DeepSeek released DeepSeek-V4-Flash-0731 to public beta on July 31, 2026, as an official release of the model that beat V4-Pro-Preview on all nine published agent benchmarks, with Terminal Bench 2.1 scoring 82.7 at the same $0.14 per 1M input and $0.28 per 1M output pricing.
The V4-Flash-0731 uses the exact same architecture and size as the April preview version (284B total params, ~13B active in Mixture-of-Experts)—only the post-training was redone. Agent benchmarks included Cybergym 76.7, NL2Repo 54.2, DeepSWE 54.4, Toolathlon verified 70.3, Agent Last Exam 25.2, Automation Bench 25.1, DSBench-FullStack 68.7 and DSBench-Hard 59.6.
The official V4-Flash natively supports the Responses API format and has been specifically adapted for Codex, allowing direct integration without translation shim. At $0.14 per million input tokens, V4-Flash is the most cost-effective agent model on the market, delivering 82.7 on Terminal Bench 2.1 versus Opus-4.8's 85.0, a gap of just 2.3 points, at a fraction of the price. For builders: a cheaper model beating the expensive tier on agentic work through post-training signals a narrowing quality gap on standard agent tasks. Verify against your own workload before swapping.