DeepSeek API Deadline: deepseek-chat and deepseek-reasoner Retire July 24, 2026

If your product calls DeepSeek's API directly, mark July 24, 2026 on your calendar. That's the DeepSeek API deadline: at 15:59 UTC, the legacy model identifiers deepseek-chat and deepseek-reasoner stop working entirely, and any request still using those names will simply fail.

What's actually changing

DeepSeek rolled out its V4 generation earlier this year, led by deepseek-v4-pro and the lighter deepseek-v4-flash. Both replace the old chat and reasoner endpoints, and both ship with a much larger context window: 1,048,576 tokens (roughly 1M), a sizable jump from the older models' limits. V4 Pro is the heavier option, a mixture-of-experts model with 1.6 trillion total parameters and about 49 billion activated per request, aimed at complex reasoning and long-document work. V4 Flash trims that down to 284 billion total parameters for faster, cheaper responses on simpler tasks.

The good news for anyone maintaining an integration: DeepSeek has kept the base URL and API key scheme identical across the migration. This isn't a full platform switch, it's a model-name swap. Wherever your code currently sends "deepseek-chat" or "deepseek-reasoner" as the model parameter, it needs to send "deepseek-v4-flash" or "deepseek-v4-pro" instead. No new authentication flow, no new SDK.

Why the deadline still catches people off guard

Model deprecation deadlines are common across the industry, but they're easy to miss when a model is buried three layers deep in a backend service rather than sitting in a UI a human checks daily. Teams that treat an LLM API the way they'd treat a stable cloud API, something you configure once and forget, are the ones most likely to get a wall of failed requests on July 25 if nobody updated the string.

The pricing shift is also worth planning around, not just the technical swap. Per trackers like OpenRouter and PricePerToken, V4 Pro runs close to $0.435 per million input tokens and $0.87 output, while V4 Flash is markedly cheaper at roughly $0.09 input and $0.18 output per million tokens. Neither the identical dollar amount you were paying under the old names, so it's worth re-checking your cost assumptions rather than assuming a like-for-like swap.

The bigger pattern: API-first AI is a moving target

This kind of forced migration isn't unique to DeepSeek. Every major AI lab has, at some point, deprecated a model name, changed pricing, or shifted a default. What makes DeepSeek's case notable is the scale of adoption behind it: DeepSeek has become one of the most widely used API-first models globally in 2026, in part because of aggressive pricing relative to Western frontier models. A deadline affecting a model this widely embedded touches a lot of production code at once.

For teams that call a single provider's API directly, the practical lesson is that model names are not permanent infrastructure. They get renamed, retired, and repriced on the vendor's timeline, not yours. This is the same friction that pushes some developers and end users toward tools that abstract the provider layer entirely, picking a model from a list rather than hardcoding an identifier string. That's the same premise behind bring-your-own-key tools like ByteChat, where switching between DeepSeek, Claude, GPT, or GLM models is a dropdown choice rather than a code change.

Frequently asked questions

Did DeepSeek really set a hard deadline for its old API models?

Yes. According to migration guides tracking the change, deepseek-chat and deepseek-reasoner stop functioning entirely after July 24, 2026 at 15:59 UTC, with no extension being discussed as of this writing.

What do I replace deepseek-chat and deepseek-reasoner with?

Use deepseek-v4-flash for faster, lower-cost requests or deepseek-v4-pro for more demanding reasoning tasks. Both support a 1M-token context window, and your existing base URL and API key continue to work unchanged.

Is DeepSeek V4 cheaper than the old models?

Pricing has shifted rather than simply dropped across the board; trackers list V4 Flash as low as roughly $0.09 per million input tokens, while V4 Pro costs more at around $0.435 input and $0.87 output per million tokens, so it's worth checking your specific use case against current rates rather than assuming the old price carries over.

Whichever side of this migration you're on, the underlying lesson is the same: don't let a vendor's model-name string become a single point of failure in your stack.