Deepseek Gave Away Its Agent Software and Raised API Prices on the Same Day
Deepseek Harness v0.1 is out under the MIT license, a free open alternative to Codex and Claude Code. On the same day the company announced steep API price increases, with cache hits jumping sixfold.
Deepseek did two things on Thursday that point in opposite directions. It released Deepseek Harness v0.1, the agent software it had been using internally, as open source under the MIT license, which is about as permissive as licences get. And it announced that API prices are going up on August 16, with the steepest increase landing exactly on the kind of usage that agent software produces.
Start with the software. A harness is the layer that turns a language model into something that can actually do work: it gives the model a shell, a file editor, a sandbox to run things in, and a memory of what it has already tried. Without one, a model can only talk. Harness is built on a plugin system called Cordis, where every part is swappable, from the tools to the user interface. It keeps a continuous session log of every prompt, every tool call, and every result, so a run can be paused, branched, or replayed. There is a minimal mode that strips it back to just a shell and a file editor, which is what Deepseek uses for its own benchmark runs. It launches through npx into a local web interface, and Deepseek warns of compatibility issues, which is fair for a v0.1 developer preview. When the team asked for beta testers in early August, 712 projects signed up in three days.
The flagship model got an update at the same time. The deepseek-v4-pro endpoint now serves build V4-Pro-0813, same parameter count, same one-million-token context window. Independent testing by Artificial Analysis moves it from 45 to 53 on the Intelligence Index, which ties GLM-5.2 but still trails Qwen 3.8 Max at 58, Kimi K3 at 60, and Claude Opus 5 at 63. Worth knowing: the weights for this build have not been published, only the older April preview is on Hugging Face.
What is actually going on here
Now the prices. From August 16, Deepseek splits its rates into peak and off-peak, with peak hours matching the Chinese workday. Off-peak, V4-Pro input goes from $0.435 to $0.66 per million tokens and output from $0.87 to $1.98. At peak those double again. The biggest change is cache hits, the discounted rate you pay when the model re-reads text it has already seen: from $0.003625 to $0.022 off-peak, roughly six times more. Agents re-read the same files constantly, so this is precisely the cost that Harness users will feel. Giving away the harness while repricing the fuel it burns is not a contradiction, it is a business model, and Deepseek is raising capital ahead of a stock market listing.
What this means for you: if you are curious rather than building, the useful takeaway is the concept. Harness is a clear, readable example of how a chat model becomes an agent, and it is free to look at. If you are actually using the Deepseek API, check your bill after August 16 and pay attention to your cache hit ratio, because that is where the increase hides. One small consolation for anyone in Europe: peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC, so most of the European afternoon falls under the cheaper rate.
Sources
Source: https://deepseek.com/harness/en/
The Best Model on the Market Is the One Companies Are Not Buying
Spending data from Ramp shows Anthropic's Fable 5 taking just six percent of the tokens companies buy from Anthropic. The most capable model is losing on price, and that says something about the whole industry.
