What You Actually Get from the Free Tiers in 2026: ChatGPT, Claude, Gemini, and Perplexity

My cousin installed four AI apps on a Sunday because someone on YouTube told him to. By Tuesday evening he messaged me, genuinely annoyed: "One of them stopped talking to me. One changed into a different model without telling me. One said I'd used up my research thing for the day. And the fourth is still working fine — why is the fourth still working fine?" That message is the entire reason this article exists. Every comparison I could point him to was either written in 2024, copied from a pricing page, or so hedged it said nothing. So I did the boring thing instead: I ran all four on their free plans, side by side, for months, doing real work. This is what I found.

First, a Warning About Numbers

I'm going to give you numbers in this article, and I want to be upfront about how much to trust them.

Free tier limits are the most volatile thing in this entire industry. They change when a new model launches, when servers are under load, when a company is trying to convert people, and sometimes for no announced reason at all. I have watched a limit change twice in one month. Any article — including this one — that quotes exact message counts as permanent fact is lying to you.

So here's the deal. Where I quote a figure, treat it as "roughly this, as of mid-2026, verify before you plan around it." What doesn't change quickly is the shape of each plan: what each company is willing to give away, what it guards, and how it behaves when you run out. That structural stuff is the genuinely useful part, and it's what I've weighted this article towards.

Why Every Company Gives You Something for Free

It's worth understanding the incentive, because it explains the limits better than any changelog.

Free plans are not charity and they're not really marketing either. They're habit formation. Every one of these companies has worked out that a person who uses their assistant daily for two months will find it painful to switch, not because the product is better but because their history, their saved projects, and their muscle memory all live there. The free tier's job is to get you to that point.

That's why the limits sit where they do. They're calibrated to be generous enough that you form the habit, and tight enough that a serious user feels the wall often enough to consider paying. When a limit feels weirdly, specifically annoying — like it cuts out right as you get into flow — that's not an accident. That's the product working as designed.

There's a second reason, less cynical: real usage is training data for product decisions, and free users generate an enormous amount of it. And a third, newer one — OpenAI began testing advertising on its free tier in early 2026, which means free users are now a revenue stream rather than purely a cost. Expect that model to spread.

None of this is a reason to avoid free plans. It's a reason to use them deliberately.

The Quick Comparison

If you read nothing else, read this table. The "Overall Value" column is my opinion after months of use, not a spec.

Look at that "what happens at the cap" row again. It's the row nobody puts in comparison tables and it's the one that determines how the plan actually feels to use.

ChatGPT Free: The Comfortable Default

ChatGPT free is the plan I'd hand to someone who has never used an AI assistant, and that's a real compliment rather than a backhanded one.

What you get is broad. OpenAI currently puts its mainstream model in front of free users — not the reasoning-heavy variant, but the everyday one — along with web browsing, image generation, file uploads, voice, memory across conversations, and Projects for organising related chats. Almost nothing is entirely missing. Most things are just smaller.

The message cap is the thing you'll meet first. Free accounts get a limited run on the good model within a rolling window of a few hours. The number moves around and OpenAI doesn't publish it prominently, which is itself informative. Recent reporting puts it in the low tens of messages per window, sometimes lower under load.

Here's what matters more than the number: you don't get blocked, you get downgraded. The interface doesn't change. There's no modal, no countdown, nothing obvious. Your answers just start getting thinner. I have watched people conclude that ChatGPT "has got worse recently" when what actually happened is they crossed a line nobody drew for them. If quality falls off a cliff mid-session, check the model selector before you blame the product.

Image generation exists on free but the allowance is small and shrinks when servers are busy. Fine for a birthday card. Not a workflow.

File uploads are capped per day at a number low enough that a student with a stack of lecture PDFs will hit it. This is the second wall people meet.

Deep Research — the mode that goes off and reads dozens of sources before writing a cited report — is effectively a paid feature. Free users get, at most, a token amount. If research is your main use, this alone pushes you towards Gemini or Perplexity.

The ads. Since early 2026 OpenAI has been testing advertising for free and low-tier users, starting in the US and expanding to other markets. Ads are labelled and separated from the answer, and OpenAI says they don't influence responses. There's also reportedly an option to opt out in exchange for fewer daily messages. Whether this bothers you is personal; I mention it because no pricing page leads with it.

