On 24 July 2026 at 15:59 UTC (17:59 Paris time), the deepseek-chat and deepseek-reasoner identifiers become permanently inaccessible. Any request using them will return an error. No grace period, no automatic fallback. If you have code calling the DeepSeek API, now's the time to check it.
There's a category of tech news that makes no noise but breaks systems in production: the end of life of an API identifier. Today, it's DeepSeek's turn. Here's what's changing, how to adapt in a few minutes, and the trap waiting for those who migrate too quickly.
What's happening exactly
On 24 April 2026, DeepSeek released its fourth generation of models, temporarily keeping the old names alive as aliases to give teams time to adapt. That 90-day window closes today.
In practice, the old identifiers have already been pointing to the new models since April, without anyone having to do anything. In other words, you're probably already using version 4 without knowing it. What's disappearing isn't the model—it's the name. And a name that no longer exists in a production configuration brings an entire chain to a halt: automated pipeline, agent, scheduled report, background task.
The mapping to remember is simple. The old deepseek-chat becomes deepseek-v4-flash. The old deepseek-reasoner becomes the same deepseek-v4-flash, but with reasoning mode enabled. The API endpoint, key, and request structure don't change.
Look closely at the second mapping. The old reasoning model points to the Flash version—the fastest and most economical—not the Pro version. If you were using that identifier specifically for the quality of its reasoning, you risk a silent degradation of your results: the code will work, but the answers will be worse. The workaround is to test your outputs after migration, and switch to deepseek-v4-pro with reasoning enabled if quality drops. An outage is visible immediately; a degradation surfaces three weeks later.
Another subtlety that hits the wallet: reasoning mode is no longer determined by the choice of model—it's become a request parameter. If you were using the old chat identifier for simple, high-volume tasks, remember to explicitly disable reasoning after migration. Otherwise you'll pay for reasoning tokens you don't need—a mechanism we detailed in our article on reasoning models.
The two models, to choose right
| Model | Parameters | Best for |
|---|---|---|
| deepseek-v4-flash | 284 billion total, 13 billion active | Volume, speed, lower cost |
| deepseek-v4-pro | 1.6 trillion total, 49 billion active | Demanding tasks, high-quality reasoning |
Both are released under the MIT licence with open weights, and have a one-million-token context window. The Flash version is priced at around $0.87 per million output tokens—a rate that remains 60–90% lower than comparable leading US models.
What this deadline reveals
Beyond the practical side, this migration says a lot about a shift. According to a recently reported analysis, Chinese models would now account for between 30 and 46% of tokens consumed by enterprises via major US development platforms. A simple maintenance operation at a Chinese provider is now capable of breaking production systems worldwide.
It's a very concrete illustration of the dependency we discussed in our article on open models. Ironic twist: DeepSeek publishes its weights under a free licence, meaning nothing forces you to go through its API. Those who self-host the model simply aren't affected by today's deadline. Using an open model via its creator's API means enjoying the convenience while keeping the dependency. Freedom exists—you just have to exercise it.
If you have code in production, the check takes five minutes: search for the two old identifiers in your codebase, gateway configurations, and infrastructure files. Five minutes today beats an on-call page tonight.