On July 24, 2026, a quiet but consequential shift took place in the AI infrastructure landscape: DeepSeek officially retired its legacy API model names deepseek-chat and deepseek-reasoner. After a three-month compatibility window, these endpoints—familiar companions throughout the V3 era—stopped responding entirely. In their place, deepseek-v4-pro and deepseek-v4-flash now serve as the canonical entry points for DeepSeek’s next-generation models.
For developers who integrated DeepSeek into production systems over the past year, this is not a breaking change to ignore. Here is what you need to know and how to adapt your DeepThink-powered workflows.
Why the Shutdown Matters
The old endpoints were more than just names. deepseek-chat routed to DeepSeek V3’s general-purpose model, while deepseek-reasoner connected to the DeepThink R1 reasoning engine. Both have now been superseded by V4-class models that are faster, more capable, and significantly cheaper to run.
The migration is straightforward at the API level—change the model name—but the implications for DeepThink reasoning pipelines run deeper. V4 introduces architectural improvements that affect how reasoning traces are generated, how tokens are counted, and how costs accumulate over long chain-of-thought sequences.
V4-Pro vs. V4-Flash: Choosing the Right Model
DeepSeek V4 ships in two variants, and understanding the distinction is critical:
deepseek-v4-pro is the flagship model with 1.6 trillion parameters. It delivers the strongest reasoning performance and is the direct successor to deepseek-reasoner. If your application relies on DeepThink-style extended reasoning traces—mathematical proofs, multi-step planning, or code generation—this is your target.
deepseek-v4-flash packs 284 billion parameters and is optimized for speed and cost efficiency. It replaces deepseek-chat for fast-turnaround tasks where deep reasoning is unnecessary. At roughly 1% of the cost of competing frontier models, Flash is compelling for high-volume, latency-sensitive workloads.
Migration Checklist
- Update model names in every API call. Replace
deepseek-chatwithdeepseek-v4-flashanddeepseek-reasonerwithdeepseek-v4-pro. - Review token accounting. V4 models have a 1-million-token context window. Token pricing has changed—V4-Pro input costs $0.435/M tokens with cached input at $0.0036/M, making prompt caching more impactful than before.
- Test reasoning trace formats. DeepThink R1 traces under V4 may include new structural elements. Validate that your parsing logic handles these correctly.
- Adjust timeout and retry policies. V4-Pro’s extended reasoning can produce longer outputs. Ensure your client SDKs and gateway configurations accommodate the increased response times.
- Monitor cost and quality. Run A/B comparisons between V3 and V4 outputs before fully cutting over, especially for edge-case prompts in your production traffic.
What This Signals for DeepThink’s Future
The API retirement is not merely a cleanup exercise. It signals DeepSeek’s confidence that V4 has matured beyond preview status into a stable, production-grade release. The permanent price reduction on V4-Pro—locking in the 75% promotional discount—further underscores the company’s strategy: make DeepThink-class reasoning so affordable that choosing a closed, opaque alternative becomes difficult to justify.
For the broader ecosystem, this migration is a reminder that the AI infrastructure layer is moving fast. Model names are not permanent, API contracts evolve, and the applications that thrive will be the ones designed for adaptability. If you have not yet updated your integrations, do it now—the three-month grace period has expired, and the old endpoints are gone for good.