DeepSeek has placed its V4 Flash API into public beta, describing the release as an upgrade aimed at improving AI agent performance. The company announces the update via its official X account and says the V4 Flash model improves results on agent-related benchmarking compared with a prior “V4 Pro Preview.” Tech outlets report that DeepSeek positions the release as being integrated with the Responses API and adapted for Codex-related workflows.
According to published benchmark figures cited in coverage, DeepSeek reports a score of 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE, alongside additional unspecified benchmarks. Coverage also notes that the V4 Flash line includes a version labeled “V4-Flash-0731.”
Overall, the reports agree that the public beta focuses on faster tooling for agent tasks, with added support for the Responses API and performance claims based on standardized benchmarks. DeepSeek’s communications emphasize comparative improvement against earlier preview tooling.
