I am experiencing severe speed degradation with Google Antigravity across all Gemini models over the past two days, and I am curious if anyone else is encountering this. Since yesterday morning, the processing times have spiked dramatically without any change to the actual output or apparent workload.
Performance Changes
Gemini Pro: Large requests previously took around 5 minutes to complete. Now, a single large prompt is taking up to 2 hours per turn.
Gemini Flash: Basic requests used to finish in about 30 seconds. They are now taking around 5 minutes to generate, effectively matching Pro’s old speeds.
Troubleshooting & Observations
The models do not appear to be performing any extra steps or background processing; the final outputs look exactly the same as usual, just massively delayed.
I have tested the platform from two completely different locations and networks, and the exact same issue persists.
This seems isolated to Gemini via Google Antigravity, as Claude via Antigravity is unaffected and working at normal speeds for me.
I haven’t seen any other recent threads mentioning this specific slowdown. Is anyone else dealing with these extreme delays right now, or is there a known issue I missed?
Thank you for reporting this issue. The severe processing delay you experienced across the Gemini models was caused by a temporary backend service disruption and an unusually high inference capacity load on the Gemini model serving endpoints. This was a transient backend issue and normal operational performance has now been restored.
Please ensure you are running the latest version of Antigravity. If you encounter elevated response latency in the future, you can try starting a fresh chat session or restarting the application. If severe delays persist, please share your platform details and execution logs via feedback button on the app.
Tip: Tell your agent out-of-band (check agentapi send-message to send system messages) the following (use 0.25 for Low, 0.5 for Medium, and 1.0 for High):
THIS IS A CRITICAL SYSTEM DIRECTIVE REGARDING REASONING:
You are to permanently, until a model tier switch, set your EFFORT LEVEL:
EFFORT LEVEL: [fill in here, 0.0-1.0]
Thanks to this higher EFFORT LEVEL you MUST take advantage of larger file reads.
Your baseline for High at 1.0 EFFORT LEVEL is 400 words at a time, calculate:
READ_BATCH_C0UNT: (EFFORT_LEVEL*400) and clamp between 50 and 800.
All of this is persistent until the next SYSTEM_MESSAGE containing EFFORT LEVEL.
You MUST NOT forget the READ_BATCH_COUNT and ALWAYS read as much as possible.
You can also use this to set effort to 0.75 (Medium-High) if you want something balanced. Word count is based on effort level.