Best AI assistant: the free tiers decide it, not the $20 ones

Contents

The short answer

ChatGPT if you are going to pay, on breadth rather than on winning any single category. Gemini or Grok if you are not, because both hand a free account a model picker where ChatGPT and Claude hand it nothing. And Claude if your output is prose somebody reads.

Assistant	Alley Rating	Free tier gives you	Entry paid
ChatGPT	4.6, Category Leader	No model choice	$20
Claude	4.4, Power Tool	No model choice	$20, or $17 annual
Perplexity	4.4, Power Tool	Ten models, all padlocked	$20
Google Gemini	4.3, Power Tool	Four selectable modes	Via Google One
Grok	Not rated, free tier tested	Five selectable modes	$30, or $300 a year

The ranking above is the usual one. The third column is not, and it is the more interesting half of this comparison.

The thing every ranking gets backwards

Rankings for this term compare the $20 tiers, because that is where the feature tables live and where the affiliate money is.

The tests I ran this week were free-versus-free, and they point somewhere unexpected. Running the same prompts through ChatGPT’s free tier and Grok’s free tier, both answered two factual questions correctly, including one most of page one of Google still gets wrong, and both followed four explicit writing constraints without breaking one.

I have not used Grok’s paid product at all, so that is a claim about the bottom of the ladder rather than the top. What it suggests is that the output gap is already small before anyone pays. Above that line, on the evidence of the four reviews behind this post, the $20 tiers are close too, and the differences that survive are mostly matters of taste and of which ecosystem you already live in.

The free tiers are where they diverge, and the divergence runs opposite to the reputations.

Free tier	Model choice	What you actually get
Grok	Five modes	Fast, Build (beta), Auto, Expert, Heavy
Gemini	Four modes	Including advanced reasoning, none padlocked
ChatGPT	None	One Think toggle
Claude	None	The assistant, no picker
Perplexity	None usable	Ten models displayed, a padlock on each

I opened four of these free accounts to check rather than reading it off a comparison page. The exception is Claude: I hold a paid account, so its free-tier row comes from Anthropic’s own plan table, which does not list the model picker below Pro. The two most famous assistants give a non-paying user the least choice, and the two most often described as challengers give the most. Perplexity’s is the strangest of the five: ten models are visible and every one of them is locked, which is a showroom rather than a free tier, and our Perplexity review says so at more length.

Two of those rows need a caveat before you act on them. A mode appearing in a picker is not the same as unmetered access to it: Grok’s own upgrade screen sells smarter answers in Expert mode as a paid benefit, which implies the free version is capped in a way the interface does not spell out. And Gemini’s four modes are genuinely selectable, but the thing most people want from Gemini — reaching Gmail, Docs and Drive — is not on the free tier at all.

There is also a difference in how the two restricted products restrict you. ChatGPT’s free tier gives you one good model and no choice, which is a coherent product. Perplexity’s shows you ten and locks all ten, which is a shop window. Those are not the same experience even though both reduce to “no model choice” in a table.

This matters more than a feature-table row because the free tier is where almost everyone forms their opinion. If you tried ChatGPT free and Gemini free and concluded ChatGPT was better, you were comparing a product with no model choice against one with four, and the conclusion you reached may have been about the interface rather than the intelligence.

How this was tested

Four of the five carry an Alley Rating earned on a real account over an extended period, and those reviews hold the per-tool detail: ChatGPT at 4.6, Claude at 4.4, Perplexity at 4.4 and Gemini at 4.3. This post adds one thing they do not: a same-day comparison of what each free tier hands a new user.

What	Detail
Free tiers opened	Four of five, one sitting, 27 August 2026
Claude free tier	Not opened — documented from Anthropic’s plan table
Grok free, prompts run	Four, once each, 27 August
Also run on ChatGPT free	All four of them, same day
Also run on Claude	Two of the four, the pricing and writing prompts
Paid tiers used	ChatGPT, Claude, Gemini, Perplexity. Not Grok
Agent tools tested	None

Two limits worth stating rather than burying. Each cross-tool prompt ran once, so those results are an anecdote with screenshots rather than a benchmark, and I would not extrapolate from them to a general claim about which model reasons better. And “not rated” for Grok means exactly that: I have seen its free tier across four prompts and nothing of its paid product, so it sits in the ranking on a narrower basis than the other four.

