This is my third post about this. Antigravity is capping Claude Opus 4.6 at 64k output tokens when the model natively supports 128k. This needs to be fixed now.
Here is what is actually happening and why most of you probably have not noticed. You give Opus a task, walk away, come back, and see it completed. Looks fine. But what you did not see is the model failing ten times in between. It generates its real response, hits “The model’s generation exceeded the maximum output token limit,” then tries a completely different approach. That fails too. It retries again and again until it finally strips everything down enough to fit under the cap. The code you got back is not what Opus intended to write. It is a compromised, squeezed down version of it.
Opus 4.6 was clearly trained to use its full 128k output. On nine out of ten tasks it blows past 64k on the first attempt. Every retry burns tokens you are paying for and the final output is degraded. You are effectively getting Sonnet 3.7 quality out of an Opus model because the platform will not let it finish its actual response.
If you have been shipping code with Antigravity in the last couple of days, go back and look at it. Seriously. The quality is not what you think it is.
If you are seeing this, post about it. This is not a niche problem, it affects every Opus user on the platform. The more visibility this gets the faster it gets fixed.