Where it's genuinely strong: being decent at everything. It reads PDFs well, follows formatting instructions reliably, explains things at whatever level you ask, and doesn't need coaxing. When I want an answer rather than a project, this is what I open.

Where it lets you down: it is confidently wrong with the same tone it uses when it's right, and it drifts towards bland textbook phrasing unless you push back. And free users don't reach the reasoning models, so genuinely hard problems — a tricky algorithm, a multi-step proof — get an answer that looks right and isn't.

Who should stay free: almost everyone who uses it a few times a day. Who should pay: anyone who does research reports, generates images regularly, or hits the downgrade more than twice a week.

Claude Free: Best Answers, Fewest of Them

Claude free is the plan I have the most complicated relationship with, because it's simultaneously the best and the most frustrating.

Anthropic gives free users its mid-tier model. The genuinely capable top model is paid, full stop. But the mid-tier model is very good, and on writing tasks it is — in my honest opinion after a lot of side-by-side testing — the best of the four by a clear margin. It writes prose that sounds like a person wrote it. The others write prose that sounds like an assistant produced it.

You also get Projects, Artifacts, web search, memory across conversations, file uploads, code execution, and extended thinking on a limited basis. On paper that's a strong free plan.

Then you hit the cap, and it is tight. Tightest of the four, comfortably. And unlike ChatGPT, Claude doesn't downgrade you — it stops. You get a message telling you when your limit resets, usually a few hours out, and until then that's that.

Honestly? I've come to prefer being blocked over being downgraded. It's ruder and much more useful. You know exactly where you stand instead of quietly getting worse answers for an hour.

The thing nobody explains about Claude's limits: they're driven by how much text is being processed, not how many times you press enter. Every reply reprocesses the whole conversation. So a fiftieth message in a long thread costs dramatically more of your allowance than a first message in a fresh one. Two people can send the same number of messages and one runs out three times faster.

Which gives you the single most valuable free-tier trick in this article: start a new conversation when you change topic. Not for tidiness — for allowance. Dragging a 40-message thread about your CV into a question about Python is paying a tax for context you don't need. I roughly doubled my usable Claude free time by doing nothing except opening new chats more often.

Artifacts deserve a mention. When Claude writes code or a document, it appears in a side panel you can iterate on rather than being buried in chat. For drafting something you'll revise several times, it's the nicest interface of the four.

Where it's strong: writing, editing, long documents, and honest feedback. When I paste my own paragraph and ask what's unclear, Claude tells me. The others tell me it's great. That is not a small difference if you write for a living.

Where it lets you down: capacity, and only capacity. It hedges more than some people like, though I'd argue accurate hedging beats confident nonsense.

Who should stay free: writers who use it for the hard paragraph rather than the whole draft. Who should pay: anyone who hits the wall daily — and if you're using it for real work, you will.

Gemini Free: The Most Stuff for Nothing

If the question is purely "which free tier gives me the most," Gemini currently wins and it isn't especially close.

Google restructured its limits in 2026, moving from flat daily prompt counts to a compute-based allowance that refreshes every few hours under a weekly ceiling. In practice: free users get a real amount of usage, can reach the higher model tiers within that allowance rather than being locked out of them, and get features the others charge for.

Deep Research is available to free users. That's the headline. Google's own documentation notes it may be unavailable during heavy demand, which is a fair trade for something the competition treats as a paid feature. Image generation is also generous — free users get a meaningful daily count with Google's current image model, far more than ChatGPT free. Canvas, Google's editing workspace, is fully available. Video generation is not free.

The ecosystem is the real argument. Gemini connects to Gmail, Drive, and Docs. "Find the invoice from that vendor in March and summarise what we agreed" is a question only Gemini can answer, and the first time it works you understand why Google will be fine in this market regardless of who has the best model. NotebookLM, separately free and excellent, pairs naturally with it.

Now the catch, and it's a good one. Google's own support documentation states that free-tier users get a 32k-token context window for uploaded files, against 128k for the entry paid tier and 1M for higher ones. Read that again, because it's the most consequential free-tier limit nobody discusses. Every article about Gemini mentions million-token context windows. Free users get a small fraction of it for their documents. So you upload a 200-page report, ask a question about chapter nine, and get a confident answer drawn from material that got truncated. It doesn't tell you it truncated. You just get a plausible wrong answer.

