AI Pro Subscription

Update: An hour after the RPD on Argon on my second account, I was able to generate another message using Argon. So, it’s possible that the previous limit behavior has returned to that account.

November 5, 2026

The chat costs approximately 450k–500k tokens.

About 25–30 messages on Gemini 3.1 Pro before hitting the limit, and the same number on 3.0 Flash before it hits its limit.

Both models report that they’ve reached the token usage limit, but if you upload a PDF to the new chat with a token count comparable to the number of tokens in the chat, they continue to work normally.

Honestly, this TPD (or is it not TPD?) is extremely annoying and interferes with the normal use of AI Studio.

November 8, 2026

About 30 messages from Gemini 3 Flash in a chat, totaling roughly 450k tokens. After that, I hit the token limit. Of course, as usual, when I open a new chat and upload a PDF containing 470k tokens, the model responds normally.

@Aakash_Bansal @Mustan_lokhand @Logan_Kilpatrick @Payal_Sharma2 @Gemini_AIS_Support @Jon_Matthews

Hello, Google team! It’s been over a month without a response—what’s going on with the limits? In fairly long chats, it feels like the limits are being handled incorrectly, to put it mildly. After, at best, about 30 messages, the model returns a “daily token limit reached” error. However, if I open a new chat and upload a PDF, video, or other document with a number of tokens comparable to the number of tokens in the chats where the model hits the limit, the model generates a response normally. Would you be so kind as to answer the question: is this intended behavior or a bug?

  1. If this is intended behavior, can you explain how the paradox I described above occurs in this case? Also, if this is intended behavior, can you explain why the system does not take into account the token cost for the model, and whether, for example, Gemini 3.1 Pro and Gemini 3.0 Flash, the number of requests before the token limit is comparable?

  2. If this is a bug in how chats with extensive context work, can you indicate when it will be fixed? This significantly impacts the user experience of AI Studio with a Google AI subscription.

Please do respond. Any response is better than silence.

Quick update regarding the limits issue:

I tried saving a 450k-token conversation into a Markdown (.md) file, which came out to ~300k tokens of raw text.

In a brand-new chat, uploading this .md file immediately triggers the “daily token limit reached” error on the very first prompt. However, uploading a 470k-token PDF file into a fresh chat processes and generates responses completely fine.

This further proves the theory that the quota limit isn’t actually being reached at the model level (since all tokens reaching the model’s attention mechanism should be computationally equivalent). Instead, the API gateway / rate-limiting middleware seems to be bugged: it specifically panics and flags raw inline text payloads while bypassing File API references.

Here because I have the same issue with the seemingly decreasing quota. Within the first few weeks of Pro subscription I was never afraid of running out of tokens. But now only half way after the weekly reset and I am straining tokens for the rest of the week. I am afraid my quota will run out by next mid-week.

November 11, 2026

A chat costing about 900k tokens. After about 15 requests to Gemini 3.1 Pro, a token limit error occurred.

I think I’m starting to understand: it’s unclear why the system counts how many tokens are sent via RAW text, and if the total number exceeds a certain threshold when sending such tokens, the site returns an error.

Given that large PDF models continue to be processed in new chats, if PDF tokens and RAW tokens are equivalent for model processing, the model’s own limit hasn’t been exhausted.

@Logan_Kilpatrick @Mustan_lokhand @Jon_Matthews @Aakash_Bansal @Gemini_AIS_Support

Google team, would you please explain what’s going on and stop keeping us in the dark? In my opinion, a limit on the number of requests is sufficient. A limit on the number of tokens is completely unnecessary.

Hello,

Thanks for the detailed reports and feedback. As a Google AI Pro or Ultra subscriber, you get higher per-model daily limits in AI Studio Playground (and Build) than on the Free tier, and limits reset daily rather than using rolling multi-hour windows (Google AI plans). Very heavy use of one model in a single day—such as long conversations or large text uploads, where each turn includes the chat history—can still reach that model’s daily limit while a shorter new chat continues until the remaining daily balance is used (Daily usage limits for models in AI Studio).

Regarding model availability, Gemini 4 Argon is included with Google AI Pro in both Playground and Build (Gemini 4 Carbon requires Google AI Ultra or a paid API key); for retired models such as Gemini 2.5 Pro, see Gemini deprecations.

  • Switch to another available Gemini model in Playground (limits are tracked per model), or start a fresh chat with a shorter context.
  • For sustained long-context or high-volume work without Playground limits, use your own paid Gemini API key via AI Studio Billing (eligible subscribers can also claim monthly Google Cloud credits on the Google Developer Program dashboard).
  • If you are hitting limits after only a few short prompts in a brand-new chat or still seeing “Permission denied. Please check your account permissions and project settings” with no API key selected, please share your Google account email used for AI Studio using this form: Google AI Developer Forum: Share Details About Your Issue. Please only use this form if your issue is the same as the one in this thread; if it’s different, start a new post instead.

Thanks for the response! However, the “remaining daily balance” theory doesn’t match the mathematical data.

If a 900k-token chat supposedly uses up ~99% of my model’s daily balance to trigger the limit, uploading a fresh 500k-token PDF in a brand-new chat should fail immediately due to insufficient remaining balance. Yet, the 500k PDF processes and generates responses without any issues.

This proves it is not a shared account-level daily balance, but rather a bug in how turn-by-turn inline text state accumulation is evaluated by the rate-limiting proxy.