DeepSeek's official changelog confirms that the legacy model names deepseek-chat and deepseek-reasoner stop working on July 24, 2026 (end of day Beijing time). If anything in your business calls the DeepSeek API by those names, it fails next week. The fix takes about twenty minutes. Finding everything that needs the fix is the real work.
What is changing
DeepSeek launched its V4 models in April 2026 and kept the old model names alive as aliases during a three-month transition. That transition ends July 24. The old names currently map to V4 Flash in its non-thinking and thinking modes, so code using them still works today, which is exactly why this deadline is easy to miss. After the cutoff, requests to the legacy names return errors instead of answers.
The replacement names are deepseek-v4-flash for fast, inexpensive work and deepseek-v4-pro for the stronger reasoning tier. DeepSeek also now serves an Anthropic-compatible interface alongside its OpenAI-style one, which matters if your tooling speaks one dialect but not the other.
"We don't use DeepSeek." Are you sure?
Most businesses that hit this deadline will not know they are using DeepSeek. It is one of the cheapest capable APIs on the market, which makes it a favorite default inside other things:
- Automation platforms. Custom API steps in n8n, Make, or Zapier that an employee set up last year and nobody has opened since.
- Internal scripts and side tools. The classification script, the report summarizer, the "quick" chatbot experiment that quietly became load-bearing.
- Vendor software. Some budget SaaS products route to DeepSeek behind the scenes. Their problem in theory, your outage in practice.
The twenty-minute check: search your codebase, environment variables, and automation platforms for the strings deepseek-chat and deepseek-reasoner. Replace them with deepseek-v4-flash or deepseek-v4-pro, run one test call each, and watch the first day of usage costs.
Three things to verify while you are in there
- Behavior, not just connectivity. The thinking mode on V4 models can change latency and output length compared to what your workflow was tuned for. Run a real task through it, not just a ping.
- Cost per call. V4 pricing differs from what you may have budgeted, and cache-hit discounts are deep. A quick before-and-after on one day of traffic tells you where you stand.
- Aggregator routing. If you reach DeepSeek through OpenRouter or a similar gateway, you are mostly insulated, but pinned model names in your gateway config can still go stale. Verify the pin.
The bigger lesson
Model deprecations used to be an annual event. In 2026 they happen quarterly, across every provider. Any business process that depends on an AI API now needs the same thing every other integration needs: an inventory of what calls what, and one person who reads the changelogs. That is not a product you can buy. It is a habit, and it is cheap compared to a Monday morning where the quoting tool stops answering.
If you are not sure what your automations actually call, that is a one-day audit for us. Tell us what is running and we will map it, or compare current models and prices yourself with our LLM Ranking Tool.