"You've Reached Your Limit": What Nine AI Services Don't Say About Their Free Plans

Nine services, 70 official pages preserved and timestamped, and one shared result: none of them publishes a numeric cap for free chat messages.

Update (29 September 2026): We updated this report to reflect changes across several official help centers, including clarified phone number requirements and revised cross-surface documentation.
Subsequent updates will be made on the Doxascope website, on the specific lens page.
Prefer reading in Arabic? Read the Arabic version on Substack

You type your question, and the system answers with one line: you've reached your limit.

When does it come back? Could you have known in advance? And the service you are about to switch to, is it any better, or is its limit waiting for you a few messages later?

The answers should be published: every company maintains pricing pages, help centers, and terms of use. But the real answers are scattered across dozens of pages, many with no clear publication date, and the values can change without clear notice. In fact, two of them changed while we were working on this report.

We collected what nine companies say about their free plans: ChatGPT, Gemini, Meta AI, DeepSeek, Copilot, Claude, Grok, Perplexity, and Mistral Vibe. We saved a copy of every page we relied on, and tied each value to the sentence that states it. We don't rank the services or recommend one. We show what each company says, when we last checked, and what it leaves out.

What the free plan gives you today

This is what we were able to document as of 29 September 2026; the live record stays updated on the lens page, with the quote, source link and last-verified date for every cell.

How to read the tables: An unmarked value comes from the company's own page. * means the source is the app's store listing, which the company also writes. † means an independent press report. "Disputed" means two official pages of the same company disagree. "Not published" means the company does not publish the specific value requested, or gives a description that leaves it unresolved. "Unknown" means we searched and found no source that settles the field at the required scope; where we looked is on record.

Models and limits

Tools

Context, availability and account

The context window is how much text the model can keep in view within one conversation. It is measured in tokens; a token is roughly part of a word.

What we found

You cannot know when the service will stop until it stops. None of the nine publishes a numeric cap for free chat messages, so the "limit reached" message cannot be predicted from the official pages. Companies take three positions on the question. The first publishes numbers for single tools, not for messages: OpenAI says everyday text chats are not capped by count and limits file uploads to 3 a day; Perplexity gives 3 Pro Searches a day and one Research query a month. The second publishes the reset time but not the count: Google says the Gemini limit refreshes every 5 hours up to a weekly cap, and Anthropic says the Claude limit resets on a 5-hour window. The third gives adjectives only: Mistral says "limited" and adds that Pro gives up to 6x, xAI says "within generous limits" that reset on their own schedule, Meta says it is testing limits on some features, and Microsoft says "a limited free version". For DeepSeek's free app we found no published limit at all.

Published figures for the context window range from 27K to 1M tokens — but the comparison is not equal in its recency. Published context windows range from 27K to 1M tokens. Two companies publish a fixed figure covering the free plan: OpenAI publishes 27K for Instant mode and Google 32K. DeepSeek's last official figure is 1M, in the notice of 24 April 2026, published in the developer documentation rather than on a consumer page, and it is not a measurement of today. Meta, Microsoft, Grok, Perplexity and Mistral Vibe publish no context figure for their free plans.

And the value you read today may change tomorrow, with no announcement. That is what happened while we were working: Anthropic used to publish a fixed figure for the free Claude context window, and our saved copy of 15 September shows "200k" in its table. In the copy of 22 September 2026 the table reads "Up to 1M, varies by model" for every plan. We did not erase the old value. We closed its record, wrote down what changed, and opened a new record with the current value. The change itself became a documented fact, instead of a value quietly replaced.

The four tools are available on most services, and the difference is in the exceptions. Image generation is confirmed on six services: ChatGPT, Gemini, Meta AI, Copilot, Grok and Mistral Vibe, and it is not available on Claude. Perplexity's plans page says no, while its image-generation help article says the Free plan gets limited image generations, and only two days separate the dates of their last edits. DeepSeek remains unknown. Web search is confirmed on all nine, but on Gemini only within Deep Research, and no official page we collected names a search tool for ordinary chat. File upload is confirmed for the free plan on six services, while the pages of Claude, Copilot and Grok describe uploads for the whole product without saying the free plan includes them, and the same holds for voice in Perplexity's iOS app; so we record those cells as unknown for the free plan.