The five AI assistants, axis by axis

Axis	ChatGPT	Claude	Gemini	Perplexity	Grok
Alley Rating	4.6	4.4	4.3	4.4	Not rated
Free model choice	None	None	Four modes	Ten, all locked	Five modes
Entry paid	$20	$20	Google One	$20	$30
Annual discount	None offered	$17/mo	Via Google One	Annual-equivalent shown	$300/yr
Top consumer tier	$200	$200	Google One tiers	$200	$30
Generates images	Yes	No	Yes	No	Yes
Cites sources by default	On retrieval	On retrieval	On retrieval	Always	Always
Best single strength	Breadth	Long prose	Free capability	Checkable answers	Free model range
Weakest thing	Restricted free tier	No images	Tiered adjacency	Locked free tier	No paid testing here

Two rows carry most of the decision. Free model choice is where the products actually diverge, and best single strength is what you are really buying once you have decided to pay.

Which of these I have actually used

The cross-tool evidence in this post comes from two write-ups published this week: ChatGPT vs Grok, which ran four prompts through both free tiers, and Claude vs Grok, which found a real difference in which sources each one trusted.

What I did not test is the other half of this category, and that needs saying plainly rather than in a footnote.

Assistants and agents are not the same product

Search for the best AI assistant and you will get two different kinds of software mixed into one ranking.

The five tools here are assistants. You ask a question, they answer, you take the output and do something with it. The loop closes with you.

Tools like Motion, Reclaim, Lindy and Notion’s agent features are agents. They connect to your calendar, inbox or CRM and change things without you approving each step. The loop closes without you.

Every page-one competitor for this term ranks both groups together. One of them ranks fifteen tools that include ChatGPT and Gemini alongside a scheduling tool, a meeting transcriber and two voice assistants, scored against criteria that cannot apply evenly to all of them. A calendar tool cannot be judged on writing quality and a chat assistant cannot be judged on how well it blocks focus time.

We have not tested the agent tools. Testing them properly means connecting a real inbox and a real calendar to software and letting it act, which is a different kind of commitment from opening a chat window, and it is not something to fake with a screenshot of a marketing page.

So this post ranks the assistants and stops there. If what you actually want is something that clears your inbox overnight, the tools above are the wrong shortlist — but you can leave equipped rather than empty-handed.

	Assistant	Agent
Closes the loop	You do	It does
Needs write access	No	Yes, to inbox or calendar
Cost of a bad output	A wasted minute	An email that went out
Judge it on	Answer quality	What it does unsupervised
In this ranking	All five	None

What I can offer on that half is the structural difference to check for, because it decides more than any feature list. Ask whether the tool needs write access to something you care about. An assistant needs nothing: you paste in, you copy out, and the blast radius of a bad answer is a wasted minute. An agent needs your calendar, your inbox or your CRM, and the blast radius of a bad decision is an email that went out or a meeting that moved.

Four questions are worth asking any agent vendor, and none of them apply to a chat assistant:

1. What does it do without asking? The whole value is unsupervised action, so the scope of that autonomy is the product.
2. What can it undo? A sent email and a moved meeting have different reversibility, and vendors rarely lead with this.
3. What access does it need on day one? Read-only calendar access and full inbox write are very different commitments.
4. What happens when it is wrong? Not whether it errs, which it will, but what the recovery looks like.

For orientation rather than ranking, these are the names that recur in other guides:

Tool	Acts on	Access it needs	Tested here
Motion	Calendar, tasks	Calendar write	No
Reclaim	Calendar	Calendar write	No
Lindy	Email, CRM	Inbox and CRM write	No
Notion Agent	A workspace	Workspace write	No
alfred_	Inbox	Inbox write	No

I have used none of them, so treat that as a map of the territory rather than a recommendation. The last column is the one that matters: every row is a product this post did not test.

That difference is why the two categories cannot honestly share a ranking. Those four questions are not ones a chat assistant needs to answer at all, and any list that scores both groups on one set of criteria has quietly decided they do not matter.

1. ChatGPT — best overall, on breadth

C

ChatGPT at a glance
Alley Rating	4.6 — Category Leader, reviewed here
Free tier	Yes, no model picker
Paid	Plus $20, Pro $100 or $200, monthly only
Best for	Wanting one subscription to cover everything

It does not win a single category outright in this list. It writes less well than Claude, cites less cleanly than Perplexity, and gives a free account less choice than Gemini or Grok. It is still the one to buy for most people, because it covers more ground in one subscription than anything else here.

Images, voice, a wide integration surface and a paid ladder that keeps going past $20 all count for more than a category win when you are choosing one tool rather than assembling a stack. Our ChatGPT Plus vs Pro guide covers that ladder, including a $100 Pro tier most comparisons still miss.

