Google resets Antigravity's Gemini quota after a Neon caching fix; Varun Mohan says quota now goes 22% further
Google reset everyone's Gemini quota in Antigravity on Saturday, November 21. Varun Mohan, who builds Antigravity at Google DeepMind, announced it at 09:18 UTC:
Reset everyone’s Gemini quota on Antigravity.
The caching improvements have increased the amount you can get done in quota by 22%. Enjoy building with Neon over the weekend!
"Reset everyone's Gemini quota on Antigravity. The caching improvements have increased the amount you can get done in quota by 22%."
The reset was promised the evening before. On November 20 at 18:38 UTC Mohan wrote that Antigravity was "massive increasing the cache pool size" for Gemini 4 Neon, which would raise the cache hit rate "by around 4%" and stretch what users get out of their quota "by at least 20%". Because the low hit rate had "burned quota faster," a reset would follow once the change was out. It came about fifteen hours later, and the final figure was 22%. Theo (@theo), who runs T3 Code and had been publicly pushing Google about Neon's caching, asked right after: "Did I accidentally turn Varun into Tibo by pushing the caching thing so hard?" Tibo is OpenAI's Thibault Sottiaux, whose Codex limit resets have become a regular event.
What was wrong
In a follow-up reply Mohan explained where the misses came from. The hit rate, he wrote, was "particularly bad" for workloads with more time between model invocations, because by then the cache had been evicted. An agent sends the model the whole conversation on every step. When the earlier part of that prompt is still in the provider's cache, it is cheaper to process. If the cache is too small, entries are pushed out quickly, and a session where the agent waits for a slow test run, or for the user to read a diff and reply, comes back to an empty cache and pays full price for context it already sent. A bigger pool keeps those prefixes around longer. Mohan said those workloads "should see a much higher improvement" than the average.
Why 4% turns into 20%
Theo had answered Mohan's first post with a suggestion: "reduced cache misses by 50%" sounds better than "increased cache hit rate by 4%." Mohan replied that he would use that next time. Neither gave the hit rates themselves. (After the reset Theo asked: "Did I accidentally turn Varun into Tibo by pushing the caching thing so hard?" Tibo is OpenAI's Thibault Sottiaux, whose Codex limit resets have become a regular event.) Theo's version works if the 4 is percentage points and the hit rate was around 92% before the change: misses go from about 8% to about 4% of input.
Halving the misses matters because cached input costs a fraction of fresh input. In the Gemini API, Neon's input is $3 per million tokens and implicitly cached input is $0.15 (our Neon coverage has the details). At those prices, an 8% miss rate makes the average input token cost $0.38 per million; at 4% it is $0.26, about 30% less. Output tokens don't change, so a whole session saves less. For the session as a whole to stretch by 22%, input would have to be about 60% of its cost before the fix. That is plausible for an agent that re-sends a long context on every step, but it is our assumption.
This is a rough check, not Google's accounting. Antigravity doesn't publish how it converts a request into quota. Its plans page only says the limits "correlate with the amount of work done by the agent," and nothing says the quota follows API prices.
What the reset covers
Antigravity counts quota per model group. According to its CLI documentation, Gemini models share one five-hour limit and one weekly limit, and Claude and GPT models share another pair. Mohan's post resets the Gemini quota only. It doesn't say whether that means the five-hour window, the weekly limit or both. On Google AI Pro and Ultra, the plans page says, quota refreshes every five hours until the weekly limit is reached; free accounts get a weekly refresh. Neon in Antigravity requires Google AI Pro, Ultra or an Enterprise plan.
Mohan also said Google "will continue to improve this metric globally," with no date or target. For now the only public number is the 22%, and by Mohan's own explanation the gain will be uneven: larger for sessions that sit idle between steps, smaller for ones that keep the model busy.
- Varun Mohan: Reset everyone's Gemini quota on Antigravity (Nov 21, 2026)
- Varun Mohan: change to increase cache hit rate on Neon in Antigravity (Nov 20, 2026)
- Varun Mohan: hit rate was bad for workloads with more time between invocations
- Theo: reduced cache misses by 50% vs increased hit rate by 4%
- Plans | Google Antigravity Docs
- Model quotas (/usage) - Antigravity CLI
Comments
No comments yet. Be the first.