Me and a bunch of other users have noticed that over the past few days the rate limits for 2.5 and 3 have been slashed dramatically.
I’m getting limited after approx. 18 requests.
Anecdotally, I’ve also noticed that 3 isn’t performing as well as it usually does (eg making loads of spelling and grammatical errors which it hardly did when 3 was launched , ignoring the system prompt more than usual).
Please could it be confirmed if this a temporary issue or a permanent change to the rate limits going forwards?
Hey folks! Yes the limits are shifting around given capacity issues. You can always switch to a paid API key if you want to keep a conversation going, and we are exploring allowing G1 subscriptions to work too!
Can I get some insight on context caching within the Playground? I know it does not allow explicit caching via the UI, but how effective is implicit caching? My only concern is not having to burn through cash when regenerating responses or getting a bill that’s way out of my budget.
I am probably in the wrong spot but does any one know if google is offering grants for research use? I designed a system that is worth researching further on agentic AI systems, And I believe I’m real close to a cure for pancreatic cancer, Still speculative but grounded on real techniques. Grounded:
Low metabolic activity in necrotic tumor cores is a real biological issue. Cells with minimal transcription/translation may fail to express any kill-switch proteins.
Bystander lysis via holin/endolysin release from neighboring PSPs is a real mechanism, but diffusion limitations in dense tissue may prevent reaching all dormant PSPs.
Speculative / Caveats:
Complete clearance of quiescent cells is unlikely. This is a best-effort mitigation, not a guarantee.
2. Circuit Mutation / Cheater Strains
Grounded:
Selection pressure favoring kill-switch escape mutants is well-established in synthetic biology; “cheater” strains can emerge in long-term cultures.
Linking essential functions (replication genes) to a toxin-antitoxin system is a proven method to enforce circuit integrity.
Speculative / Caveats:
Evolutionary escape is never zero-risk. Over many generations, especially in vivo, some PSPs could accumulate mutations bypassing the kill-switch.
3. Decounter Drift
Grounded:
Uneven plasmid segregation during division is real; plasmid-borne replication counters are inherently stochastic.
Speculative / Caveats:
Moving the decounter to the chromosome improves inheritance fidelity, but precise “generation counting” over multiple doublings remains beyond current synthetic biology, especially in a diverse, heterogeneous swarm.
4. Truthfulness Table
Failure Mode
Status
Notes
Quiescent Escape
Partially grounded
Dormant PSPs can survive; bystander lysis mitigates locally but cannot guarantee full swarm clearance.
Circuit Mutation
Grounded
Cheater strains emerge; toxin-antitoxin linkage is a proven containment strategy.
Decounter Drift
Partially speculative
Chromosomal integration improves stability, but precise replication counting is still unreliable.
5. Summary
The “Stochastic Leak” analysis is conceptually correct and grounded in real synthetic biology risks.
Mitigation strategies are plausible, but none can guarantee 100% fail-safe clearance due to biological variability, spatial heterogeneity, and evolutionary pressure.
This section shifts from design speculation to risk modeling, which is exactly what PhD-level system design should do.
Am rewriting this as it got deleted last time. It seems like a lot of people are also experiencing a new rate limit that’s even lower than the aforementioned 18 since today. Based on what others have said and my own testing, it appears to be 10 prompts.
Posting this in case it helps others who might be experiencing the same issue.
Nano Banana doesn’t work well. For a single request, I have to run multiple generations, repeating the same question, and after 10 requests, I finally get the result from a prompt that the system claims is correct and accurate. So, something is sabotaging our generations. Deliberately. What’s the point?
I’m willing to pay a significant monthly fee, but all the features that Nano Banana used to offer, like face swapping, must works; otherwise, what we’re seeing now is disrespectful to the users. If the service works excellently it’s good; otherwise, it’s a toy.
I hope the problem it’s temporary. I can post dozens of cases where a prompt last week used to work fine, and today the system completely gets the wrong result. People are happy to pay when things works as promised. Before we running 50-100 generations every day, now find fewer than 12 requests, and Nano Banana hasn’t gotten a single one right. The system is wrong, and I have to pay a price without getting the good result? This policy is absolutely not right.
Hope to see some major improvements, thank you staff for reading this.
you are so right and i agree with you. we used to get 50 to 100 messages with the pro models. and then that went to 21 and now recently yesterday got me 9. i also hope this is temporary cause if ai studio keeps this up. it’s not looking for them
They killed the context window and screwed with the system so much I started demanding a full refund of all past Ultra payments. Pathetic that they will allow competition to steal away most of their customers.
