Inception's Mercury API shuts down December 28: move now
Inception is winding down the Mercury API on December 28, 2026, and your logs and invoices go with it. Here's what to export and how to migrate off diffusion LLMs.
Inception is shutting down the Mercury API. The diffusion LLM service stays fully available through December 28, 2026. After that, requests stop and you lose access to the platform and to any data you submitted. A month ago we wrote about Mercury 3 at 1,068 tokens per second. Now the fastest model in that post has an expiry date.
What actually happened
Per Inception's wind-down notice and its docs home page:
- The date. The Mercury API and the Inception Platform dashboard work as normal until December 28, 2026. After that, API requests are not served.
- The data. Request logs, billing history and invoices on the platform become inaccessible after the cutoff. Inception tells customers to download them now.
- The models. The docs list Mercury 3, Mercury 2.5, Mercury Voice, Mercury 2, Mercury Edit 2 and Mercury Decide. The notice covers the Mercury API as a whole.
- No reason given. The notice says only that Inception made "the decision to wind down the Mercury API."
The likely reason sits next to it. The Information reported that Nvidia is in advanced talks to buy Inception for more than $500 million, with all three co-founders joining Nvidia. Neither company has confirmed the deal. Treat it as reported, not done.
Why it matters for your business
If you picked Mercury for speed, you probably put it in the hot path: autocomplete, code edits, routing, voice. Those are the calls that break loudly when the endpoint goes dark at 12:01 a.m.
Five weeks is enough time if you start now:
- Find every caller. Search code, env files and no-code tools for
api.inceptionlabs.ai. Inception says this in the notice too. - Export first. Pull invoices for your books and request logs for your evals before December 28. You can't get them back after.
- Swap at the client, not in the app. Mercury uses an OpenAI-compatible API, so a base URL and model name change gets you most of the way. Run your evals against the replacement before you flip it.
- Re-test latency. No autoregressive model will match diffusion speeds. Check the user-facing steps where you relied on that speed, and add streaming or caching where you need it.
This is the same lesson as every acqui-hire. A model API from a startup is a feature of that startup's runway. Keep one config line between your app and the provider, and keep a second provider tested.
Key takeaways
- Inception's Mercury API stops serving requests after December 28, 2026
- Request logs, invoices and billing history are inaccessible after the cutoff, so export them now
- Nvidia is reported to be in talks to buy Inception for over $500M; neither company has confirmed
- The OpenAI-compatible API makes the swap mostly a config change, but re-run your evals
- Expect slower responses on any replacement and plan the UX for it
Got Mercury in production and five weeks on the clock? We build model-agnostic AI layers with a provider switch in config, evals that run before every swap, and a fallback that's already tested. See how we build them or send us your call map.
Sources: Inception — Mercury API wind-down, Inception Platform docs, VKTR on The Information's report.
- #inception-labs
- #mercury-api
- #diffusion-llm
- #vendor-risk
- #llm-migration
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenRouter drops 3 providers, pulls ZDR from 125 endpoints
OpenRouter off-boarded three inference providers and removed 125 endpoints from zero data retention after Dragos canary research. What to check in your stack.
Read itOpenAI Performance Factory: Codex agents debug, engineers merge
OpenAI's Performance Factory ran 1,700+ Codex investigations into ChatGPT slowdowns. 70% of audited root causes held up; about 15 agent PRs shipped as written.
Read it