DeepSeek has expanded the scope of its competition in AI tools for developers by launching the official version of the DeepSeek-V4-Pro-0813 model, alongside releasing DeepSeek Harness v0.1 as a developer preview and an open-source project under the MIT license. Harness gives developers a customizable framework for building coding agents that interact with repositories, files, shell commands, and long-running workflows, rather than merely using the model itself.
V4-Pro is currently available through DeepSeek’s web interface, mobile app, and API. The company has added native support for the OpenAI Responses API, along with integration with Codex, and the model is now available in “Expert Mode” on the website and app. API users do not need to change the model identifier, as deepseek-v4-pro now points to the latest version of V4-Pro.
A Replaceable Agent Framework
DeepSeek describes the Harness project, also known as dsh, as an open-source framework built on Cordis, a framework based on composable plugins. The company summarizes its philosophy with the phrase “everything is a plugin,” allowing models, tools, skills, sessions, isolated environments, file systems, execution loops, orchestration, and user interfaces to be replaced.
In its current form, Harness supports repository inspection, file editing, shell command execution, file and web search, planning, skill invocation, delegating tasks to sub-agents, and enforcing approval policies. DeepSeek says Standard mode functions as a full coding agent, with a local web interface that allows users to select a workspace and approve sensitive operations.
The code is available on GitHub, and the interface can be run through npm using the command npx @deepseek-ai/dsh web. DeepSeek also provides instructions for building from source. However, the repository explicitly describes the software as a “developer preview” and warns that breaking changes may occur, so the company does not yet present it as a stable production platform. Harness is also not a complete replacement for the Claude Code or Codex experience, but rather an open, model-neutral framework focused on assembling the infrastructure surrounding the agent.
V4-Pro Focuses on Agent Tasks
The August 13 release marks V4-Pro’s transition from the preview stage to the official version, rather than the first appearance of the V4 family, which DeepSeek introduced in April. The lineup consists of V4-Pro with 1.6 trillion parameters, 49 billion of which are activated per token, and V4-Flash with 284 billion parameters, 13 billion of which are activated. Both models support a context window of up to one million tokens.
DeepSeek says the new version delivers major improvements in agent capabilities, particularly in production environments, and supports tool calling, JSON output, the Responses API, and an Anthropic-format API for both V4-Pro and V4-Flash. The company has also added clear controls for inference effort: Non-think for fast routine tasks, Think High for more complex problems and planning, and Think Max for tasks requiring intensive reasoning.
According to figures announced by DeepSeek, V4-Pro-0813 recorded 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 explains that some Code Agent task results were tested using Harness in “minimal mode,” meaning they measure the model inside an agent execution environment rather than independently measuring the model alone.
New API Pricing Based on Usage Times
Alongside the launch, DeepSeek announced that it would abandon uniform API pricing starting August 16 at 16:00 UTC. Peak periods will run from 01:00 to 04:00 UTC and from 06:00 to 10:00 UTC, while the remaining hours will be classified as off-peak. Off-peak prices will be half the new peak prices.
V4-Pro’s price per million uncached input tokens will rise from the current $0.435 to $0.66 off-peak and $1.32 during peak periods, while the output price will rise from $0.87 to $1.98 and $3.96, respectively. For V4-Flash, input pricing will rise from $0.14 to $0.22 off-peak and $0.44 during peak periods, while output pricing will rise from $0.28 to $0.66 and $1.32.
Cached-token prices will also increase: V4-Pro will rise from $0.003625 per million tokens to $0.022 off-peak and $0.044 during peak periods, while V4-Flash will rise from $0.0028 to $0.007 and $0.014. Based on one million uncached input tokens and one million output tokens, V4-Pro’s cost will rise from the current $1.305 to $2.64 off-peak and $5.28 during peak periods, with the actual cost varying according to caching and output ratios.
The new releases put DeepSeek in direct competition for the operating layer of coding agents, rather than only for model quality and token prices. At the same time, the early Harness preview and API increases mean teams will need to evaluate framework stability, caching behavior, and task scheduling, alongside the option of running the open weights on their own infrastructure when calculating total costs.