DeepSeek · API Sunset
On July 24, the
direct DeepSeek API disappears.
DeepSeek has confirmed that the current public API will shut down on July 24, 2026. It is being folded into the next release, but any stack calling DeepSeek directly has only twelve days of realistic runway. On the same day, China unveiled the Orca world model and Ant Group's LingBot-VA 2.0 robotics AI — three moves on one date that shift the map of Chinese AI another notch.
The Announcement
Shutdown on 7/24,
twelve days of runway
A "next-release replaces this one" sunset. Old-API calls after that date return errors.
DeepSeek is folding the old API into a new one. Detailed specs and pricing for the successor should land by shutdown day, but direct-callers have to move to some option within twelve days. After 7/24, calls to the old endpoint fail, so any production stack without a fallback goes down on its own.
China also revealed two other pieces on the same day: the Orca world model (research-stage, sitting between simulation and video generation), and Ant Group's LingBot-VA 2.0 robotics foundation. Beyond DeepSeek, the Chinese AI stack has more layers to keep track of.
By The Numbers
Seven facts
not to misread
Why It Matters
Not a version bump —
a signal to re-check China exposure
Sunset notices are routine. What shifts here is where the meaning sits.
DeepSeek's role has been the cost-effective Chinese frontier option — not a fallback to Claude or GPT but a real alternate that entered production stacks. Lindy dropping Claude in favor of DeepSeek on 6/27 was the clearest example. This sunset shows that the position rested on a quiet assumption of API continuity with a short expiry.
The move is less a DeepSeek-specific stumble than part of an industry trend where model transitions are also API transitions. Buyers now need to add sunset cadence and runway length to the same evaluation grid as model performance.
Who It Hits
Where the shock actually lands
| Deployment pattern | What happens on 7/24 |
|---|---|
| Direct API in your own backend | Calls fail after 7/24. Swap to the next-gen API or a different model is mandatory. |
| Via SaaS (Cursor, Lindy…) | Vendors handle the swap. Users mostly just wait for the notice and reconfirm settings. |
| Experiment / PoC only | Low impact — a good moment to re-evaluate Qwen and the next-gen DeepSeek side by side. |
| Not using DeepSeek | No direct hit. Still worth watching how the Chinese stack (Orca, LingBot-VA) fills out. |
Before the model's price,
check the API's expiry date.
What To Do Next
Four moves for this week
Enumerate direct-call sites
Grep the codebase for deepseek.com/v1-style calls and confirm they can be swapped at the env-var level. Any hardcoded endpoint left in place takes production down on 7/24 — this is the top priority.
Prepare a fallback without waiting for the new spec
Assume the next-gen spec might land late. Wire in a fallback path to Qwen and other models via Bedrock / Vertex so you can swap targets on the day of shutdown.
Re-cost SLO and per-1K pricing
For each candidate, measure per-1K-token cost and p95 latency yourself. DeepSeek's "cheap" edge may not survive substitution; get updated numbers to finance early.
Log Chinese-stack picks in an internal memo
Give Orca and LingBot-VA 2.0 a one-page write-up in your tech notes. Next time you evaluate, performance, cost, and geopolitical risk sit side by side.
Counterview
Counterview and limits
Two caveats. First, DeepSeek's announcement includes a "next-gen equivalent" clause; the swap may effectively be a rename. The old-API sunset may not translate into real pain. Second, Orca and LingBot-VA 2.0 are not yet full commercial-grade; they're not ready to run production workloads today.
Still, treating "model migration = API migration" as the default is a discipline this event reinforces. Not DeepSeek-specific; frontier-wide.