of a significant overall increase without a final number [3]. And whether anyone outside DeepSeek reproduces 87.9 on Terminal Bench using the published harness [5][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 [12].

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence55
Adoption
Not scored
Hype gap+25
Incentives70
Confidence58

Perspective Coverage

4 publishers
Builder
Builder 51%
Operator
Operator 31%
Investor
Investor 18%
Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

  1. [1]

    DeepSeek's official changelog confirms that V4-Pro reached DeepSeek's app, web interface and API on August 13th.

    Reported SupportedView cited source
  2. [2]

    DeepSeek will replace its flat international API rates with peak and off-peak pricing at 16:00 UTC on August 16th. Peak periods run 01:00 to 04:00 UTC and 06:00 to 10:00 UTC; all other hours use the off-peak rate.

    Reported SupportedView cited source
  3. [3]

    DeepSeek's pricing page warns that it plans a significant overall API price increase without giving a date or final rate card.

    Reported SupportedView cited source
  4. [4]

    DeepSeek released an open-source runtime for building coding agents on August 13th, pushing beyond model APIs into the developer tooling controlled by Anthropic's Claude Code and OpenAI's Codex.

    Reported SupportedView cited source
  5. [5]

    DeepSeek says the general-availability release scored 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 60.0 on Humanity's Last Exam with tools; these are vendor-reported results and several comparisons include DeepSeek's internal test sets.

    Reported Supported Source: DeepSeek, via runtimewire.comView cited source
  6. [6]

    VentureBeat reported that DeepSeek Harness v0.1 entered developer preview on August 13th under the MIT license.

    Reported Supported Source: VentureBeat, via runtimewire.comView cited source
  7. [7]

    For V4-Pro the pricing change is an increase at every level; even the discounted rate raises cached-input pricing by about 507% and uncached-input pricing by 52%.

    Reported SupportedView cited source
  8. [8]

    Simon Willison wrote that the latest DeepSeek Pro model is available via API only and that he had to link to OpenRouter because DeepSeek had no obvious announcement page for the new model.

    Reported Supported Source: Simon WillisonView cited source
  9. [9]

    DeepSeek Harness, also called dsh, is built on the Cordis plugin framework, and its core design rule is that practically every component can be replaced: models, tools, skills, sessions, sandboxes, storage, agent loops, orchestration and user interfaces.

    Reported SupportedView cited source
  10. [10]

    The Harness release on August 13th precedes the August 16th pricing change by three days.

    Derived Supported
  11. [11]

    Harness standard agent mode can inspect and edit files, run shell commands, search local files and the web, maintain a plan, invoke reusable skills, delegate tasks to subagents and apply approval policies; DeepSeek also provides a Python SDK and modes for tool composition, stripped-down benchmarking and custom agent creation.

    Reported SupportedView cited source
  12. [12]

    DeepSeek's official documentation does not disclose a change log, benchmark delta, training update or exact rollout timestamp for the 0813 build; because applications calling the moving deepseek-v4-pro alias may now receive it, teams should record the returned model version and repeat critical regression tests.

    Reported SupportedView cited source
  13. [13]

    Developers can start a local web interface with the command npx @deepseek-ai/dsh web.

    Reported SupportedView cited source
  14. [14]

    In its July 31st V4-Flash release note, DeepSeek said its public code-agent evaluations used the then-unreleased Harness in minimal mode, configured with only a persistent shell and a file-editing tool.

    Reported SupportedView cited source
  15. [15]

    Publishing Harness gives developers a path to inspect and reproduce more of the runtime around the model, where tool definitions, prompts and execution loops can materially affect agent scores.

    Reported SupportedView cited source
  16. [16]

    DeepSeek's API pricing page lists V4 Pro at $0.003625 per million cached input tokens, $0.435 per million uncached input tokens and $0.87 per million output tokens, with a 1 million-token context window and a 384,000-token maximum output.

    Reported SupportedView cited source
  17. [17]

    The stated peak windows total seven hours per day, leaving seventeen off-peak hours.

    Derived Supported
  18. [18]

    The model is another plugin in Harness; DeepSeek's provider documentation supports Anthropic, OpenAI, Codex, Amazon Bedrock, Google Vertex AI, Microsoft Azure and custom endpoints alongside DeepSeek's API, making Harness model-agnostic while giving DeepSeek a distribution channel for V4-Pro and V4-Flash.

    Reported SupportedView cited source
  19. [19]

    Harness occupies a different product tier from Claude Code and Codex: Anthropic and OpenAI have built managed coding products across terminals, development environments, cloud tasks and repository workflows, while DeepSeek is shipping a composable runtime and local interface that developers can inspect, fork and rebuild.

    Reported SupportedView cited source
  20. [20]

    DeepSeek warns that Harness remains a developer preview and will receive compatibility-breaking changes; the MIT license removes a legal barrier to commercial use while the unstable interfaces raise the engineering cost of deploying it in production today.

    Reported SupportedView cited source
  21. [21]

    GitHub displayed about 94,000 stars and 8,600 forks for the Harness repository on August 14th; those figures measure launch attention rather than active installations or production adoption.

    Reported SupportedView cited source
  22. [22]

    Cached input for V4 Pro is currently about 120 times cheaper per million tokens than uncached input.

    Derived Supported
  23. [23]

    Applying the stated discounted-rate increases to the current card puts off-peak uncached input near $0.66 per million tokens and cached input near $0.022 per million, narrowing the cached-to-uncached ratio to roughly 30x.

    Derived Supported
  24. [24]

    Simon Willison said he had not been able to confirm whether DeepSeek plans to release open weights for the new build, noting that weights are available for April's DeepSeek-V4-Pro and July's DeepSeek-V4-Flash-0731.

    Reported Contested Source: Simon WillisonView cited source

Sources

4 independent publishers whose own reporting we read for this story.

  1. dev.to

    2 articles · August 16, 2026

    DeepSeek V4 Pro 0813 发布：新一代混合推理大模型带来哪些升级与行业影响
  2. letsdatascience.com

    1 article · August 12, 2026

    DeepSeek Lists V4 Pro 0813 Build in Its API Catalog
  3. runtimewire.com

    1 article · August 14, 2026

    DeepSeek open-sources coding-agent harness before V4-Pro API price hike
  4. simonwillison.net

    1 article · August 12, 2026

    DeepSeek V4 Pro 0813 (on OpenRouter)

Topics and entities

Topics

Entities