Two things to know before you commit. There is no annual billing on any individual ChatGPT plan, so a year costs exactly twelve times the monthly rate with no discount to reach for. And the free tier is the most restricted of the five on model choice, which makes it a poor trial for what the paid product does.

The ladder is worth understanding before you buy into it, because it is stranger than it looks. Pro is not one price but two, $100 and $200, and OpenAI’s own help centre puts it this way: “Pro $100 unlocks 5x higher usage than Plus, while Pro $200 unlocks 20x usage than Plus.” Work that per unit and the middle tier is not a discount at all — it costs the same per unit of usage as Plus does, and only the $200 tier improves the rate.

There is a second oddity in the specification table. Go and Plus publish identical context windows on every row OpenAI lists: 54K on the instant model, 256K on reasoning, roughly forty and three hundred and twenty pages of input respectively. Paying to move from Go to Plus buys models and features, not room. Pro is where the number finally moves, to 128K and 400K.

The Go tier has one detail worth knowing before it tempts you: OpenAI’s own plan card says it may include ads, which makes it the only rung where you pay money and still see them.

For the head-to-heads it keeps losing and winning, ChatGPT vs Claude covers the $20 tier most people are actually deciding between, and ChatGPT vs Gemini covers the free tiers where the gap is widest.

2. Claude — best for prose somebody reads

C

Claude at a glance
Alley Rating	4.4 — Power Tool, reviewed here
Free tier	Yes, no model picker
Paid	Pro $20 or $17 annual, Max $100 or $200
Best for	Long writing held to rules

This is the clearest single-axis result across everything this site has tested. On prose that has to hold a style guide across length, Claude wins, and it is not close.

The reason so many rankings miss it is that the difference does not appear in a quick trial. Give two assistants a set of prohibitions and a long brief, then read the last third of each draft rather than the first. Claude still respects the rules there; the others drift back toward their house structure once the material gets complicated. A three-prompt comparison will never show you this.

What holds the rating at 4.4 is metering rather than output. The limits are real and the naming is confusing, and our Claude Max plan guide works through why the multiplier in the plan name describes a five-hour session rather than a month, which is behind a lot of surprise bills.

The metering deserves a paragraph of its own because it is the most misunderstood thing in this category. Anthropic’s Max tiers are named 5x and 20x, and the natural reading is five or twenty times as much use per month. The actual wording describes a five-hour session window, and there is a separate weekly cap sitting above it that applies across every model. Heavy users tend to meet the weekly ceiling rather than the session one, which is precisely the limit the marketing number says nothing about.

Two hard limits beyond that. Claude generates no images at all, so if your assistant is doing double duty that is a second subscription rather than a feature gap. And Anthropic now sells five separately named products under the Claude brand — Code, Cowork, Design, Science, and Security on Enterprise — which makes shopping for a replacement unexpectedly confusing. Our Claude alternatives post untangles which one you actually mean before you start comparing.

One genuine advantage in the pricing: Claude Pro is the only plan in this list that discounts for paying annually at the entry tier, at $17 a month against $20. ChatGPT offers no annual billing on any individual plan at all.

3. Google Gemini — best free tier for reasoning

G

Gemini at a glance
Alley Rating	4.3 — Power Tool, reviewed here
Free tier	Four selectable modes, none padlocked
Paid	Via Google One AI plans
Best for	Getting real capability without paying

The free tier is the argument, and it is a strong one. Four selectable modes including advanced reasoning, none of them locked, where ChatGPT and Claude give a free account no choice at all.

The second argument is adjacency. Gemini reaches Gmail, Docs, Drive and Calendar natively, and no competitor can match that because the obstacle is ownership rather than engineering. The catch is that this advantage is tiered: a free account reaches none of your Google apps, so the thing people most want from Gemini sits behind a Google One plan.

The caveat we documented with a screenshot is retention. Google’s own support page states chats are kept for 72 hours even with history switched off. That is disclosed rather than hidden, and it is not nothing if you paste client material. Our Gemini alternatives roundup covers who that should bother.

Rated 4.3 rather than higher largely on packaging: the paid tiers are storage plans with an assistant attached rather than assistant subscriptions, which is a strange way to buy one and makes it oddly hard to answer “what does Gemini cost” without a paragraph of caveats.

There is one more thing Gemini does that nothing else in this list does, and it rarely appears in assistant rankings because it sits outside the category. It generates images on the free tier with unusual generosity, and it reaches Google’s Veo models for video. If you were using one subscription to cover writing, images and video, moving to a chat-only assistant leaves you short in a way a feature table will not show you.

The three-way against its closest rivals is in Claude vs ChatGPT vs Gemini, which covers the tier that is not priced like the other two at all.