If you take one practical thing from this entire article, take that.

The workaround is free too: Google AI Studio. Same underlying models, no subscription, and long-context work actually works there. It's a developer-facing tool and it looks like one — no memory, clumsier interface, no polish. But for one enormous PDF, it's the most capable free option that exists across all four platforms. I keep it bookmarked purely for that.

Where it's strong: volume, breadth, images, research, and anything touching your Google account.

Where it lets you down: consistency. Ask the same question twice and the two answers can differ in quality more than they do on ChatGPT or Claude. For coding it's capable but I trust it least of the three for anything I won't test.

Who should stay free: most people, genuinely. Who should pay: people doing serious document work, where that context limit is the whole reason to upgrade.

Perplexity Free: The Specialist

Perplexity is the odd one out, and comparing it to the other three by message count misses what it's for.

It's a search engine that answers you. You ask, it retrieves live sources, writes a synthesised answer, and cites each claim with a clickable link. Basic searches are unlimited on free — genuinely unlimited, which none of the others offer for anything meaningful.

What's capped is the good stuff. Free users get a small number of Pro searches per day, where it does a more thorough retrieval pass, and a small number of Deep Research runs. Both are in the handful-per-day range and both move around. When you're out, it falls back to standard search, which is still useful — a softer landing than the others.

Why the citations change how you work. With the general assistants, verifying a claim means going to Google and starting over. With Perplexity the source is right there. I don't think Perplexity's underlying model is more truthful than the others. I think it's the only one where checking is fast enough that I actually do it, and that's the difference that matters in practice.

It's not immune to error — it will occasionally cite a source that doesn't support the sentence attached to it, so click through on anything important.

Why people pair it rather than switch to it. Perplexity is bad at the things chat assistants are good at. Ask it to draft an email in your voice, restructure an argument, or work through code with you and it's mediocre, because it's built to retrieve and summarise rather than to think alongside you. Nobody I know uses Perplexity instead of ChatGPT. Plenty use it with ChatGPT. That's the natural shape of the product, not a criticism.

Where it's strong: current information, fact-checking, product research, academic sources, anything where "where did that come from?" matters.

Where it lets you down: creative work, long-form writing, sustained coding, and anything needing memory of what you discussed last week.

A Real Day: Which One Survives?

Specs are abstract. So I spent a day running one realistic student workload through all four, using each as if it were the only tool I had. Same tasks, same order, morning start.

Task one, summarise a 40-page lecture PDF. All four handled it. Claude gave the best summary. Gemini gave a decent one but I couldn't tell how much of the document it had actually read, which — knowing the context limit — bothers me more than the answer's quality suggests it should. Perplexity was the weakest here; it's not built for your files.

Task two, turn that into revision notes. Claude and ChatGPT both good. Gemini fine. First real cost of the day: Claude was now maybe a fifth into its allowance, two tasks in, because both involved a lot of text.

Task three, debug a Python script that wouldn't run. Claude found it fastest and explained why. ChatGPT got there in two exchanges. Gemini fixed it but also rewrote three things I hadn't asked about. Perplexity was not useful, as expected.

Task four, write an email to a professor. Trivial for everyone. But notice: I'd now spent four premium interactions on something a downgraded model could have handled. That's the mistake almost everyone makes on free plans — spending good-model capacity on tasks that don't need it.

Task five, research a phone before buying. Perplexity won outright — current prices, real reviews, citations. Gemini's Deep Research produced a longer and more thorough report and took several minutes. ChatGPT free was the weakest here, which is entirely down to research being a paid feature.

Task six, compare three laptops in a table. Perplexity again, for the same reason.

Task seven, build a revision timetable. All fine. Gemini's went straight into my Calendar, which none of the others could do.

Task eight, analyse a 90-page PDF. Here it separated. Claude handled it properly but this one task ate a serious chunk of what was left. ChatGPT managed but I'd nearly used my daily uploads. Gemini answered confidently and I have real doubts about how much it read. Google AI Studio, which I'd been ignoring, handled it best of anything — free.

Task nine, generate images for a presentation. Gemini, comfortably. ChatGPT gave me two before slowing down. Claude can't. Perplexity, no.

