DeepSeek launches Harness as API prices jump

DeepSeek is moving up the developer stack. On August 13, the Chinese AI lab released DeepSeek Harness v0.1, an open-source agent framework, alongside the general-availability version of DeepSeek-V4-Pro-0813.

As VentureBeat reports, Harness is designed as an alternative to integrated coding-agent products such as Anthropic’s Claude Code and OpenAI Codex. The Next Web likewise describes it as the scaffolding that lets a model inspect files, edit code, browse the web and continue working through a task.

The project is available under the MIT license and can be launched with npx @deepseek-ai/dsh web. It is built on DeepSeek’s Cordis framework, with a deliberately modular architecture summed up by the company’s phrase: “Everything is a plugin.” Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and user interfaces can all be replaced or extended.

Recommended reading

Google’s Gemini 3.7 Flash targets coding agents

Harness can currently:

• Inspect repositories and edit files
• Execute shell commands and search files or the web
• Maintain plans and invoke skills
• Delegate work to subagents
• Enforce configurable approval policies

Its model-agnostic design means developers can connect DeepSeek, Anthropic, OpenAI or compatible custom endpoints. That gives it a wider technical scope than a model-specific coding assistant, but it is not yet a direct replacement for the broader Claude Code or Codex experience. DeepSeek lists the software as a developer preview and warns that “THERE WILL BE COMPATIBILITY-BREAKING CHANGES.”

The repository had approximately 27,500 GitHub stars and 2,000 forks on August 13, although those figures are only a launch-day snapshot, not an adoption measure. For developers comparing agent tools, the contrast with Claude Code’s established terminal, IDE, desktop, browser, mobile and Slack interfaces is significant; our recent survey coverage found that 75% of 138 self-selected developers preferred Claude Code, though Codex scored better on price and usage limits.

DeepSeek API prices rise sharply

The Harness release arrives with a major change to DeepSeek’s API pricing. Starting August 16 at 16:00 UTC, DeepSeek will replace its flat rates with peak and off-peak pricing. Peak periods run from 01:00–04:00 UTC and 06:00–10:00 UTC; all other hours are off-peak, with off-peak rates set at half the new peak prices.

For one million output tokens, the prices will change as follows:

Input prices also increase. V4-Pro cache-miss input rises from $0.435 per million tokens to $0.66 off-peak and $1.32 at peak. V4-Flash rises from $0.14 to $0.22 off-peak and $0.44 at peak.

Cached input becomes substantially more expensive as well: V4-Pro goes from $0.003625 per million tokens to $0.022 off-peak and $0.044 at peak, while V4-Flash rises from $0.0028 to $0.007 and $0.014, respectively.

For a simple workload using one million cache-miss input tokens and one million output tokens, V4-Pro currently costs $1.305. That same mix will cost $2.64 off-peak or $5.28 at peak. V4-Flash moves from $0.42 to $0.88 off-peak and $1.76 at peak. Actual bills will depend on the proportion of cached input, uncached input and generated output.

The new rates remain below several premium alternatives cited by the outlets: V4-Pro’s $3.96 peak output rate is lower than Moonshot’s Kimi K3 at $15 and Anthropic’s Fable 5 at $50 per million output tokens. But the change ends DeepSeek’s earlier position as an exceptionally cheap provider. The Next Web notes that even the new off-peak V4-Pro output rate is more than twice the previous $0.87 price.

V4-Pro adds agent controls and integrations

DeepSeek-V4-Pro-0813 is the official release of a model family first previewed in April. V4-Pro has 1.6 trillion parameters, with 49 billion active per token; V4-Flash has 284 billion parameters, with 13 billion active. Both support context windows of up to one million tokens.

The updated V4-Pro is available through DeepSeek’s website, mobile app and API, with an Expert Mode option in the app and web interface. It adds native support for the OpenAI Responses API, tool calling, JSON output and an Anthropic-format API. DeepSeek also says it is optimized for one-click integration with Codex.

Developers can now choose among three reasoning levels: Non-think for routine work, Think High for more complex planning and problem-solving, and Think Max for difficult tasks. That lets an agent reserve heavier reasoning for debugging or multi-stage changes instead of using it for every file inspection or tool call.

DeepSeek reports V4-Pro-0813 scores of 87.9 on Terminal Bench 2.1, 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard. The company says some public Code Agent results were produced with the upcoming Harness in “minimal mode,” meaning those measurements reflect both the model and its execution environment.

The benchmarks are not directly interchangeable across the two reports. The Next Web cites a separate model-card result of 67.9% on Terminal Bench 2.0, compared with 75.1% for GPT-5.4, as well as 80.6% on SWE-bench Verified. The different benchmark versions and configurations make a simple ranking unreliable, and the reported figures remain vendor-supplied.

DeepSeek’s two releases therefore make a clear strategic statement: the company wants to compete not only on model capability and token prices, but also on the software layer that manages tools, files, permissions, sessions and long-running agent workflows. The immediate trade-off is equally clear—developers get a flexible, open harness, while API users face prices that can be more than four times higher at peak than the rates available before August 16.