Google doesn’t usually move this fast. Just days after rolling out one of the most significant changes to how Gemini usage limits work across its entire AI ecosystem, the company found itself tripling those limits for Antigravity, its agentic coding IDE, not once, but twice in the same week. It’s the kind of rapid, public course-correction you rarely see from a company Google’s size, and it tells you everything you need to know about how developers actually felt about what launched at I/O 2026.

What changed with Gemini limits after I/O 2026

The story starts with a billing model shift that, on paper, sounded reasonable. Rather than counting a fixed number of daily prompts, Google announced it was moving to a “compute-based” usage system for Gemini. Under the new system, Gemini limits are calculated based on the complexity of your request, the features you use, and the total length of your conversation history. In Google’s own words, it’s “a better way to allocate limits, because a simple text prompt uses far less compute than a complex video or coding prompt.”

The change went live alongside the I/O 2026 announcements, and the new subscription tiers looked clean on a chart. The $7.99-a-month AI Plus plan doubles the limits of the free tier. AI Pro at $19.99 a month gets you four times the free-tier allowance. A new AI Ultra tier at $100 a month offers five times the AI Pro limits, sitting alongside a higher-end $250-a-month plan. Weekly limits refresh every five hours until you hit the weekly cap.

What Google may not have anticipated was how fast those caps could disappear in actual use.

Developers hit the walls fast

The frustration landed almost immediately. Paid subscribers across Reddit and Google’s official developer forums started reporting that a handful of prompts, or a few video generations, was enough to completely drain their quota. One Reddit user shared that a basic five-post back-and-forth burned through 50 percent of their entire five-hour compute limit. That’s not a power user abusing the system. That’s a normal work session.

I’ve been watching agentic coding tools closely for a while, and honestly, this outcome wasn’t hard to predict. Agentic tasks are computationally expensive by design. A single Antigravity session can spawn multiple agents, run parallel tasks, and iterate across your terminal, browser, and editor simultaneously. Pricing and limit systems designed around traditional chat prompts were never built for that kind of workload.

The complaints piled up fast. Some users said failed media generations were still counting against their limits. Others reported being automatically switched from the Pro model to Flash even after explicitly selecting Pro preferences. The backlash was loud enough that Varun Mohan, a Director at DeepMind who works directly on Antigravity, had to step in publicly.

Google’s response: 3x, twice

What Mohan announced was direct. On Wednesday, Google tripled the Gemini model rate limits across all paid tiers in Antigravity and reset everyone’s weekly quota. The message that appeared in premium dashboards was short: “If you are on a paid plan, your Gemini quota has been reset for the week and increased by 3x moving forward. Keep building!”

That part was already notable. But then it happened again.

The second round of increases, also reported by 9to5Google, specifically targeted the weekly quota limit. Mohan acknowledged on the record that users were hitting their weekly Gemini limits “after a couple work sessions,” and confirmed that Google had reset quotas for all paid plans for the second time in a single week. The Google AI Developers account on X also confirmed the triple, noting that the higher rate limits across all paid Antigravity tiers are now permanent.

After looking into this more closely, I can tell you that this kind of public, twice-in-a-week reversal is genuinely unusual for Google. The speed of the response suggests the developer backlash hit hard enough to override whatever internal calculus produced the original limits in the first place.

What this means for Antigravity and Gemini users

There’s still a real caveat worth naming. Even with the 3x increase, users quickly pointed out that the new limits are still lower than what they were used to before the compute-based system kicked in. The 9to5Google reporting was clear on this: the higher Antigravity quotas reduce the immediate friction for paying developers, but those limits will inevitably still have an impact. And critically, the limit increases only apply inside Antigravity at this time. Gemini’s usage caps everywhere else, including the main Gemini app and other Google tools, have not changed.

There’s also a broader technical context that makes the situation harder to fix with a simple multiplier. Gemini 3.5 Flash, the model powering much of Antigravity 2.0, appears to be significantly more resource-intensive than the previous Flash generation. Early data is already showing that even single prompts with 3.5 Flash can consume a meaningful percentage of a user’s available compute budget.

This is one of those stories I genuinely got excited about the moment I started digging into it, not because of the drama, but because of what it reveals about where AI tooling is heading. The old model of flat daily prompt counts was already on borrowed time. Agentic systems don’t work that way.

What comes next for Gemini limits

Sources suggest Google is still actively working on the compute metering formula and plans to refine how limits are calculated based on feedback from Antigravity users. The Antigravity team has acknowledged making early mistakes and pledged to continue acting on developer input, which at the very least signals that this won’t be the last adjustment to Gemini limits in the near future.

Industry observers also hint that the launch of Antigravity 2.0, with its native CLI tool, updated desktop app, and SDK for custom agent workflows, brings with it a fundamentally different usage profile than anything Google has had to meter before. Developers building custom subagent workflows and scheduling background tasks are in an entirely different category from someone asking Gemini a question in a chat window.

The 3x reset is a meaningful short-term fix. But the harder work, getting the compute model right for a world where agents run in parallel for hours at a time, is still very much in progress. What’s clear is that Google is listening, and Antigravity users are making sure their voices carry.