Free AI Tiers in 2026: What ChatGPT, Claude, Gemini and Perplexity Actually Give You

A friend installed four AI apps in one weekend because a video told him to. By Tuesday he was annoyed: one had stopped replying, one had swapped itself for a different model without saying so, one had used up his "research thing" for the month, and the fourth was still going. He wanted to know why. That question is worth more than a feature list, because the answer is not which app is best — it is that these four free plans ration entirely different things. Once you know which thing each one rations, you stop hitting walls by accident.

⚡ Quick answer

  • Most talking for free: ChatGPT — text chats became genuinely unlimited on 6 August 2026.
  • Best answers per message: Claude — and the fewest messages of the four, roughly 15–40 per five-hour window.
  • Most features thrown in: Gemini — 100 images a day, audio overviews, some Deep Research, on a compute-based budget.
  • Best sourced research: Perplexity — unlimited basic search with citations, but Pro Search is capped at a handful a day.
  • Longest documents: Claude — 200,000-token context on free, versus 32,000 on Gemini free.
  • If you install one: ChatGPT. If you install two: ChatGPT plus Claude, or ChatGPT plus Perplexity if your work is research-heavy.

What actually changed in 2026

If you last compared these plans even three months ago, two of the four have moved under you.

ChatGPT stopped counting your messages. On 6 August 2026 OpenAI announced that Free and Go users get unlimited text chats on GPT-5.6 Luna, its new default model for those tiers, with the rollout completing the week of 10 August. The cap it replaced was around ten messages per five-hour window, after which you were quietly moved to a lighter model. Free and Go users also got a "Think" button for pushing harder reasoning on a specific question. Limits remain on images, file uploads and voice, and OpenAI notes that abuse guardrails still apply — so "unlimited" means unlimited typing, not unlimited everything.

Gemini stopped counting prompts and started counting compute. Google's Gemini Apps limits now flex with how complex your prompt is, which model and features it touches, and how long the conversation has run. Your allowance refreshes every five hours until you hit a weekly ceiling. After a round of user complaints, Google capped how much of your quota a single Pro-model prompt can consume, and made Flash-Lite prompts free of quota entirely. This is arguably more honest than prompt counting — a one-line question genuinely costs less than a 40-page document analysis — but it is much harder to plan around, because you cannot see the meter.

Claude and Perplexity held their shape. Claude still rations tightly on a rolling five-hour window and still reserves its flagship model for paying users. Perplexity still gives away unlimited basic search and meters the premium modes. Both remain the most predictable plans of the four for exactly that reason.

The pattern worth noticing: the two companies with the largest free user bases are both moving away from simple, visible message counts. One went unlimited on the cheap thing; the other made the meter invisible. Neither change makes limits go away — it makes them harder to see, which shifts the cost from frustration to surprise.

The comparison table

Checked 27 August 2026. The "what happens at the cap" row is the one that most affects daily use, and the one comparison tables usually skip.

Feature ChatGPT Free Claude Free Gemini Free Perplexity Free
Default model GPT-5.6 Luna Sonnet 4.6 / Haiku 4.5 Flash tier, limited Pro access Perplexity's own search models
Text chat volume Unlimited (since 6 Aug 2026) Tightest — roughly 15–40 per 5 hours Compute budget, 5-hour refresh, weekly ceiling Unlimited basic search
What happens at the cap Only non-text features cap, and you are told Blocked until the window resets Moved down the model ladder Falls back to standard search
Context window Not published for free tier 200,000 tokens 32,000 tokens Per-query, not chat-based
Reasoning mode Yes — "Think" button Adaptive thinking toggle Within the compute budget Capped per day
Image generation Yes, small allowance No Up to 100 per day Peripheral
File uploads Yes, capped per day Up to 20 files, 500MB each Yes, but small context applied Yes, limited
Deep Research Effectively paid No dedicated mode Small monthly allowance A few runs per day
Workspaces Projects Projects + Artifacts Gems Spaces
Ads on free tier Yes, expanding by country No No No
Best at Everyday volume Long documents, code, writing Images, breadth, multimodal Sourced answers

ChatGPT Free — now the obvious default

What you get

Unlimited text conversation on GPT-5.6 Luna, plus web browsing, memory across chats, Projects, voice, a limited image allowance and capped file uploads. The August change also added a Think button, which lets you spend more reasoning on one question instead of hoping the default model gets there on its own.

