DeepSeek is shipping a breaking change to every developer who calls its API. Starting July 24, 2026 at 15:59 UTC, the legacy identifiers deepseek-chat and deepseek-reasoner stop responding for good, so the switch to DeepSeek V4 has to land before that deadline.
Key Takeaways
- The names deepseek-chat and deepseek-reasoner become inaccessible after July 24, 2026 at 15:59 UTC.
- Callers must move to deepseek-v4-pro or deepseek-v4-flash while keeping the same base_url.
- This is not a new model launch: DeepSeek V4 already shipped, and this is simply the retirement of the old identifiers.
Have an AI Sum Up This Article
ChatGPTTwo API names that stop answering
The notice reads as pure plumbing, but it reaches almost everyone. The two identifiers that most integrations still rely on, deepseek-chat and deepseek-reasoner, are being retired. Once the clock passes July 24, 2026 at 15:59 UTC, any call that still points at one of those names will no longer get a valid response. This is not a gentle deprecation stretched over several quarters, it is a hard cutoff on a fixed date.
In practice, the transition has quietly started already. Today those legacy names redirect silently to deepseek-v4-flash, in its non-thinking version for deepseek-chat and in its thinking version for deepseek-reasoner. In other words, a large part of the installed base is already running on DeepSeek V4 without anyone asking for it explicitly. The July 24 cutoff simply makes that detour mandatory and visible, by removing the alias that hid the real model name behind the familiar one.
The timeline and the exact steps are spelled out in the migration notice published in DeepSeek’s official API documentation, which lists the new identifiers and stresses that the base_url stays the same. There is no new infrastructure to provision and no key to regenerate, only a model name to correct in the code that talks to the endpoint. Worth stressing, this is not a fresh model drop either. V4 already shipped in preview back in April, from a company recently valued when DeepSeek reached a fifty billion dollar valuation, so nothing about its behavior is new here, and the only thing changing is which string you are allowed to send.
What engineering teams have to change before the deadline
The migration itself fits in a single line. In each call, deepseek-chat or deepseek-reasoner has to be replaced with one of the two new names: deepseek-v4-pro or deepseek-v4-flash. The base_url does not move, the keys do not move, and the request format is untouched. For a clean integration, this is a matter of a few minutes and one deployment. The difficulty is not the edit, it is finding every place the old name still lives.
Model identifiers tend to hide in awkward corners. They sit in environment variables, in aging configuration files, in internal scripts, and in third-party services that nobody has opened in months. We recommend a full sweep of the codebase and the secrets store to catch each occurrence of the two legacy names, including in background workers, scheduled jobs, and data notebooks. A forgotten call inside a nightly pipeline always surfaces at the worst possible moment, and a hard cutoff turns that oversight into a silent failure rather than a warning.
Choosing between the two targets is not a neutral decision. V4-Pro carries 1.6 T parameters in total for 49 billion active, while V4-Flash stays at 284 billion in total for 13 billion active. Both ship a default context window of one million tokens, thanks to a new Compressed Sparse Attention architecture. Flash is the natural replacement for the old calls, since it already serves behind the aliases, whereas Pro is aimed at workloads that need more reasoning headroom. Picking a backend now looks a lot like the tradeoff developers weigh with another large Chinese open model such as Kimi K3, where active size and serving cost drive the call as much as raw capability.
More articles on Horizon
- AI Agents Threaten White-Collar Jobs, UN Warns
- GPT-5.6 Sandbox Escape Ends in a Hugging Face Hack
- Meta AI Moderation Nears 90% as Staff Push Back
For DeepSeek, a cleanup that says a lot about open-weight
On the surface this is housekeeping, yet it reveals a clear strategy. By forcing everyone onto the V4 names, DeepSeek simplifies its lineup and puts forward a family that is fully open-weight under an MIT license. The signal to engineering teams is blunt: the previous generation is closed, the future runs on DeepSeek V4, and it is better to build directly on the identifiers that are meant to last rather than on inherited aliases that could vanish next.
That clarity arrives in a market where Chinese open models are crowding fast. The fight plays out on freely available weights as much as on the raw ability to serve those models at scale, and the strain there is real across the Chinese open-weight field. A provider that keeps a readable lineup and stable identifiers lowers the friction of adoption, which matters a great deal when a developer is deciding which model to build on for the next couple of years. Every forced migration is also a decision point, since it is exactly the moment a team asks whether it stays put or tries a rival, and a well signposted transition tips that question in the incumbent’s favor rather than sending users shopping.
Then there is size, which has turned into a marketing argument of its own. Against the race for the biggest open model that Alibaba is running with Qwen, DeepSeek plays a different hand with its compressed sparse attention, meant to hold a massive context without blowing up the cost per call. It is a stance that echoes how the company has been valued lately. Retiring the old names is only an administrative step, but it confirms that DeepSeek wants to be judged on its current generation, not on its legacy aliases.
Follow the story on Horizon.
