This week DeepSeek did two things at once: it open-sourced Harness, its first agent runtime, and it switched V4 API billing to peak and off-peak tiers. One hand gives, the other takes. If you're building on DSH — or just watching this space — both moves are worth understanding, because they tell you where DeepSeek is heading.
The big story: Harness is open source
Late on August 13, DeepSeek pushed Harness (dsh) to GitHub under the MIT license, as a v0.1 developer preview. The whole pitch is a single sentence: everything is a plugin. Models, tools, the agent loop, memory, context — each one a swappable component.
It crossed 33,000 stars within hours and roughly 50,000 within twelve. People keep reaching for "the Android of agents"; we'd only half-agree. Android is an operating system — Harness goes a step further and makes the operating system itself a set of replaceable parts.
Underneath is Cordis, a meta-framework built around three properties: observable, composable, rollback-able. That last one is the part we keep thinking about — an agent that can rewind is an agent you can actually let rewrite itself.
The week, in order
The late-night triple drop: the V4 pricing announcement, the Harness open-source release, and the companion paper, all at once.
V4-Pro moves out of preview into commercial use. HPCwire, The Register, and VentureBeat all weigh in.
The paper deep-dives arrive — the headline framing: letting agents rewrite themselves at runtime.
The community starts moving: third-party projects like dsh-handbook and dsh-continual-evolve appear.
The V4 peak/off-peak pricing takes effect. Peak windows (09:00–12:00 and 14:00–18:00) cost the most; off-peak runs at half price.
The price change, because it actually matters
From the 17th, V4 drops flat pricing for time-of-day tiers: peak hours cost the most, off-peak runs at half. Some tiers land roughly 11x their original price, which is why "the end of cheap AI" has been floating around — for once, that's not just a headline.
For anyone who calls these APIs daily, this is the most concrete change of the week. Retiming workloads off-peak stops being a money-saving trick and becomes a scheduling necessity. If your jobs can tolerate a queue, move them.
Worth reading, in the order we'd open them
Official — V4-Pro GA announcement (api-docs). The authoritative one for pricing and the hard specs: 384K context, tool calling, Responses API, FIM.
HPCwire — "DeepSeek Open-Sources the Missing Layer Between AI Models and Agents". The clearest framing of the week.
The Register — "DeepSeek's innovative harness treats everything as a plug-in". The mainstream-tech read on the core design.
VentureBeat — Harness as an open-source Claude Code rival. Bluntest on the commercial angle.
机器之心 (Machine Intelligence) — the companion paper, "letting agents rewrite themselves". Read this one for the self-evolution direction.
澎湃 (The Paper) — "it works, but you have to watch it". A rare balanced take that doesn't hype.
品玩 (PingWest) — "taking down Claude Code the Minecraft way". The most entertaining metaphor of the week.
Key numbers
Our take
Reading the open-source move as pure PR — "free stuff right after a price hike" — misses the point. When models, tools, and the loop are all plugins, developers can swap models at will. DeepSeek is betting on ecosystem lock-in, not model lock-in.
That said, preview is preview. "It works, but you have to watch it" is the honest summary. Completion, stability, and documentation are all early. Our advice: play with it, internalize the plugin mindset, and don't ship it to production yet.
On pricing: time-of-day tiers will push compute into off-peak windows. If your workloads tolerate latency, retime your jobs — the savings are real, and this is a change you can act on today.