Where it caps: everything that is not text. Image generation stays a small daily allowance and tightens under load. File uploads are capped per day, and that is the limit most people now hit — you can discuss a document endlessly, you just cannot feed it many new ones.

The catch: Luna is the smallest model in the GPT-5.6 family. Unlimited access to a lighter model is a real upgrade for casual use, but it is not the same as unlimited access to what paying users get. If your frustration with ChatGPT was answer quality rather than message count, this change does not fix your problem. Note too that ChatGPT's free tier now carries advertising in a growing list of countries, including on the paid $8/month Go tier.

Good fit: anyone who has never used an AI assistant, anyone doing high-volume routine work, students asking a lot of small questions. Poor fit: long document analysis, or work where the gap between a light model and a frontier model actually shows.

Claude Free — best answers, fewest of them

What you get

Sonnet 4.6 and Haiku 4.5, web search, file uploads (up to 20 files at 500MB each), Projects, Artifacts, and the adaptive thinking toggle. Crucially, free accounts get the same 200,000-token context window as standard paid plans, which is the most undervalued fact in this whole comparison.

Where it caps: volume, hard. Anthropic publishes no fixed number; most free users land somewhere around 15 to 40 messages per rolling five-hour window, varying with prompt length, attachment size, history and demand. The window starts on your first message, not at midnight. When you hit it you are blocked outright rather than downgraded — annoying in the moment, but at least honest.

Use it differently: long threads drain Claude free far faster than short ones, because the entire conversation is reprocessed on every reply. Starting a fresh chat per task, instead of continuing one endless thread, is the single change that most extends a free Claude day. That is the opposite of how most people use a chatbot, which is why the plan feels stingier than it is.

The free tier carries no ads. By default, conversations may be used for model training, which you can turn off in privacy settings — worth doing before you paste anything work-related. Claude Code, Cowork and Design are paid-only.

Good fit: long PDFs, codebases, serious writing, anything where you want one very good answer instead of ten adequate ones. Poor fit: casual all-day chatting, and image work, since it does not generate images at all.

Gemini Free — the most stuff, on an invisible meter

What you get

The widest feature spread of any free plan here: image generation at up to 100 per day, up to 20 audio overviews per day, a small monthly allowance of Deep Research reports, Gems, and deep integration with Google Search, Docs and Gmail if you already live in that ecosystem.

Where it caps: compute, not prompts. Your allowance drains according to prompt complexity, model, features used and chat length, refreshing every five hours until a weekly ceiling. Google capped how much quota one Pro-model prompt can consume and exempted Flash-Lite prompts from quota entirely after user pushback, both of which help. When the better model runs out you drop down the ladder rather than getting blocked.

The catch that costs people most: the 32,000-token context window on free accounts. That is roughly a sixth of what free Claude gives you. Upload a long report and Gemini will still answer confidently — it just may not have all of the document in view. This is the failure mode nobody notices, because no error appears when a summary quietly omits chapter four.

Good fit: image generation, multimodal work, anyone inside Google's ecosystem, and anyone who wants to try Deep Research without paying. Poor fit: long-document analysis, and anyone who wants a predictable, countable allowance.

Perplexity Free — the specialist that knows what it is

What you get

Essentially unlimited basic search with inline citations, search history, Spaces, and limited file uploads. Every answer arrives with clickable sources, which is the entire reason to use it — you can check the claim rather than trusting it.

Where it caps: the premium modes. Independent trackers put free Pro Search at roughly three to five per day, plus a small number of Research runs. Perplexity does not publish these prominently and reported figures disagree, so treat any specific count as an estimate rather than a promise. Free accounts also do not get manual model selection, the deeper financial and scientific sourcing, or the Comet assistant.

