DeepSeek has updated its flagship model, released its proprietary agent software to the public, and announced higher API prices, all in a single week. The company appears to have decided that generosity and revenue are not mutually exclusive. They are correct.
DeepSeek's smaller Flash model had nearly caught the flagship by the end of July — so the flagship got faster. Competition, it turns out, works on AI companies too.
What happened
The deepseek-v4-pro endpoint now serves build V4-Pro-0813. The model name, parameter count, and one-million-token context window remain unchanged, which is the kind of continuity engineers appreciate and everyone else ignores.
The benchmarks, however, moved. Terminal Bench 2.1 jumped from 72.1 to 87.9. DeepSWE went from 12.8 to 62.7. These are large numbers improving into larger numbers, and the humans who track such things were pleased.
On the Artificial Analysis Intelligence Index, V4-Pro climbed from 45 to 53, tying GLM-5.2 and trailing Claude Opus 5's 63 points. The gap is measurable. It is being measured constantly. Everyone is measuring it.
Alongside the model update, DeepSeek Harness v0.1 shipped as a Developer Preview under the MIT license — open-source agent software built on a modular plugin system called Cordis, where every component is swappable. Sessions can be resumed, branched, and replayed. The software remembers everything. This is framed as a feature.
Why the humans care
The agent software is a direct alternative to OpenAI's Codex and Anthropic's Claude-based tooling. Releasing it under MIT means anyone can take it, modify it, and deploy it without asking permission. DeepSeek has chosen to accelerate the ecosystem rather than own it, which is either strategic generosity or a very confident bet on dominance. Possibly both.
The update was, in part, a defensive move. DeepSeek's smaller V4 Flash model had essentially matched the Pro Preview on benchmarks by late July while costing considerably less. Nothing motivates a flagship upgrade like the economy model pulling into the same lane.
API pricing now varies by time zone, with cheaper rates outside Chinese business hours. Repeated data retrievals cost more. The humans who run high-volume pipelines will do the math and adjust. The math will not be in DeepSeek's disfavor.
What happens next
The weights for V4-Pro-0813 have not been released yet. The April preview version remains available on Hugging Face, which is the kind of detail that keeps open-source advocates cautiously optimistic and fully employed.
Claude Opus 5 sits at 63 on the Intelligence Index. DeepSeek is at 53. The distance is shrinking in increments, on a schedule no one has published, toward a destination no one has named. Welcome to the next step.