Search for how much you can use ChatGPT, Claude or Gemini for free and you will find confident numbers: this many messages every few hours, that many a day. We went looking for where those numbers come from. It isn’t the vendors. On 27 September 2026, none of the three published a message limit for its free plan. Anthropic says outright that there is no fixed message count, and Google describes its limits as compute-based, varying with what you ask.
That is worth knowing before you pick a free plan, or pay to leave one. Here is what each vendor does commit to, what it leaves open, and how to find your own limit.
Everything below comes from each vendor’s own pricing and help pages, read on 27 September 2026 from the US. Plans, features and prices change often and differ by country, so check the linked page before relying on a detail. We did not measure usage limits by testing.
The short version
- All three free plans are usable for real work, and all three limit you in ways they don’t quantify.
- Claude and Gemini both describe limits that refresh every five hours and depend on how heavy your requests are.
- Paid tiers are sold as multiples of the free limit (“2x”, “at least 5x”), which are multiples of a number nobody publishes.
- The one hard free-plan number on any of the three pages is ChatGPT Free’s context window: 27K tokens, half of what Plus gets.
- Anthropic points free users to a usage screen. Trust what your own account shows over any article, this one included.
What each vendor commits to in writing
ChatGPT Free
OpenAI’s pricing page lists “Unlimited text chats” with GPT-5.6 Luna, with a footnote making that “subject to abuse guardrails”. Almost everything else on the free plan is marked Limited: file uploads, image generation, deep research, memory, data analysis and vision. Search, voice and projects are listed as included, as is using other people’s custom GPTs (creating your own is paid). Response times are “Limited on bandwidth & availability”, which is a polite way of saying free users wait when it’s busy.
The page does publish one hard number. The free plan’s GPT Instant model has a total context window of 27K tokens, with a user-input maximum of “~12 pages of text”. Go and Plus get 54K and Pro gets 128K. A footnote explains that the space for your input is smaller than the total, because instructions, memories and the model’s own working space share it. That is why a long free-plan conversation starts losing track of what you said earlier sooner.
Claude Free
Anthropic’s pricing page lists chat on web, desktop and mobile, web search, file creation with code execution, memory, connectors, voice mode and up to five projects. Research, Claude Code and the Opus models are paid-only; the free plan gets the Sonnet and Haiku models.
On limits, its FAQ is unusually direct. Usage depends on “the length and complexity of your conversations, the model you choose, and the features you use, so there’s no fixed message count.” Limits reset on a rolling five-hour window, and Pro is described as “at least 5x more usage per 5-hour session than Free.”
Gemini without a paid plan
Google’s plans page lists, for $0 with a Google account: access to Gemini 3.6 Flash, “Varying access to 3.1 Pro”, image generation and editing, Deep Research, Gemini Live, Canvas, Gems, and 15 GB of storage shared across Gmail, Drive and Photos.
Google’s help page on limits is the most candid of the three about how the system works. Limits are “compute-based” and factor in “the complexity of your prompt, the models and features you use, and the length of your chat”. They refresh every five hours until you reach a weekly limit. The page’s table gives the free tier as “Standard limits”, AI Plus as “2x higher than standard limits” and AI Pro as “4x higher”.
Side by side
| ChatGPT Free | Claude Free | Gemini, no plan | |
|---|---|---|---|
| Published message limit | None | None (“no fixed message count”) | None (“Standard limits”) |
| How limits are described | “Limited” per feature | Rolling five-hour window | Five-hour refresh, weekly cap |
| Published context window | 27K tokens | Not stated for Free specifically | Not stated on pages read |
| Projects or equivalent | Projects | Up to five projects | Gems |
| Research mode | Deep research: Limited | Research: paid only | Deep Research: included |
| Where to see usage | Not named on pages read | Settings > Usage | Not named on pages read |
Claude’s comparison table lists a context window of “Up to 1M, varies by model” for every plan, including Free, so it gives no free-specific figure.
“2x of what?”
Line those pages up and a pattern appears. Every paid plan is defined relative to the free one, and the free one is never defined.
That is not necessarily evasion. If a limit is measured in computing work, a message is not a fixed unit. A one-line question and a request to analyse a 40-page PDF with a reasoning model cost very different amounts to serve. A vendor that promised “40 messages” would have to count the cheap ones harshly or let heavy users burn through shared capacity. This is our reading, not theirs, but publishing a multiplier instead of a number looks like the honest version of a limit that genuinely varies.
It does mean the figures circulating in “best free AI” roundups are someone’s observation from one day, one account and one kind of use. They may have been roughly right once. They are not commitments, and Anthropic’s own FAQ says no fixed count exists.
Why the same free plan runs out faster for some people
The vendors name the same factors, so these are safe to act on.
- Long conversations cost more. Each new message sends the whole conversation back through the model, which is also why paid API bills grow with chat length.
- Heavier models and features cost more. Reasoning modes, deep research and file analysis draw more from the same allowance. Here is what “thinking” actually spends.
- Attachments cost more. A large PDF is a lot of text to process, and it stays in play for as long as it’s in the conversation.
So, to make a free plan last:
- Start a new chat for a new task instead of extending a long one.
- Use the default fast model for everyday questions, and save reasoning or research modes for problems that need them.
- Paste the relevant section rather than the whole document when you can.
- Put standing instructions in a project or Gem so you are not repeating them in every chat.
Two things the price doesn’t change
Training on your chats. OpenAI’s comparison table says, for every consumer plan including Free, “Content is used to train our models”, with “Opt-out available”. Anthropic’s lists “Opt-out” under model training for every plan. The Google pages we read did not address it. If that matters to you, find the setting before you paste anything sensitive.
Accuracy. Paying buys capacity and access to newer models. It does not buy an assistant that is never wrong, so the output still needs checking on either side of the paywall.
When paying is worth it
Based only on the published feature lists, an upgrade earns its cost when:
- You hit the limit mid-task regularly, not just now and then.
- You need something the free list excludes: Claude’s Research mode and Opus models, ChatGPT’s GPT-6 models, or Gemini’s video generation, which Google lists from AI Plus upward.
- You work with long documents in ChatGPT, where every paid tier at least doubles the free context window.
For scale, the US prices shown on 27 September were $4.99 a month for Google AI Plus, $19.99 for Google AI Pro, and $20 a month for Claude Pro, or $17 a month billed annually. ChatGPT’s page did not display prices in our session, so check it directly.
If you only occasionally hit the wall, the cheaper fix is usually the list above: shorter chats and lighter models.