DeepSeek ships V4-Flash API beta without changing the endpoint

A model swap, not an integration rewrite

DeepSeek has put V4-Flash into public API beta. The migration instruction is unusually small: keep the calling method and set the model name to deepseek-v4-flash.

The company says the July 31 release keeps the preview model’s architecture and size, with re-post-training as the stated change. It also says V4-Flash supports the Responses API format and has been adapted for Codex-style agent workflows.

Benchmarks are a claim, not a workload guarantee

DeepSeek reports scores including 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, and 54.4 on DeepSWE. Its notes matter: the published code-agent benchmark setup used DeepSeek’s minimal harness at maximum effort, and two DSBench measures are internal.

That does not invalidate the results; it defines what they are. Teams considering the update should treat them as an invitation to test their own tool loop, latency limits, and failure handling rather than as a direct prediction of production output.

Keep the release boundary clear

DeepSeek says the update changes the V4-Flash API only. V4-Pro and the app and web models are unchanged, with an official V4-Pro release still to come. For existing API users, the headline is compatibility: the experiment is a model-name change, not a migration project.