So today the previous Antigravity version got marked unsupported. One version back. I was on it because the current release had this “agent exceeded output tokens” bug that kept killing sessions. Couldn’t use the latest, downgraded, worked fine.
Today that option disappeared. Forced upgrade. And now nothing works.
Opus 4.6 gets instant 503 Agent Terminated. Not after a wait. Not after the IDE tries to queue. Instant. The request fires and dies in the same breath. Meanwhile if Gemini hits a capacity wall, the IDE actually holds the connection and waits for a slot. Same IDE, same session, same machine. One model gets a queue, the other gets killed on contact.
I checked. Claude’s status page shows Opus 4.6 had a brief incident earlier today, resolved hours ago. Gemini API is fully operational. Both upstream providers are healthy right now. This isn’t a real capacity problem. This is Antigravity’s routing layer reporting exhaustion when the actual APIs have availability.
The timing is what gets me. They pulled support for one IDE version and suddenly the capacity system is broken across models? How does deprecating a version break quota routing? What changed in the backend that made this happen?
I had about a month of completely smooth usage. No terminations, no issues, nothing. Whatever shipped today undid all of it.
Steps to reproduce:
- Update to the latest Antigravity build (the one forced today after the previous version got deprecated).
- Set model to Opus 4.6.
- Send any prompt.
- Watch it instantly return 503 Agent Terminated. No wait, no retry, no queue.
- Switch to Gemini. Hit a capacity limit and watch the IDE actually wait for availability instead of killing the session.
Both APIs are up. The IDE is the bottleneck.