On Aug. 13 DeepSeek released the official general‑availability version of DeepSeek‑V4‑Pro and launched DeepSeek Harness v0.1, an MIT‑licensed open‑source agent harness (dsh). The moves signal that DeepSeek is pushing beyond raw model development into the software layer developers use to run agentic workloads.
What’s new in V4‑Pro and V4‑Flash
The V4 family, introduced in preview in April, includes two main variants:
- DeepSeek‑V4‑Pro: a 1.6‑trillion‑parameter model with roughly 49 billion parameters activated per token.
- DeepSeek‑V4‑Flash: a smaller 284‑billion‑parameter model with about 13 billion activated per token.
Both models support context windows up to one million tokens. DeepSeek says the GA release of V4‑Pro emphasizes improved agent capabilities, and the model now includes native support for the OpenAI Responses API and integration with Codex. The model ID deepseek‑v4‑pro will resolve to the latest V4‑Pro version.
DeepSeek Harness: modular, plugin‑centric agent runtime
DeepSeek Harness (dsh) is built on Cordis, a composable plugin framework whose core idea is that virtually every runtime component can be swapped as a plugin. That includes models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and user interfaces. The project is published as developer preview under the MIT license; the code is available on GitHub and can be launched from npm with npx @deepseek‑ai/dsh web. The repository warns that "THERE WILL BE COMPATIBILITY‑BREAKING CHANGES," indicating it is not yet a stable production drop‑in.
On launch day the repository drew notable attention: roughly 27,500 GitHub stars and about 2,000 forks as of Aug. 13 (a snapshot figure).
Capabilities and positioning relative to Claude Code and Codex
DeepSeek positions Harness as a model‑agnostic alternative to the agent infrastructure behind products such as Anthropic’s Claude Code and OpenAI’s Codex. It already supports inspecting repositories, editing files, executing shell commands, searching files and the web, maintaining plans, invoking skills, delegating to subagents and enforcing approval policies. DeepSeek describes Standard mode as a full coding agent with file editing, shell access, search, planning, subagents and workflows; its local web UI supports workspace selection and approval of sensitive operations.
Compared with Claude Code and Codex, Harness emphasizes exceptional modularity: the model is just another plugin rather than the center of a vertically integrated stack. However, DeepSeek does not claim feature parity—Claude Code and Codex are mature commercial products with broader ecosystems and hosted services that DeepSeek has not yet documented as equivalents.
Benchmarks and reasoning controls
DeepSeek reports that V4‑Pro‑0813 shows improved performance on agent‑oriented tests, but the company notes some results depend on harness configuration. Reported scores for V4‑Pro‑0813 include 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 also clarifies that public Code Agent tasks were tested using the upcoming DeepSeek Harness in "minimal mode," meaning some results reflect model performance inside an agent execution environment rather than a pure model benchmark.
DeepSeek also introduced an explicit "reasoning effort" control with three levels: Non‑think for fast routine tasks, Think High for more complex problem solving and planning, and Think Max for allocating substantially more reasoning to difficult problems. This lets developers tune computation and token use across agent steps.
API pricing overhaul: peak and off‑peak rates, effective Aug. 16
Alongside the product releases, DeepSeek is abandoning its flat API pricing and introducing time‑based peak and off‑peak rates starting Aug. 16 at 16:00 UTC. Peak windows are 01:00–04:00 UTC and 06:00–10:00 UTC; off‑peak hours are priced at half the corresponding peak rate.
Examples of pricing changes (per 1M tokens, cache‑miss input / output):
-
Previous example pricing listed by DeepSeek:
- deepseek‑v4‑flash: $0.14 input / $0.28 output (total $0.42 for 1M in + 1M out)
- deepseek‑v4‑pro: $0.435 input / $0.87 output (total $1.305)
-
New pricing examples:
- deepseek‑v4‑flash: off‑peak $0.22 input / $0.66 output (total $0.88); peak $0.44 / $1.32 (total $1.76)
- deepseek‑v4‑pro: off‑peak $0.66 input / $1.98 output (total $2.64); peak $1.32 / $3.96 (total $5.28)
Cache‑hit prices also rise substantially: for V4‑Pro cache hits move from $0.003625 per 1M tokens to $0.022 off‑peak and $0.044 peak. Depending on model, token category and time of use, Reuters calculated increases ranging from roughly 50% to more than 1,100% over existing rates.
As a concrete example: a workload with 1M cache‑miss input tokens plus 1M output tokens costs $1.305 today on V4‑Pro; under the new tariffs it would cost $2.64 off‑peak (about twice as much) or $5.28 at peak (over four times as much).
Strategic implications
The releases shift DeepSeek’s competitive focus from pure model performance and low token prices toward control of the agent execution layer. By offering an MIT‑licensed harness, DeepSeek aims to shape how models call tools, edit files, maintain sessions and coordinate long‑running workflows—areas where Claude Code and Codex have become important developer products.
At the same time, the sizeable API price increases make workload scheduling, caching strategies and the option to run open models on private infrastructure more salient for organizations considering DeepSeek. The company now pursues two parallel advantages: expanding and opening the agent stack while making its hosted API materially more expensive.
For many enterprise teams the Harness may be the more consequential development in the long run, since the harness that orchestrates tool use, persistence and agent reasoning can be harder to replace than the model behind it. DeepSeek is now competing for that layer as well.
