Varun Mohan, the former Windsurf CEO now running Google's Antigravity team at DeepMind, posted on X on May 21 that Google was tripling Gemini rate limits across all paid Antigravity tiers and resetting everyone's weekly quota. The change took effect immediately. What he didn't lead with: this was the company patching a problem it had created days earlier at I/O.
And here's the part the cleaner write-ups skipped. Google did it twice.
One announcement, two scrambles
The first 3x landed Wednesday, alongside a quota reset, after users started hitting walls almost immediately. Then, in the same week, Mohan said the team was tripling the weekly quota a second time and resetting paid-plan quotas again. He admitted users were exhausting their allowance "after a couple work sessions," which is not a sentence you want to be saying about a paid product you launched last week.
Stack the two boosts and you get roughly a 9x increase over where things landed after the original cut. Sounds enormous. It mostly tells you how far the floor had dropped to begin with. Several users pointed out the obvious: even after all this, the limits sit below what they had before the change. You don't reset something three times in a week because it was working.
What actually broke
At I/O 2026, Google swapped how it meters Gemini. The old system counted prompts. The new compute-based approach weighs the complexity of each request, which features you touch, and how long your chat history runs. The pitch is that it allocates server capacity more fairly. The result, for anyone doing heavy agentic coding, was that a Pro or Ultra allowance could evaporate in minutes, followed by five-hour cooldowns and weekly lockouts.
For an agent-first IDE, that's close to fatal. The whole premise of Antigravity is handing long, autonomous coding jobs to the model. Meter it by compute and the most-intended use case becomes the one that burns out fastest. One developer wrote up an agent hitting the wall in eight minutes. Eight.
Mohan's team owned the misstep, more or less, and said more changes are coming. To his credit the response was fast, measured in hours, not the usual quarter. Whether "the 3x is forever," as he put it in a follow-up, survives contact with Google's actual server economics is a separate question. Permanent is a strong word from a team that revised the same number twice in five days.
Why Mohan, of all people
There's a backstory worth a sentence. Mohan co-founded Windsurf, the AI coding startup, then joined DeepMind in July 2025 through a talent deal reported around $2.4 billion, after OpenAI's $3 billion bid for the company fell apart. So the person now apologizing for restrictive limits on a developer tool is the same person who spent years building developer tools that lived or died on whether developers felt nickel-and-dimed. He knows exactly how this looks. That's probably why the fix came as fast as it did.
The catch nobody at Google is putting in a headline
The boosts apply inside Antigravity only. Everywhere else in the Gemini suite, the new compute-based caps stand untouched. So if you're using Gemini for deep research, or in the app, or anywhere outside the coding tool, none of this reaches you. The squeaky wheel got grease. The rest of the cart did not.
I'm not sure the compute-metering idea is wrong, for what it's worth. Charging by actual load is more honest than pretending every prompt costs the same. The execution was the problem: ship aggressive caps on paid tiers with no cushion, watch the forums catch fire, then walk it back in public. Twice.
The open question is what happens to the broader Gemini limits once the Antigravity noise dies down. Google hasn't said whether the other tools get relief or whether Antigravity stays the exception because that's where the loudest, most technical users live. For now it's the latter. Watch whether that holds past the next billing cycle.