Task ten, practice interview questions. Around 4pm. Claude had been out for over an hour. ChatGPT was on the smaller model and I could feel it — answers technically responsive, noticeably shallower. Gemini was still going. Perplexity was still going but was never the right tool.

The result: Claude ran out first, roughly two-thirds of the way through, and it wasn't close. ChatGPT never stopped but stopped being good, which is a different problem and arguably a worse one. Gemini lasted the whole day. Perplexity never ran out of what I mostly used it for.

If that day were my normal, the honest answer is: Gemini as the daily driver, Perplexity for research, Claude saved deliberately for the two or three things that most needed quality. Which is not a conclusion I expected when I started, because Claude gives me the best individual answers of any of them.

The Limits Nobody Puts on the Pricing Page

Message caps are the advertised limit. These are the ones that actually cost you.

Silent model switching. Covered above, worth repeating. On several platforms hitting the cap downgrades rather than stops you, and the interface doesn't shout about it. This single behaviour causes more "this AI got worse" complaints than anything else.

Context truncation without notice. The Gemini 32k free-tier file context is the clearest example, but all of them do a version of this. When a document exceeds what the plan will process, you don't get an error. You get an answer based on part of it. There's no visual difference between a complete answer and a confidently truncated one.

Long conversations burn allowance faster. Because context is reprocessed each turn, a long thread costs multiples of a short one per message. Most people never learn this and wonder why their limits feel inconsistent.

Peak-hour throttling. Free tiers are the first thing cut when infrastructure is strained. Your allowance can be effectively smaller at 3pm on a weekday than at 6am on a Sunday, and nothing tells you.

Rolling windows, not midnight resets. Most of these reset a few hours after you started, not at a fixed daily time. So "I'll wait until tomorrow" is often wrong — you might be fine in three hours.

Staggered feature rollouts. A feature announced in a blog post may take weeks to reach your account, and free tiers are last in the queue. If you can't find something you read about, it may simply not have arrived.

Regional differences. Availability, features, and ads vary by country. Comparisons written for a US audience routinely describe a product that doesn't match what someone in India or the EU sees.

Pricing pages don't mention any of this, and I don't think that's malicious. It's that "up to 32k context on uploaded files, sometimes, depending on load" doesn't fit in a marketing bullet. But it's what you're actually buying, or in this case not buying.

Which Free AI Is Best For You?

Students. Gemini as the base, because volume matters more than peak quality when you're grinding through material, and free Deep Research is genuinely useful for assignments. Add NotebookLM for course material — it answers only from what you upload and cites the page, which makes verification a two-second click. Keep Claude for the concepts that won't click. I've written up the full free study stack I used across two semesters if you want the workflow rather than the comparison.

Programmers. ChatGPT daily, Claude for the hard one. Most coding questions are routine and don't need the best model; you need availability. Save Claude's scarce turns for the bug you've been stuck on. Neither free tier is enough for a full working day, and if you code professionally, one subscription is the easiest ROI on this list. For interview practice specifically, the way you prompt matters more than which model you're on.

Writers. Claude, and it's not close. It's the only one whose default prose doesn't read as machine-written, and the only one that gives useful critical feedback rather than praise. The free cap hurts, so use it for the paragraph you can't crack rather than the whole draft.

Researchers. Perplexity plus Gemini. Perplexity for targeted questions with sources; Gemini's Deep Research for the broader sweep. Both free. This combination replaces a surprising amount of manual reading.

Teachers. Gemini, for Workspace integration and image generation for materials. If you're generating worksheets and slides, the free image allowance alone justifies it.

Job seekers. Claude for the CV and cover letters — the writing quality difference is most visible exactly here — and Perplexity to research the company. Keep the CV work in one Project so you're not re-explaining your background every time.

Content creators. Gemini for volume and images, Claude for anything with your name on it. The economics of free tiers suit this well: bulk work where quality is negotiable, careful work where it isn't.

Business owners. Whichever ecosystem you're already in. If your company runs on Google Workspace, Gemini's integration beats a marginally better model elsewhere. Do check your provider's data-handling terms before putting client information into any free tier — free plans generally have weaker data protections than paid ones, and that's a genuine business consideration rather than a technicality.

