Google raised usage ceilings for Gemini users and Google AI subscribers using Antigravity at Google I/O 2026, and the complaints from developers and heavy users haven't slowed down. The core grievance isn't that the new limits are too low. It's that nobody can tell, before starting a session, whether they'll finish it.
Starting May 17, Google replaced fixed prompt allowances with a compute-based consumption model where the cost of any interaction varies based on request complexity, active tools, and accumulated conversation length.
No pre-request cost estimates are published at any tier. The same documentation notes that limits may change without notice and can be adjusted during periods of high activity. That's not a footnote buried in terms of service. It's a description of how the system is designed to work.
This piece focuses on developers and heavy Antigravity users, who are most exposed to what the compute model actually does in practice. The broader subscription changes are context. Antigravity workflows are where the unpredictability breaks real work.
What changed at Google I/O: new tiers, removed credits, relative limits only
The headline announcement: a new $100/month AI Ultra tier offering up to 5x higher usage limits in Gemini and Antigravity compared to AI Pro, plus a price cut on the existing top-tier plan from $250 to $200/month. These are real changes. For casual users with short, light sessions, more headroom likely makes a difference.
The multipliers are relative, not absolute. Google's support page lists AI Plus at 2x the standard limit, AI Pro at 4x, and AI Ultra at "5x or 20x higher than AI Pro limits depending on your subscription," per the documentation. What Google has not disclosed is the underlying compute allowance those multipliers are applied to. Five times an unknown number is still an unknown number.
Google also removed the 1,000 monthly AI credits previously bundled with AI Pro. Extended access to Antigravity and Flow now requires purchasing additional credits separately. The net result: more plan options, higher relative ceilings, fewer included benefits, and no published numbers that would let anyone independently verify the value of any of it.
How the Gemini AI quota changes work and where they fall apart
Google's new system grades each request on three factors: how computationally complex the prompt is, which tools and features are active, and how long the current conversation has been running. A short message in a fresh thread costs less than the same message sent deep into a tool-heavy session. Google does not publish pre-request cost estimates, so users have no way to check how much a given interaction will cost before sending it.
For developers running multi-step agentic sessions, where Antigravity is coordinating subagents, maintaining extended context, and executing code across a long thread, those are exactly the conditions the compute model penalizes most. Quota burns faster as session length and tool use increase. The longer and more complex the work, the harder consumption is to estimate a pattern consistent with user reports and confirmed by how Google describes the system in its own documentation.
The refresh mechanism adds another layer of friction. Google advertises a five-hour rolling quota refresh alongside a broader weekly cap, per MakeUseOf. What users have actually encountered since earlier this year is different. In posts on Google's AI Developers Forum from earlier this year, users described countdown timers showing hours until refresh that abruptly reassign to a "baseline quota refresh" date four to ten days out, with no explanation for the jump, per the forum thread. Support tickets have largely produced generic replies pointing back to documentation that doesn't explain the discrepancy; the same thread shows.
The volume of complaints is hard to dismiss. Some Gemini 3.1 Pro users report consuming 60% of their available quota within a single session before hitting a forced wait. Multiple developers reported exhausting their weekly limit after only a handful of requests. The same pattern appears across Reddit, Google's own developer forum, and support threads.
A larger ceiling doesn't change any of this. Five times an undisclosed baseline doesn't make consumption estimable, and it doesn't make the five-hour refresh behave as advertised. These are separate problems that a quota increase doesn't address.
What happens when Gemini usage limits run out
Subscribers who exhaust their quota aren't paused or queued. The system automatically shifts them to the lighter Gemini 3.5 Flash model. Google's position is that 3.5 Flash now matches 3.1 Pro's reasoning quality at faster speeds — a claim that users broadly reject.
Across forum reports and user feedback, 3.5 Flash is described as hallucinating more frequently than prior versions and underperforming on complex coding and reasoning tasks, according to MakeUseOf and developer forum posts. These are user-reported observations, not benchmark results. But the consistency of the pattern across independent sources makes it difficult to attribute to noise.
For a developer mid-session, a forced model switch is a workflow failure. The session that started with 3.1 Pro's reasoning capabilities continues with a weaker model, at the point where context is longest, and the work is already underway. Paid subscribers aren't just losing access when quotas run out; they're losing the model they paid for, without warning, in the middle of a task.
The June 18 CLI migration adds a hard deadline
On June 18, Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for AI Pro, AI Ultra, and free-tier users, replaced by the newly announced Antigravity CLI. Google confirmed directly that there will be no one-to-one feature parity between the two tools at launch.
The licensing shift matters too. Gemini CLI carried an Apache 2.0 open-source license. The repository will remain available under those terms, and Google says it will continue receiving model updates, bug fixes, and security patches for enterprise customers. Antigravity CLI does not appear to be open source, based on what Google has published. It launches with support for agent skills, hooks, subagents, and extensions. But the migration is forced, not voluntary, and the feature gap is Google's own acknowledged fact, not speculation.
This matters for the quota story because it compounds the timing problem. Developers are being asked to adopt an unfamiliar tool with acknowledged gaps, at the same moment the underlying quota system is behaving inconsistently. Both issues reflect the same pattern: changes to a professional workflow, on Google's schedule, without the predictability that professional work depends on.
What a real fix would require
Google's I/O announcements are substantive. A new $100/month entry point, a price reduction at the top tier, and higher relative limits across plans are meaningful changes, particularly for lighter users. The structural problem they don't solve is predictability.
Higher ceilings don't explain why five-hour refresh windows have been producing multi-day lockouts documented in developer forums since earlier this year. They don't close the gap between what Google's documentation describes and what users are actually experiencing. And Google's own support policy explicitly preserves the right to adjust limits without notice, per the documentation.
The benchmark for a real fix is specific: Google publishes usable compute thresholds at each tier, and the refresh behavior matches what the documentation says. Until both of those are true, upgrading to a higher tier buys more headroom, not reliability. The complaints will keep coming. They won't tell you when the underlying problem has actually been fixed.
Comments
Be the first, drop a comment!