Overview
- DeepSeek launched the DeepSeek‑V4‑Flash official‑version API public beta on July 31, 2026, and published a nine‑item scorecard showing large gains on Agent‑style tasks such as terminal operations, repo analysis, security testing and full‑stack development.
- The firm says the V4‑Flash‑0731 update did not change model architecture or size but applied targeted post‑training to the preview model, and the company disclosed test settings using its soon‑to‑be‑released DeepSeek Harness framework.
- The V4‑Flash official build now natively supports the Responses API format and is adapted for Codex to simplify integration for developers, while DeepSeek‑V4‑Pro API and app/web models were left unchanged and V4‑Pro official release is pending.
- Independent platforms updated evaluations showing competitive scores — Artificial Analysis gave V4‑Flash‑0731 an AII of 50 — and DeepSeek’s claimed ~98% cache‑hit discount on its API was cited as making per‑task costs about 60% lower than a top rival in some comparisons.
- The release highlights a broader industry shift toward optimizing agent‑level workflows and standardized agent evaluation, which may push vendors to favor post‑training and tool‑use tuning over raw model scale going forward.