Google Overhauls AI Usage Limits with New "Compute-Based" System
Following significant user backlash, Google has rolled out and subsequently adjusted a new "compute-based" quota system for its Gemini AI models and tools. The system, which took effect in late May, replaces a less defined model and assigns limits based on the complexity of prompts, conversation length, and features used.
"A simple text prompt consumes less compute than a complex video or coding prompt."
New Compute-Based System
On May 20, Google announced the shift to a compute-based usage limit for the Gemini app. Key aspects of the new system include:
- Dynamic Calculation: Limits are calculated based on prompt complexity, tools used, and chat length.
- Quota Refresh: The quota refreshes every five hours until a user reaches an overall weekly limit.
- No Penalty for Errors: Quota is only consumed for successful completions; errors do not count.
- Prompt Capping: The amount of quota a single prompt can use is now being capped for the Gemini 3.1 Pro model, addressing issues where complex prompts with large files rapidly depleted the limit.
- Bug Fixes: A bug where one or two Omni videos drained quotas for certain users has been fixed. AI Ultra users now have double the number of Omni generations.
Tier Structure and Paid Plans
The new compute-based limits replace a previous system where Google used vague terms like "more" and "higher" without quantifying usage differences. The current paid plan limits are as follows:
- Free Plan: Serves as the baseline.
- AI Plus ($7.99/month): Offers double the free plan limit.
- AI Pro ($19.99/month): Offers quadruple the free plan limit.
- AI Ultra ($100/month): Offers 5x the Pro limit.
- AI Ultra ($200/month): Offers 20x the Pro limit.
User Incident and Reported Frustration
In a notable incident, AI Pro subscriber Ashutosh Shrivastava reported on social media that a single prompt using Gemini's avatar-based video generation feature consumed their entire five-hour compute-based limit. The video generation failed after running for three to four minutes, at which point usage reached 100%.
Google's Gemini lead, Josh Woodward, acknowledged the complaint, stating, "Yikes, let us take a look!"
Multiple users on the Gemini subreddit have criticized the new system for being less predictable and more restrictive. Some users have expressed frustration that the new limits are compared against the free plan rather than prior Pro allowances, and some have questioned the value for money.
Adjustments to Antigravity Coding Tool
In response to user complaints and reports of rapid limit exhaustion, Google increased the limits for Gemini models within its AI-powered coding tool, Antigravity, twice within one week.
- First Adjustment: On Wednesday, Google raised the Gemini model rate limits for Antigravity by a factor of 3 and reset the weekly quotas for all users. This came after many users reported hitting the limits within an hour of use.
- Second Adjustment: Later that week, Google increased the weekly quota by another 3x after acknowledging that users could reach the limit after a few work sessions. Quotas for all paid plans were reset for the second time.
- No changes to usage limits have been made outside of Antigravity.
Separate Limits for Thinking and Pro Models
Google has implemented independent usage limits for its Gemini 3 Thinking and Pro models, separating them from a previously shared pool. Previously, both models operated under a single shared limit (100 prompts per day for AI Pro subscribers and 500 per day for AI Ultra subscribers).
Google stated the change was in response to user feedback requesting more precision and transparency.
The new daily prompt allowances are:
AI Pro Subscribers:
- 300 Thinking prompts per day (optimized for complex problems)
- 100 Pro prompts per day (optimized for advanced math and code)
AI Ultra Subscribers:
- 1,500 Thinking prompts per day
- 500 Pro prompts per day
Independent limits also apply to free users, with both models listed under "Basic access."
Future and Free Features
- More Transparency: Google has stated it will provide more detailed usage breakdowns and notifications for heavy tasks like Deep Research.
- Pay-As-You-Go: In the future, Gemini app users will be able to buy pay-as-you-go top-up AI credits.
- Free Model Usage: Prompts using the Gemini 3.1 Flash-Lite model are now free and do not count against quota.
- Session Memory: When a user selects a specific model, that choice is remembered across sessions unless manually changed or a cap triggers an automatic fallback to a lighter model.