DeepSeek Opens V4-Flash-0731 API Public Beta for Agent Workloads
The re-post-trained API model keeps the preview architecture and endpoint while adding native Responses API support and a sharper agent focus.
DeepSeek released DeepSeek-V4-Flash-0731 into public beta through its API on July 31. The company says developers can continue calling `deepseek-v4-flash`, which now resolves to the updated build without a new invocation pattern. The change matters most to teams running coding and tool-using agents because DeepSeek positions this version as an agent-focused update rather than a new consumer-chat rollout.
What changed
DeepSeek says V4-Flash-0731 keeps the same model architecture and size as V4-Flash-Preview. The distinction is additional post-training, not a new underlying architecture. According to the official DeepSeek API changelog, the served model also supports the Responses API format natively and has been adapted specifically for Codex-style workflows.
That combination lowers migration friction for existing API users. Applications already sending the `deepseek-v4-flash` model name should receive the latest version through the unchanged call pattern, while developers building agent loops can use the Responses API surface rather than creating a separate compatibility layer.
DeepSeek reports substantial gains across a group of terminal, repository, cybersecurity, tool-use, automation and full-stack agent evaluations. Those results are vendor-reported, however. The company also notes that its public code-agent tests used a minimal DeepSeek Harness configuration with maximum effort, and that two cited coding tests are internal datasets. Independent results may differ when developers use other harnesses, prompts, budgets or execution environments.
What is unchanged
This is an API-only update. DeepSeek says V4-Pro and the models served through its app and web products are unchanged. V4-Flash-0731 also retains the preview model's architecture and size, so the release should be read as a targeted serving and post-training revision rather than a full new model generation.
DeepSeek says an official V4-Pro release will follow, but it gives no date on the changelog. That future release is not part of this confirmed update.
Why it matters
Model updates that preserve an alias can improve capability without forcing immediate application changes, but they also make regression testing important: the same identifier can begin returning a materially different build. Teams using DeepSeek for production agents should therefore re-run their own task suites and safety checks before assuming the vendor's reported benchmark gains transfer to their workloads.
Status
Confirmed. DeepSeek's official API documentation records the July 31 public beta and its scope. Internal confidence is medium because the technical and benchmark claims come from the vendor and have not been independently corroborated here.
Sources
Update note: Last reviewed 2026-08-03. We will revise this post if DeepSeek changes the public-beta scope or publishes a V4-Pro release date.
Sources
- DeepSeek API change log — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.