Analysis
DeepSeek released a retrained build of its V4-Flash model on Thursday, designated V4-Flash-0731, that outperforms the company's own more expensive V4-Pro-Preview flagship on all nine agent and coding benchmarks DeepSeek has published -- including an 82.7 score on Terminal Bench 2.1 versus V4-Pro-Preview's 72.1, a 14.7% improvement despite the budget model costing a fraction of the flagship to run.
What makes the release notable is what didn't change: V4-Flash-0731 keeps the identical 284-billion-total, 13-billion-active-parameter mixture-of-experts architecture DeepSeek shipped with the original V4-Flash-Preview back in April. The entire performance gain came from retraining the same-size model rather than scaling it up, a data point that cuts against the industry's dominant assumption that better benchmark scores require bigger models or new architectures.
The release also adds native support for the Responses API format and full compatibility with Codex, a deliberate design choice that lowers the switching cost for developers already building against OpenAI's API conventions. Pairing a cheaper, higher-performing model with drop-in compatibility for a competitor's tooling is a direct play for developer migration rather than just a benchmark headline.
The update is scoped narrowly: it applies only to the V4-Flash API interface, while DeepSeek's V4-Pro API and its official consumer app and web products remain unchanged for now. That leaves an unusual situation where DeepSeek's cheaper model is publicly benchmarked ahead of its own premium tier, at least until the company presumably applies the same retraining approach upstream.
For the broader AI market, the release reinforces a pattern that has defined 2026's open-model competition: Chinese labs continuing to compress the gap to closed Western frontier models on price and efficiency even without larger training runs, echoing the DeepSeek playbook that first rattled markets over a year ago. What to watch: whether DeepSeek applies the same retraining gains to V4-Pro, and how quickly Western labs respond on pricing given a competitor is now beating its own flagship with a cheaper, unchanged-size model.
