DeepSeek has moved the official DeepSeek-V4-Flash API into public beta, with the same model name developers already use: deepseek-v4-flash.
The July 31 change log says the API calling method is unchanged. Developers set the model name to deepseek-v4-flash to use the latest version. The update is focused on the API, not DeepSeek’s app or web model surfaces, and the V4-Pro API is unchanged for now.
The bigger developer change is format support. DeepSeek says V4-Flash now natively supports the Responses API format and has been adapted for Codex. A separate Responses API guide says support currently applies to deepseek-v4-flash, uses https://api.deepseek.com as the base URL, and that V4-Pro support is planned for early August 2026.
DeepSeek also says V4-Flash-0731 keeps the same architecture and size as V4-Flash-Preview and was only re-post-trained.
The benchmark claim is about agents
DeepSeek’s own change log frames the update around agent capability. It reports higher results than V4-Pro-Preview across Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon verified, Agent Last Exam, Automation Bench Public, DSBench-FullStack, and DSBench-Hard.
Those numbers should be treated as DeepSeek-reported until independent replications appear. The footnote matters too: for public code-agent benchmarks, DeepSeek says the official model was tested with its DeepSeek Harness minimal mode, max effort level, top_p=0.95, and temperature 1.0. Two listed DSBench sets are internal.
The current The AI Feed models page now includes DeepSeek V4 Flash 0731 as a ranked entry from Artificial Analysis. In the refreshed dataset for this run, DeepSeek V4 Flash 0731 appears at rank 21 with an intelligence index of 49.9 and a listed blended price of about $0.058 per million tokens.
The narrow change is the useful one
This is not a new public app launch, and it is not the official V4-Pro release. It is a coding-agent API release with a compatibility story.
That makes it easier to test. Teams already using Responses API-shaped agent loops can compare V4-Flash against their current coding model without rewriting the whole harness. The first checks should be boring and concrete: accepted patches, test pass rate, tool-call failures, total tokens, retry count, and reviewer edits.
For buyers, the key question is whether the low listed cost survives real agent workloads. Cheap tokens help only if the model can finish multi-step work without adding failed attempts, brittle tool calls, or human cleanup.
