Deepseek V4 Pro Update Outperforms Flash, Weights Withheld

Deepseek pushed out a new build of its flagship model this week, and the timing says as much as the specs. V4-Pro-0813 replaces the testing version that's been running on the deepseek-v4-pro endpoint, with Terminal Bench 2.1 scores climbing from 72.1 to 87.9 and DeepSWE jumping from 12.8 to 62.7, according to the company's own comparison table. Parameter count and the one-million-token context window stay the same, and Deepseek says current integrations won't need any changes.

Alongside the model update, Deepseek open-sourced "Deepseek Harness," the plugin system it uses internally to turn its language models into agents that can act on their own. That's a notable move for a company that's kept its agent tooling close until now. The release lands as Deepseek also raises API prices, introducing time-based rates that reward usage outside Chinese business hours while making repeated data pulls more expensive.

Third-party benchmarking from Artificial Analysis paints a more measured picture than Deepseek's internal numbers, putting the gains in context against rivals like Claude Opus 4.8. That comparison is worth sitting with before deciding how much V4-Pro has actually closed the gap.