The upgraded DeepSeek-V4-Flash is here: the model's API enters public beta with the same architecture as the April preview but a new post-training round, shifting the offering's focus from raw cost to autonomous execution. DeepSeek didn't launch a new model — it made the existing one far more capable, aiming straight at the agent market. Details are in the official API changelog.

Psst!Add AlexTech.ai asPreferred Source on Google

V4-Flash overtakes the old V4-Pro-Preview

The striking figure is the inverted internal hierarchy. The official version of the lightweight model, dubbed DeepSeek-V4-Flash-0731, keeps the preview's structure and size (284B total parameters, 13B active) but received retraining focused on agentic capabilities. On several coding-agent benchmarks it beats V4-Pro-Preview, the flagship from three months ago: 82.7 on Terminal Bench 2.1, 76.7 on Cybergym, 70.3 on Toolathlon, 68.7 on DSBench-FullStack, 54.4 on DeepSWE and 54.2 on NL2Repo.

It's a break from the "Pro strong, Flash weak" logic: the cheap model outguns the previous-generation flagship. But the overtake is probably temporary — the two aren't competing within the same generation. The defeated V4-Pro-Preview is three months old, and when the official V4-Pro ships (with 1.6 trillion total parameters and 49B active), the hierarchy is likely to flip back, returning Flash to the role of the efficient option.

Responses API and Codex: the gateway to agents

DeepSeek's move has a clear target: holding the agentic application market with a cost-effective model. The official version natively supports the Responses API format and ships an adaptation for Codex, OpenAI's coding assistant, with the first public appearance of DeepSeek's own Harness as a test framework. For developers, that means migrating OpenAI-built applications with minimal changes — and for DeepSeek, a direct attack on the most profitable segment of enterprise AI.

Only the Flash API changed

The beta leaves the rest of the lineup untouched: the update applies only to the V4-Flash API (model name deepseek-v4-flash, same base URL). The V4-Pro API and the models behind the app and website are unchanged.

Outlook: the new V4-Pro could flip everything

The Flash upgrade fits DeepSeek's strategy of holding the agent segment with a light, fast model. But the official V4-Pro is expected soon, and once the updated flagship arrives it's plausible the internal balance returns to the traditional order: Pro on top, Flash as the efficient alternative. Today's overtake reads as a snapshot of a transition, not the new hierarchy.