And in some Arab countries, you may not find your country on the support list at all. Only three services publish a country list: the ChatGPT and Gemini (web) lists name 21 of the 22 Arab League countries, and Claude's names 20, leaving out Syria and Yemen. An absence from a list means the company does not name the country among those it supports, not that it announces a block, although OpenAI does warn that use outside its list may get an account suspended. The rest publish no list at all: Grok, Perplexity and Mistral Vibe have none, Microsoft says "over 170 markets" without naming them, DeepSeek says availability varies by jurisdiction, and Meta publishes no list; what we have for it are press reports from February 2025.

And before you create an account, two conditions differ between services: age and a phone number. Seven services publish a minimum age of 13, Claude stands alone requiring 18, and DeepSeek publishes no age. As for the phone number: Claude confirms it is required, and three services confirm that it is not required. OpenAI says it is not generally required to use ChatGPT, although it may be requested when accessing Codex in the desktop app. DeepSeek varies by version and region¹³, and we found no official answer for Grok, Perplexity or Mistral Vibe.

What about the same model answering differently across the app, web and API? That is a separate question, and it turned out to be broader than a feature comparison. We have held that investigation back until it includes a deeper review of company policies, model behavior and direct cross-surface testing.

How we know

Each value climbs a ladder of sources and carries the name of the rung it came from:

1. An official page: pricing, help center or terms.
2. An official source inside the product, such as the app's store listing.
3. A press report, or independent testing with a published method.
4. Our own measurement with a free account. This rung is paused for now.
5. Several user reports, shown with their count.

We write "Unknown" only after searching, and we record where we looked. The full method is on the method page.

We keep a copy of every page and check it again each week, because official pages change, and so do their dates, and because a page's own date may not match the date of the event it describes¹⁶.

Tell us

If you have a source that contradicts a value here, or a measurement from your own free account, send it to us. And tell us which question to answer next: why the answer differs across interfaces, or what paid plans actually add?

1. ChatGPT's free plan also lists GPT-5.6 Terra, but only as limited access in Work and Codex on desktop. ↩

2. The File Uploads FAQ says OpenAI may lower this limit during peak hours. ↩

3. The limits DeepSeek publishes (2500 and 500 concurrent requests) apply to its API, not to the free chat app. ↩

4. Microsoft's help says pressing "Quick response" under the prompt box changes the mode. It names no model for Quick response, and names GPT-5 for Smart mode only. ↩

5. Gemini's two app-store listings mention Gemini Live without tying it to a paid plan, while they explicitly tie video creation to paid plans. ↩

6. The source is DeepSeek's app launch notice of 15 January 2025. File upload is also confirmed by an App Store version note (July 2026). For web search we found no newer official page that mentions it. ↩

7. An official page describes the feature for the whole product but does not say the free plan includes it: file upload on Copilot, Claude and Grok, and voice in Perplexity's iOS app. So we record it as unknown for the free plan. ↩

8. Perplexity's plans page (last modified 5 September 2026) says the Free plan has no image generation. Its image-generation help article (last modified 3 September 2026) says: "Free Plan: Get access to limited image generations." ↩

9. OpenAI may request phone verification when accessing Codex in the ChatGPT desktop app; email or an authenticator app cannot replace this step. ↩

10. WhatsApp's help center says Meta AI is available in "limited countries", with no list. TechCrunch (24 February 2025) names 10: Algeria, Egypt, Iraq, Jordan, Libya, Morocco, Saudi Arabia, Tunisia, the UAE and Yemen. A Wamda report confirms 5 of them: the UAE, Saudi Arabia, Egypt, Morocco and Iraq. ↩

11. On meta.ai you need a Meta account, which can be created with an email address or a phone number. This does not cover Meta AI inside WhatsApp. ↩

12. DeepSeek's V4 notice of 24 April 2026, published in the developer documentation, says a 1M-token context is the default across all its official services. No consumer page restates the figure for the free chat app. It is the last official figure we found, not a measurement of today. ↩

13. The Chinese privacy policy (updated 10 February 2026) says registration and login need the user's own mobile number under China's real-name rules. The English terms and the app notice name an email or a third-party account, and neither names a phone number. ↩

14. From Microsoft's Copilot Supplemental Terms (effective 18 August 2026): at least 13, and sometimes older depending on the country's laws. ↩

15. In our copy of 15 September 2026, the plan table read "200k" for the Free plan. In the copy of 22 September 2026 it reads "Up to 1M, varies by model" for every plan. ↩

16. An example: the current copy of Anthropic's official Citations page carries the date 23 June 2025, while Anthropic's official changelog records the launch of the Citations API on 23 January 2025. So we do not rely on the current page alone to reconstruct the timeline. ↩