Honest scoping: Perplexity is not a Google replacement and is not trying to be. For "synthesise this from several sources and show me where each claim came from", it saves real time against opening eight tabs. For finding a specific site, a product page or a map, ordinary search is faster. Most regular users keep both and switch based on the question.

Good fit: research, fact-checking, anything you will need to cite. Poor fit: coding, long-form writing, casual conversation.

The row nobody puts in comparison tables

Look again at what each plan does at the cap, because it determines how a plan feels far more than the size of the allowance does.

Claude blocks you. That is the most frustrating in the moment and the least frustrating in retrospect, because you always know exactly what happened and exactly what your last answer was worth.

Gemini moves you down the model ladder. ChatGPT used to do this too, which is why "ChatGPT has got worse recently" became such a common complaint — people crossed a threshold nobody drew for them, kept typing into an identical interface, and concluded the product had degraded. If quality falls off a cliff mid-session on any tool that behaves this way, check the model indicator before you blame the product.

Perplexity falls back to standard search, which is graceful and clearly signposted.

ChatGPT, since August, mostly does not cap you on the thing you use most. That is a genuinely better experience, and it also removes the main signal people had that they were on a free plan at all.

How to get more out of free plans

Five habits that do more for a free-tier day than switching tools does:

1. Decide what a task needs before you type it. Reformatting a list does not need your best available model. The paragraph going into a job application does. Most free-tier frustration comes from spending scarce capacity on tasks that never needed it, then hitting a wall on the one that did.

2. Start new chats. On Claude especially, and on Gemini's compute budget, long threads cost more per message than short ones because the whole history is reprocessed. One task, one chat.

3. Front-load context. Paste the document, the constraints and the desired format in one message rather than dribbling them across six. Six clarifying rounds cost six messages; one well-built prompt costs one. Our guide on writing prompts that actually work goes deeper on this.

4. Match the tool to the job. Documents and code to Claude, volume and images to ChatGPT and Gemini, anything you will cite to Perplexity. Two apps cover this; four is friction without benefit.

5. Check the model indicator when quality drops. It takes two seconds and it explains most "this got worse" moments.

When free stops being enough

Free plans genuinely cover most personal use in 2026, more than they did a year ago. There are four honest reasons to pay:

You need agentic research repeatedly. Every free Deep Research allowance here is a taster, not a workflow. If this is your main use, it is the clearest reason to subscribe to something.

You need frontier-model quality daily. Free tiers give you the light or mid models. If the gap between a good answer and a very good answer has commercial consequences for you, that gap is what you are paying for.

You need long documents regularly. Claude free covers this surprisingly well, so try it before paying — but if you also need volume on top of length, free will not stretch.

Predictability matters. If getting blocked mid-task costs you a deadline, a subscription is buying certainty rather than capability.

Everyone else: two free accounts, used deliberately, will get you further than one subscription used carelessly. If you are curious how far free tools stretch across a whole workflow, we tracked that in ten hours a week, measured, and built a zero-cost study workflow in the free AI study stack.

Frequently Asked Questions

Is the ChatGPT free plan really unlimited now?

Unlimited on text, not unlimited overall. Since 6 August 2026, Free and Go users get unlimited text chats on GPT-5.6 Luna, replacing the old cap of roughly ten messages per five-hour window. Images, file uploads and voice still have their own caps, and abuse guardrails still apply.

How many messages does Claude free give you?

Anthropic publishes no fixed number. Most free users report roughly 15–40 messages per rolling five-hour window, varying with prompt length, attachments, history and demand. The window starts on your first message rather than resetting at midnight, and long threads drain it faster than short ones.

What models does Claude free include?

Sonnet 4.6 and Haiku 4.5, with web search, file uploads, Projects, Artifacts and adaptive thinking. Opus 4.8 is paid only, as are Claude Code, Cowork and Design. No ads. Conversations may be used for training unless you opt out in privacy settings.

