DeepSeek announced three related updates on August 12–13, 2026: the open-source agent framework Harness v0.1 under the MIT license, the flagship model V4 Pro reaching general availability, and the introduction of new peak and off-peak pricing starting August 16, which will double the cost of output tokens.

image
image

What happened

DeepSeek Harness v0.1 was released — an open-source agent framework with an 'everything is a plugin' architecture: models, tools, agents, and UI can be swapped via configuration without changing code. The harness operates on top of the Cordis meta-framework (reversible effects model). The repository has gathered 27,900 stars on GitHub. Simultaneously, the DeepSeek V4 Pro model (build 0813) exited preview to GA: Sparse MoE architecture, 1.6 trillion parameters with 49 billion active per token, context up to 1 million tokens. Benchmark results: SWE-bench Verified 80.6%, GPQA Diamond 90.1%, LiveCodeBench 93.5%. Starting August 16, peak and off-peak hour billing is introduced: V4 Pro output will rise from $0.87/million to $1.98/million during off-peak hours and $3.96/million during peak hours (01:00–04:00 and 06:00–10:00 UTC), input token costs will increase by approximately 1.5x.

Context

DeepSeek Harness is the first attempt by a model developer to offer an open agent framework based on the Cordis architecture, which could become the standard for plugin-first agents, similar to how LangChain and Agents became the standard for chain architecture in 2023–2024. V4 Pro weights are published under the MIT license, allowing local deployment and fine-tuning. The sparsity coefficient of the V4 Pro architecture is approximately 32.7:1 (1.6 trillion / 49 billion active), which requires serious routing optimization, especially with a 1 million token context. Prior to this, DeepSeek pursued an aggressive pricing policy: the V4 Flash model processed up to 8 trillion tokens per day, and now the infrastructure has hit GPU capacity limits.

Why this matters for the industry

Raising prices by up to 2.2x for output tokens signals the end of the API market dumping phase: DeepSeek is shifting from price expansion to load management through price elasticity. The peak/off-peak hour model will likely be copied by other providers. The release of V4 Pro weights under MIT and the launch of Harness create a window for startups: solutions can be built on free weights and an open framework, but preparation for increased managed API costs is necessary. If Harness reaches a stable version and gathers a plugin ecosystem, it could become the de facto standard for agent applications.

Why this matters for users

DeepSeek API users need to recalculate budgets by August 16, move batch tasks to off-peak hours, and set up backup providers for fallback. Harness is available for testing right now via the npx @deepseek-ai/dsh web command (Node.js required). It works not only with DeepSeek models but also with OpenAI, Anthropic, Vertex, and other providers via OpenAI-compatible endpoints. V4 Pro weights are available on HuggingFace — local deployment and fine-tuning can begin without being tied to the API.

What is still unknown / limitations

There is no methodological document for V4 Pro: benchmarks are presented as finished figures without a description of testing conditions. It is unknown whether chain-of-thought was used, whether multiple-attempt evaluation was applied, and under what conditions the tests were run. The SWE-bench Verified 80.6% and GPQA Diamond 90.1% results have not been verified by independent laboratories.

Sources

Author

Look at AI, editorial team