If You Only Install One

Gemini. That's my honest recommendation for most people, and it's not because I think it's the best AI — I don't. Claude gives better answers and ChatGPT is the smoother experience.

It's because a free plan's value is capability multiplied by how often you can use it, and Gemini currently wins on the second term by enough to overcome losing on the first. A very good answer you can't get because you're capped out is worth nothing.

If you install two, though, the combinations matter more than the individual picks:

ChatGPT + Perplexity — the best general-purpose pairing. One assistant that does everything adequately, one specialist that handles current facts with sources. Almost no overlap, which is the point. If you're not sure, pick this.

Claude + Gemini — quality plus quantity. Gemini absorbs the routine volume; Claude handles what needs to be good. This is closest to what I actually use, and the discipline it requires — deciding which tier a task belongs to before you type — is a skill worth having anyway.

Gemini + ChatGPT — the highest-volume combination. Two generalists, roughly double the allowance, no specialist. Good if you use AI constantly for undemanding work and just need it to keep answering.

What I'd avoid is installing all four and using them randomly. That's my cousin's situation, and the result isn't four times the capability — it's four half-remembered histories and no idea where you left anything.

Getting More Out of Free Plans

• New chat per topic. The single highest-impact habit. Long threads burn allowance on context you don't need.
• Match the task to the tier. Don't spend a premium turn reformatting a list. Do the cheap work on whichever tool is downgraded anyway.
• Batch your uploads. Where daily upload counts are the limit, one combined PDF beats five separate ones.
• Do heavy work off-peak. Free tiers are throttled first when load is high. Early mornings are noticeably better.
• Remember Google AI Studio. Free, and the best long-document option anywhere on this list.
• Front-load before deadlines. Limits bite hardest the night before something is due. Generate on Tuesday, not at 11pm on Thursday.
• Check the model before blaming the tool. Sudden quality drops are usually a downgrade, not a regression.
• Buy a month, not a year. If you need a subscription for a specific project, pay for that month and cancel. These are monthly products; use them that way.
• Look for student pricing. Several providers discount heavily or run free periods for verified students. Check before you pay full price.
• Write better prompts before you buy more capacity. A lot of what people mistake for a model limitation is a vague request. Sharpening the prompt is free and usually helps more than upgrading.
• Consider a small local model for the routine stuff. A model running on an ordinary laptop with no graphics card has no caps at all, which makes it a decent overflow valve for simple questions.

Frequently Asked Questions

Is the ChatGPT free plan enough for most people?

For casual use, genuinely yes. A handful of questions a day, the occasional email, something explained — you'll rarely meet the ceiling. It becomes frustrating with sustained work. OpenAI currently gives free accounts a limited run on its main model within a rolling window of a few hours, then moves you to a smaller one. Nobody warns you when this happens mid-task, and the quality drop is what people actually complain about, rather than the cap itself.

Does Claude have daily limits on the free plan?

Anthropic uses rolling windows rather than a midnight reset, so your allowance refreshes a few hours after you start rather than the next day. Claude free is the tightest of the four in raw volume, and long conversations drain it faster than short ones because the whole thread is reprocessed on every reply. It also throttles harder at busy times. The practical upshot: Claude free rewards short, deliberate sessions and punishes rambling ones — which is the opposite of how most people use a chatbot.

Can Gemini replace ChatGPT for free users?

For volume, yes. Google is currently the most generous of the four, and the only one putting a research feature and a real image generator in front of free users without immediately gating them. Where it doesn't replace ChatGPT is consistency — Gemini's answers vary more between attempts, and its free tier applies a much smaller context window to uploaded files, which quietly limits long document work. Many people use Gemini for volume and keep a second assistant for answers that matter.

Is Perplexity better than Google Search?

Better at one kind of question, worse at everything else. When you want a synthesised answer from several sources with clickable citations, it saves real time against opening eight tabs. When you want a specific site, a product page, a map, or anything transactional, ordinary search is faster — just use it. Perplexity is a research layer, not a search replacement, which is why most regular users keep both and switch based on the question.

Which free AI is best for coding?