How do Gemini's free limits work now?

By compute rather than prompt count — complexity, model, features and chat length all draw down an allowance that refreshes every five hours until a weekly ceiling. Google has since capped how much quota one Pro-model prompt can use and exempted Flash-Lite prompts entirely. Published free figures alongside that include up to 100 images per day, up to 20 audio overviews per day, a small monthly Deep Research allowance and a 32,000-token context window.

What do you get on Perplexity free?

Unlimited basic search with citations, history, Spaces and limited uploads. Pro Search is capped at roughly three to five per day by independent tracking, with a small Research allowance. No manual model selection, no deeper financial or scientific sourcing, no Comet assistant. Perplexity does not publish these counts prominently, so treat them as estimates.

Which free AI is best for coding?

Claude for the hard problem, ChatGPT for everything else. Claude handles multi-file reasoning and refactoring explanations best but gives you the fewest turns. Now that ChatGPT free is unlimited on text, it is the sensible default for routine debugging where the model gap barely matters. Save Claude for the thing you have been stuck on for an hour.

Which free AI gives the most image generation?

Gemini, comfortably — up to 100 images per day on free accounts. ChatGPT's free image allowance is small and tightens under load. Claude generates none. Perplexity's image features are peripheral.

Which free AI has the biggest context window?

Claude, at 200,000 tokens on free accounts — the same as standard paid plans, and roughly six times Gemini free's 32,000. That gap decides long-document work: the same 80-page PDF Claude reads in one piece has to be chunked for Gemini free, and chunking is where summaries quietly start missing things.

Do free plans include Deep Research?

Partly. Gemini free gets a small monthly allowance on its Flash model; Perplexity free gets a handful of runs; OpenAI treats its equivalent as paid. None is large enough to build a workflow on, which makes agentic research the clearest single reason to pay for something.

Can I use all four at once?

Yes, and it is the biggest structural advantage free users have — caps are per platform and not shared. The cost is friction: four apps, four histories, and the overhead of deciding where a question belongs. Most people settle on two within a few weeks.

Why does a free AI suddenly get worse mid-conversation?

Because some providers move you down the model ladder rather than blocking you. The interface is identical, no warning appears, and answers get quietly shallower. Check the model indicator before concluding the product has degraded.

Are there ads on free AI plans?

On ChatGPT, yes in a growing list of countries, including on the paid Go tier. Claude, Gemini and Perplexity free chat interfaces do not currently carry ads. It is worth knowing which of your tools monetises your attention and which monetises an upgrade, because that shapes incentives over time.

Verdict

Which one should you install?

There is no single winner, and after this year's changes that is more true than before. Gemini gives you the most features. Claude gives you the best answers and the longest context. ChatGPT now gives you the fewest reasons to think about limits at all. Perplexity gives you sources you can check.

Install ChatGPT first — unlimited text on a free plan is a real change, and it covers the majority of everyday questions. Add Claude second if you work with long documents or code, because its 200,000-token context on a free account is the most valuable underpublicised thing in this comparison. Swap in Perplexity instead if your work is research-heavy and citations matter more than conversation. Add Gemini if you generate images or already live in Google's tools.

Two accounts, chosen for what they ration, beats four installed at random. That was the real answer to my friend's question.

Sources and how to verify these numbers yourself

Free tier terms move faster than any article can. These are the sources behind the figures above, plus where to check them before you rely on them:

Disclosure: this article contains no affiliate links, and none of these companies had any input into it. Every plan discussed was used on a personal free account.

Update log

27 August 2026 — full rewrite. Rebuilt after OpenAI removed free-tier text limits on 6 August and Google switched Gemini Apps to compute-based limits. Updated Claude's free model lineup to Sonnet 4.6 and Haiku 4.5, added the context-window comparison, and added a sources section.

11 May 2026 — first published after several months of daily use across all four free plans.

Hit a limit this article does not describe, or seeing different numbers on your account? Tell me what you saw — regional and account-level variation is real, and reader reports are how this page stays accurate.