ACTION REQUIRED — DEEPSEEK API MIGRATION BY JULY 24
● Deadline: July 24, 2026 at 15:59 UTC — no announced extension
● What breaks: deepseek-chat and deepseek-reasoner aliases both return errors after this time
● What to migrate to: deepseek-v4-pro OR deepseek-v4-flash — same base URL, same API key, one parameter change
● Critical mapping: deepseek-reasoner → V4-Flash (thinking mode), NOT V4-Pro. If you relied on deepseek-reasoner for heavy reasoning, explicitly call deepseek-v4-pro.
● deepseek-chat: Maps to deepseek-v4-pro (general use)
● Pricing: V4-Pro $0.27/$1.10/M (standard hours), $0.07/$0.28/M (off-peak) · V4-Flash $0.07/$0.28/M standard
● Time remaining: 11 days as of July 13, 2026
The Migration — Exactly What to Change
Every development team using DeepSeek's hosted API has one non-optional task: update any code that calls deepseek-chat or deepseek-reasoner before July 24, 2026 at 15:59 UTC. The mechanical change is trivial — one parameter in your API call. The risk is in the alias mapping that many developers may have assumed rather than checked.
Migration mapping table
| Old alias (breaks July 24) | Correct replacement | Important notes |
|---|
| deepseek-chat | deepseek-v4-pro | General use — standard replacement |
| deepseek-reasoner | deepseek-v4-flash | ⚠️ Maps to Flash thinking mode, NOT V4-Pro. For heavy reasoning: use deepseek-v4-pro explicitly. |
deepseek-reasoner maps to V4-Flash (thinking mode), not V4-Pro. If you were using deepseek-reasoner for heavy reasoning tasks and need equivalent capability, you must explicitly call deepseek-v4-pro, not just replace the alias name. V4-Flash is cheaper and faster but has lower reasoning ceiling than V4-Pro. Teams using DeepSeek for complex multi-step reasoning or code analysis and migrating deepseek-reasoner to V4-Flash without checking will silently degrade their pipeline quality. Test your specific workloads against both V4-Pro and V4-Flash before deploying the migration to production.
Data Residency — The Non-Optional Context
The hosted cloud API routes all data through DeepSeek's servers in China, where China's National Intelligence Law (2017, Article 7) legally requires the company to cooperate with government intelligence requests on demand — no court order, no exceptions. This applies regardless of DeepSeek's privacy policy or the physical location of any given server. For teams processing regulated data (HIPAA, GDPR, FINRA, FedRAMP), the data residency question is separate from the migration question. The API migration is a technical change that must be made by July 24 regardless. The data residency evaluation is a compliance question that should already have been resolved.
When you self-host DeepSeek V4 weights on Western infrastructure, prompts never reach DeepSeek's servers in China, and the cross-border data-flow concern disappears. The legal obligations that run with operating under Chinese jurisdiction do not attach to a self-hosted deployment on Western infrastructure. The censorship behaviour baked into the model's weights carries over regardless of deployment environment — that is trained behaviour, not a server-side filter. For regulated industries: self-host on compliant infrastructure rather than using the hosted API.
Also: Mistral Leanstral 1.5 — AI That Can Prove Its Own Code Works
With Leanstral 1.5, Mistral moves beyond code generation by providing mathematical proof that software behaves as intended using Lean 4. Strong benchmark results demonstrate significant progress in formal software verification for critical systems. This breakthrough could greatly improve the reliability of future software development.
Leanstral 1.5 is a 119 billion parameter model released under the Apache 2.0 license — fully open for commercial use. Lean 4 is a formal proof assistant: rather than generating code and hoping it works, Leanstral 1.5 generates Lean 4 proofs that mathematically verify the code's behaviour. This is not a QA or testing approach — it is a mathematical guarantee. For critical systems where bugs carry catastrophic consequences (avionics, medical devices, financial clearing systems), formal verification has historically required specialist expertise and weeks of work. Leanstral 1.5's approach — if the benchmark results hold at production scale — would make formal verification accessible without that expertise.
The practical limitation: formal verification with Lean 4 works for well-specified functions with clear mathematical properties. It is harder to apply to large-scale software systems with complex state, side effects, and external dependencies. Leanstral 1.5 is a genuine research advance for the formal verification domain — not a replacement for general-purpose code generation agents on enterprise software.
DeepSeek Migration Checklist
1. Audit all code calling deepseek-chat or deepseek-reasoner. Search your repos for both strings. Include CI/CD pipelines, agent frameworks, and any production automation.
2. Replace deepseek-chat with deepseek-v4-pro. Direct swap. Test output quality — V4-Pro is the general-purpose replacement.
3. For deepseek-reasoner: decide V4-Pro or V4-Flash based on your reasoning workload. Heavy multi-step reasoning → use deepseek-v4-pro. Fast thinking mode at lower cost → use deepseek-v4-flash. Do NOT assume deepseek-reasoner → deepseek-v4-pro without testing.
4. Test before July 20. Leave 4 days of buffer before the July 24 hard deadline to catch any unexpected quality regressions or API behaviour changes.
5. If using the hosted API for sensitive data: Evaluate self-hosting the MIT-licensed weights on your own infrastructure. The weights are available on Hugging Face.
Sources: TechTimes July 8, 2026 (DeepSeek deadline) · Medium / David Akpovi (Leanstral 1.5) · Related: China considering overseas AI restrictions → · LongCat-2.0: Chinese open-source beats GPT-5.5 → · How to build profitable AI agents →