DeepSeek V4.1 Flash is live: API access and Pro migration

DeepSeek V4.1 Flash is available through the official DeepSeek API as deepseek-flash. The September 10 release adds a new model behind the Flash name, lowers its listed prices and sets a September 14 transition for V4 Pro requests. This page updates our September 9 preview using the official release log, rather than continuing to rely on the earlier notice screenshot.

Checked September 10, 2026. API availability is confirmed by official documentation; this article does not establish that every website, app or third-party gateway has switched, and contains no paid performance test.

Which model name should you use?

Do not construct an identifier from the display name. A tutorial using a temporary expires-on-0910 identifier describes a different stage of the rollout. Start with the API setup guide and verify the endpoint and model together.

What changes for existing V4 Pro users?

The Pro transition starts September 14, 2026 at 12:00 Beijing time (04:00 UTC). From then until a future V4.1 Pro release, DeepSeek says Pro requests will be served by V4.1 Flash and charged at Flash rates. September 10’s Flash launch and September 14’s Pro routing change are separate events.

Before that boundary, retain a small baseline of the tasks that matter to your application. Include structured outputs, tool loops and billing records rather than only a chat prompt. Afterward, rerun the same acceptance checks. Our V4 Pro migration checklist explains what to record. A stable request string does not prove the underlying model stayed the same.

What can V4.1 Flash do?

The official documentation lists native visual understanding, tool calls, thinking and non-thinking modes, and both Responses and Anthropic-compatible APIs. Visual understanding means accepting images to answer questions about their contents; it is not an announcement of an image-generation endpoint.

DeepSeek’s release benchmarks are vendor-reported results. They are useful for identifying workloads to evaluate, but they do not prove your agent will be faster or more accurate under a different harness, prompt or tool set. We have not reproduced those benchmarks here.

For practical coding setups, use the dedicated Codex guide or Claude Code guide. Protocol compatibility and a model appearing in a menu are separate from a complete working agent loop.

How should you budget for it?

DeepSeek maintains separate cache-hit input, cache-miss input and output rates, with peak and off-peak periods. Use the V4.1 Flash pricing guide for the current USD table, time zones and a worked request budget. A headline cache rate is not the price of every token.

An official rate is not a quote for every reseller. If you use Ofox, inspect its model catalog and the specific route’s price and protocol before committing funds. You can create an account to prepare access, but signup alone does not establish that a particular route is available to you.

What happened to the original preview?

The original September 9 article reported a supplied usage-page notice and explicitly left the final identifier, numeric rates and exact Pro transition time unconfirmed. Those points are now covered by public official documentation. We have kept this URL so existing links lead to the current release information; the announcement does not retroactively turn the preview into an independent benchmark or a third-party availability test.

For model selection, compare DeepSeek V4.1 Flash with Gemini 3.8 Flash and Qwen3.8 Flash across pricing, modalities, API compatibility and deployment fit.