Quick Facts
- DeepSeek released Harness v0.1, an MIT-licensed open-source agent runtime, on August 13, 2026, earning more than 33,000 GitHub stars within hours.
- DeepSeek-V4-Pro reached general availability on the same date, ranking second on SWE-bench Verified at 96.40%, behind only Claude Opus 5.
- DeepSeek is raising API prices for V4-Flash and V4-Pro by between 50% and more than 1,100% depending on the tier.
DeepSeek shipped two products on August 13, 2026: a new open-source agent runtime called Harness and the general availability release of its V4-Pro model. The double launch comes with a significant asterisk for developers already building on the platform. API prices are going up, in some cases by more than ten times.
Harness v0.1, codenamed Black Whale, is built under the MIT license and positions itself as a self-hostable alternative to closed coding-agent environments. The main target is clear: Anthropic’s Claude Code. The project is led by Cui Tianyi, who joined DeepSeek in March 2026 after working at Jane Street.
How Harness Works
The runtime sits between a large language model and the tools, files, and environments an agent needs to complete multi-step tasks. Its architecture is built on a framework called Cordis, where every capability — models, tools, sessions, sandboxes, storage, and scheduling — is a swappable plugin. Developers can change any component through configuration without touching the Harness source code.
The product ships with four preset modes. Standard mode gives developers filesystem tools, shell access, web search, subagents, and a planning feature. Minimal mode strips the toolkit down to two tools: bash and a string-replace editor. Code mode and Creator mode round out the options, with the latter allowing developers to inspect the runtime and build new modes from existing plugins.
A session transparency feature logs everything the model sees in an append-only record: system prompts, reasoning, tool calls, results, and subagent scheduling. Developers can inspect, replay, fork, or resume any session from that event stream.
The release drew 608 points and 261 comments on Hacker News by August 14, making it one of the top developer discussions of the week. DeepSeek warned in the repository that compatibility-breaking changes will come. There is no enterprise support tier, no SLA, and no published compatibility guarantees for non-DeepSeek models.
V4-Pro Benchmark Results and Discrepancies
DeepSeek-V4-Pro ended a preview period that began in April. The company reports a 96.40% score on SWE-bench Verified, placing it second overall behind Claude Opus 5, at roughly one-thirty-sixth the cost per test. The model also supports the Responses API and integrates with Codex, with three selectable reasoning levels.
Independent results complicate the picture. On Terminal-Bench, DeepSeek reports 87.9% using its own internal harness. The reference Terminus 2 harness scores the same model at 54.68% — a 33-point gap. That score also falls 12 points below DeepSeek’s cheaper V4-Flash model, which scored 67.04% on the same benchmark. On LiveBench agentic coding, both V4-Pro and V4-Flash ranked last among seven frontier models tested, scoring 54.95 and 46.77 respectively against Claude Opus 5’s 65.20.
DeepSeek had not released its benchmark harness publicly before August 13, meaning its internal figures could not be replicated. The open-sourcing of Harness changes that, at least in principle.
Price Increases Hit Developers
The launch comes alongside price hikes on DeepSeek’s V4-Flash and V4-Pro API tiers. Increases range from 50% on the low end to more than 1,100% on certain tiers. Developers who built cost models around DeepSeek’s historically low prices will need to recalculate. The MIT license covers the Harness runtime. It does not cover the model API costs underneath it.
For software founders evaluating DeepSeek as a production dependency, the combination of rising API costs, benchmark discrepancies on third-party evaluations, and a v0.1 runtime with no backward-compatibility guarantees adds meaningful risk to any near-term adoption decision. As a technology watch item, Harness is among the more significant open-source agent releases of 2026.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.