DeepSeek has moved the official V4 Flash API into public beta. The company’s July 31, 2026 changelog says existing integrations can call the updated release by continuing to use the deepseek-v4-flash model name.
This is an API release, not a claim that DeepSeek’s consumer app or web model changed. The same notice says V4-Flash-0731 keeps the architecture and model size of V4-Flash-Preview and was re-post-trained. V4 Pro and the app/web models are unchanged.
DeepSeek also reports higher scores on several agent-oriented benchmarks. Those numbers are first-party results produced with vendor-controlled settings, including an unreleased harness for the public coding tests and two internal test sets. They describe DeepSeek’s evaluation, not independent validation.
Calling the beta does not require a new request shape
DeepSeek says the calling method and public model name remain the same. Its current model page lists the OpenAI-format base URL as https://api.deepseek.com and the Anthropic-format route as https://api.deepseek.com/anthropic.
V4 Flash supports tool calls, JSON output, the Responses API and the Anthropic API. DeepSeek’s page lists a one-million-token context length and a maximum output of 384,000 tokens. Thinking mode is supported and is the documented default, while non-thinking mode is also available.
The practical migration question is therefore narrower than a new endpoint rollout: teams already targeting deepseek-v4-flash should confirm that their provider settings, response handling and test fixtures behave as expected against the July 31 version.
Published limits and prices are operational inputs, not guarantees
As of July 31, DeepSeek lists V4 Flash prices per one million tokens at $0.0028 for cache-hit input, $0.14 for cache-miss input and $0.28 for output. The same page says prices may change.
It also describes a future peak/off-peak policy under which peak-hour billing would be twice the regular price, but says the effective date will be announced later. That policy should not be treated as active until DeepSeek publishes the start date.
Beta status matters beyond billing. Before moving production traffic, teams should verify rate and concurrency behavior, tool-call correctness, response compatibility, latency, cost and failure recovery on their own workloads. A documented context window or output ceiling does not establish that every application will use those limits reliably or economically.
The benchmark numbers need a vendor label
DeepSeek says V4 Flash improved across agent benchmarks including Terminal Bench 2.1, NL2Repo, Cybergym and DeepSWE. The changelog also reports results for Toolathlon, Agent Last Exam, Automation Bench and two DSBench sets.
The methodology notes limit what those figures can show. DeepSeek says the public Code Agent tests used its unreleased Harness minimal mode with maximum effort, top_p=0.95 and temperature 1.0. DSBench-FullStack and DSBench-Hard are internal test sets. The company did not present these numbers as independently reproduced results.
Bloomberg’s independent report confirms the public-beta rollout and attributes the performance statement to DeepSeek. It does not turn the benchmark table into third-party verification.
For evaluation teams, the defensible next step is an application-specific comparison with fixed prompts, tools, budgets, data and scoring. Record the exact model version and harness, not only the public model alias.
The update is narrower than a new model family launch
DeepSeek describes V4-Flash-0731 as a re-post-trained version with the same architecture and size as the preview. Only the V4 Flash API changed on July 31; V4 Pro and the consumer app/web models did not.
The company’s Codex integration guide documents a Responses API provider configuration for V4 Flash. It offers automated and manual setup paths. Developers should review vendor-supplied scripts and configuration changes before running them, especially when credentials and local tool settings are involved.
The release is useful because it makes the updated Flash endpoint broadly testable under the existing model name. The evidence supports that availability claim. Claims about better agent performance still depend on DeepSeek’s methods and should be tested independently before they shape production decisions.