Claude on quality per message, with the large caveat that you get the fewest messages. Its free model handles multi-file reasoning and refactoring explanations better than the others, and Artifacts make iterating on a script pleasant. ChatGPT is the better everyday choice because you can ask more before hitting a wall, and for a single error message the model gap barely matters. The pattern that works: ChatGPT for the routine questions, Claude for the one you've been stuck on for an hour.

Which AI gives free image generation?

Gemini and ChatGPT both do, and Gemini is currently more generous by a wide margin. Google gives free users a meaningful daily count with its current image model; OpenAI's free allowance is small and tightens under load. Claude doesn't generate images at all, and Perplexity's image features are peripheral. If images are your main use case, start with Gemini.

Which free AI has the largest context window?

Claude offers the largest usable conversational context among the free chat apps. Gemini's underlying models support far more, but the app applies a much smaller window to uploaded files on free — which surprises people who read about million-token contexts and then watch a long PDF get truncated. The real loophole is Google AI Studio: free, exposes long-context models directly, clumsier than a chat app, and the most capable free option for one very large document.

Should students pay for an AI subscription?

Most shouldn't, at least not permanently. Free tiers cover the overwhelming majority of student work, and two free accounts across two providers roughly doubles your effective allowance for nothing. Paying makes sense in bursts — one month during a dissertation or placement season is cheap and genuinely useful; twelve months for occasional homework is capacity you aren't using. Check for student pricing first, since several providers discount heavily or run free periods for verified students.

Which AI hallucinates least?

Perplexity, structurally rather than because its model is more truthful. Retrieving sources first and citing them beside each claim means you can check in seconds rather than trusting. It's not immune — it will sometimes cite a source that doesn't support the attached sentence. Among the general assistants, Claude is most willing to admit uncertainty, which some find frustrating and which is genuinely safer. The honest answer: all four invent things confidently, and the real difference is how fast each lets you catch it.

Can I use all four AI tools together for free?

Yes, and it's the biggest structural advantage free users have over someone subscribed to one platform. Nothing prevents four free accounts, and caps are per platform rather than shared, so running out on one costs nothing on the others. The real cost is friction — four apps, four histories, and the overhead of deciding where a question belongs. Most people settle on two after a few weeks: a main assistant and one specialist.

Do free AI plans have Deep Research?

Partly. Google currently makes Deep Research available to free Gemini users, noting it may be unavailable during heavy demand. Perplexity gives free users a small number of runs per day. OpenAI treats its equivalent as paid, with free access minimal at best. These allocations move frequently, so treat any specific count as a snapshot and check the provider's page before planning work around it.

Why do free AI plans get worse after a few messages?

Because most providers don't block you at the cap — they silently downgrade you to a cheaper model. You keep chatting, the interface looks identical, and answers quietly get shallower. This is the most common source of "this AI has got worse" complaints, when what happened is you crossed a threshold nobody displayed. If quality drops suddenly mid-session, check which model you're on before concluding the product has degraded, and consider resuming in a few hours when the window resets.

Final Verdict

I started this expecting to crown a winner. I didn't get one, and I've stopped thinking that was the right question.

Gemini gives you the most. Claude gives you the best. ChatGPT gives you the fewest reasons to think about it. Perplexity gives you sources. Those are four different products that happen to share an interface, and picking one "best" is like asking whether a notebook beats a calculator.

What genuinely changed how I work wasn't choosing correctly. It was noticing that free tiers have a shape, and that the shape rewards a particular habit: decide what a task needs before you type it. Routine reformatting doesn't need the best model available. The paragraph going in your job application does. Almost all the frustration people report with free AI comes from spending scarce capacity on tasks that never needed it, and then hitting a wall on the one that did.

So my cousin's fourth app was still working because it was Gemini, and Gemini is currently the most generous. But the better answer to his question was the one he didn't ask: he didn't need four apps. He needed two, and a rough sense of which one a question belonged to.

One last piece of honesty. Everything specific in this article has a shelf life. The limits will move, probably within months of publication. The structure will hold longer: Google buying market share with generosity, Anthropic guarding its best model, OpenAI optimising for the largest number of people who never notice a limit, Perplexity doing one thing properly. When the numbers change, that framing will still tell you roughly where you stand.

Free plans don't ration capability. They ration attention — yours and the model's. The people who get the most out of them aren't the ones who found the best tool; they're the ones who stopped spending their best model on their easiest questions.