Google moved fast and broke its own users' trust — then moved fast to fix it. At Google I/O 2026, the Gemini app quietly switched to compute-based usage limits, replacing the old simple per-prompt counting with a system that weighs request complexity, tool usage, and conversation length. Within days, paying subscribers were hitting their limits far faster than expected, and the backlash was loud enough that Google has already reversed course.
Per 9to5Google, Android Central, Android Authority, and Yahoo, Google is now capping how much quota a single prompt can consume and ensuring failed requests no longer count against your usage. It's a rare, fast public walk-back from Google — and a useful lesson in how AI subscription economics are getting squeezed. Here's the full story.
What Google Changed at I/O 2026
The original change, rolled out right after I/O 2026:
- Compute-based limits. Instead of counting individual prompts, Gemini began weighing factors like request complexity, tool usage, and conversation length to determine how quickly you burn through quota.
- 5-hour and weekly quotas. Google rolled out rolling 5-hour windows plus weekly caps — a drastic change from Gemini's previous, more generous prompt-based limits.
- Opaque consumption. The biggest user complaint: it became impossible to predict how much a given action would "cost" against your quota. A complex multi-tool query could eat a huge chunk in one shot.
Subscribers hit the new compute-based limits far faster than expected — and the complaints worked.
The Backlash — And Google's Fast Reversal
Paying Gemini subscribers — including AI Pro and AI Ultra tiers — found themselves hitting walls mid-workflow. The complaints flooded Reddit, X, and Google's own feedback channels. Per Android Central, "complaints worked." Google's response, confirmed across multiple outlets:
- Per-prompt consumption cap. Google is now capping how much quota any single prompt can consume — so one heavy request can't drain your whole allowance.
- Failed requests no longer count. If a Gemini request errors out, it no longer counts against your usage or available quota. (Previously, failed requests still burned quota — a major source of anger.)
- More predictable limits. Google is addressing the bugs that were exhausting quotas unexpectedly and making the system more transparent.
- Loosened overall caps. Per Yahoo, Google "loosened Gemini usage limits after subscriber backlash" — partially walking back the restrictive 5-hour and weekly windows.
The whole episode played out in roughly a week — from rollout to backlash to fix. That's unusually fast for Google, and it signals how sensitive AI subscription retention has become.
Why This Happened: The AI Economics Squeeze
Compute-based metering reflects the real cost structure of AI — but users want predictability, not accuracy.
The underlying reason Google switched to compute-based limits is honest: different requests genuinely cost wildly different amounts. A simple "what's the weather" query costs almost nothing. A long multi-tool agentic workflow with a 1M-token context costs Google real money. Per-prompt counting treats them equally — which means power users effectively subsidized by light users, and Google eating the difference.
Compute-based metering is "fairer" in a cost-accounting sense. The problem: users don't want accuracy, they want predictability. Nobody can mentally estimate how many "compute units" a query will cost. The opacity broke the core promise of a subscription — pay once, use freely.
This tension is everywhere in 2026. As DeepSeek's price war compresses margins, every AI provider is hunting for ways to control compute costs without driving users away. Google just learned the hard way where the line is.
What This Means For You
- → Gemini AI Pro/Ultra subscriber: Good news — the worst of the limits are being walked back. Failed requests no longer count, and single prompts can't drain your quota. Re-test your workflows.
- → Heavy AI power user: Expect compute-based metering to become the industry norm eventually. Google blinked this time, but the economic pressure isn't going away. Have a backup provider.
- → Building on the Gemini API: API pricing is separate from the consumer app limits — but watch for similar compute-weighted pricing models spreading to API tiers.
- → Considering an AI subscription: This is a reminder to read the fine print on usage limits. "Unlimited" rarely means unlimited in 2026. Compare Gemini, ChatGPT Plus, and Claude Pro limits before committing.
FAQ
What are Gemini's new usage limits?
After I/O 2026, Google switched to compute-based limits with rolling 5-hour windows and weekly caps. After backlash, it added a per-prompt consumption cap and stopped counting failed requests. The system now weighs request complexity rather than counting individual prompts.
Did Google reverse the Gemini limits?
Partially. Google loosened the limits after subscriber backlash — capping how much a single prompt can consume and ensuring failed requests don't count. It didn't fully return to the old system, but it addressed the worst pain points within about a week.
Do failed Gemini requests still count against my quota?
No — per Android Central, Google fixed this. Failed Gemini requests no longer count against your usage or available quota. This was one of the biggest sources of user anger in the original rollout.
Why did Google switch to compute-based limits?
Because different requests cost wildly different amounts of compute. A simple query is cheap; a long agentic workflow is expensive. Compute-based metering reflects real costs more fairly — but users found it unpredictable, which broke the subscription's core appeal.
Will other AI apps adopt compute-based limits?
Likely, eventually. The economic pressure that drove Google's move affects every AI provider. But Gemini's backlash is a cautionary tale — expect competitors to design more transparent, predictable versions if they go this route.
Final Word
Google's Gemini limits saga is a small story with a big lesson: in 2026, AI subscription economics are tight enough that providers are experimenting with how they meter usage — and users will revolt the moment the experience becomes unpredictable. Google's fast reversal shows it understands retention matters more than perfect cost-accounting, at least for now.
The bigger takeaway: as the price war grinds on, expect more of these awkward squeeze-and-retreat moments across every major AI app. The providers are all trying to figure out how to make the unit economics work — and you, the user, are the test subject.
📩 AI subscription changes + pricing wars, tracked weekly.
Subscribe to Tech4SSD. Free. Sign up →
Sources: 9to5Google, Android Central, Android Authority, Android Police, Yahoo Tech. Reporting accurate as of May 22, 2026.