4. Perplexity — best when the citation is the point

P

Perplexity at a glance
Alley Rating	4.4 — Power Tool, reviewed here
Free tier	Ten models displayed, all padlocked
Paid	Pro $20, Max $200
Best for	Research where the citation matters

This is a different product rather than a cheaper version of the others, and buying it as a general assistant is the most common mis-purchase in this category. It answers questions with sources attached and does not really draft. Most people who adopt it keep something else for writing.

Where it earns its place is any task where being able to check the answer matters more than the prose around it. When this site ran one identical question through three engines, Perplexity cited European Commission domains across ten sources where a rival leaned on a consultancy and two explainer sites. Both were correct; only one let you verify it.

The free tier is the weakest thing about it. Ten models are shown and every one is locked, which reads as an upsell surface rather than a usable product, and it is the one free tier in this list I would not recommend forming an opinion from.

Its pricing is confusing in a different way. The vendor’s own pages quote two different numbers for the same plan depending on which page you land on, because one shows monthly rates and another shows annual-equivalents without labelling either. Our Perplexity pricing guide works through which figure applies to you, and the short version is that a like-for-like comparison is not the one most third-party articles run.

Where it sits in the wider field is covered in best AI search engine, which put it against Google’s AI Mode and Brave on one identical query and compared what each one cited.

5. Grok — best free model picker, not yet rated

G

Grok at a glance
Alley Rating	Not rated — free tier tested, paid tier not
Free tier	Five selectable modes
Paid	SuperGrok $30 a month, or $300 a year
Best for	Model choice without paying

It earns a place here on one measurable thing: a free Grok account opens a picker with five entries, which is more than any other free tier in this list. That is a real difference for anyone not paying.

I tested it across four prompts on 27 August. It answered two factual questions correctly, including one where most of page one of Google is out of date, and it got there by opening the vendor’s own documentation rather than articles about the vendor. On a four-rule writing task it scored four out of four. The full runs are in ChatGPT vs Grok and Claude vs Grok.

Three honest caveats, which are why it carries no rating. I have not used the paid tier at all. It made me confirm my birth year before it would answer anything, which none of the others did. And grok.com/pricing returned a 404 when I checked on 28 August, so the price above comes from the in-app upgrade screen rather than a page you can send to a colleague.

One stylistic tic worth knowing: on the constrained writing task it produced 140 words containing no commas at all, having apparently generalised a ban on em dashes into avoiding punctuation broadly. The output complied with every stated rule and read strangely, which is the kind of thing that costs you an editing pass every time rather than once.

The finding I did not expect concerns sourcing. Asked a factual question about a competitor’s pricing, Grok ran three searches, opened the vendor’s own help centre and pricing pages, and quoted them. Claude, asked the same question, reached the correct answer through third-party articles about the vendor instead. Both were right; only one chain was checkable in a click.

That looked like a Grok advantage until I ran the same shape of test against ChatGPT’s free tier, which also went to the primary source and also got it right. So the habit is not unique to Grok, and the gap I first found was specific to Claude rather than general. It is a good example of why one comparison is never enough, and both write-ups are linked above.

Four things the whole category is bad at

None of the five solves these, and it is worth knowing which limits belong to the category rather than to the product you happen to be annoyed with.

The limit	Whose problem it is
Nothing here acts unsupervised	The category’s, not the product’s
No history moves between them	The category’s
The top tier may be the real complaint	Yours to check before switching
Every figure decays	Everyone’s, including this post’s

Anything that acts without you. All five wait to be asked. If the job is clearing an inbox or defending focus time on a calendar, no amount of chat quality helps, and you are shopping in the agent category described above.

Your accumulated context, if you switch. None of these import another assistant’s conversation history, saved instructions or remembered preferences. That loss is invisible on day one and obvious in week three. It is the strongest argument for adding a second assistant rather than replacing the first, and if you are weighing that switch anyway our ChatGPT alternatives roundup covers the same field from the other direction.

A cheaper answer than the one you are being sold. A large share of “this assistant is too expensive” is really “the top tier is too expensive”, and the top tiers here run to $200. Before migrating, check whether the complaint is about the product or about the rung you are standing on, because changing rungs is cheaper than changing tools and loses none of your history.

A stable answer. Every figure in this post was checked in August 2026, and free tiers in particular change faster than anything else here. Two of the five changed their model line-ups within the last quarter. Treat any ranking of this category, including this one, as a snapshot.

The pricing ladders, side by side

Five products, five different shapes, and the differences are structural rather than a few dollars either way.

	Free	Entry	Middle	Top	Annual option
ChatGPT	Yes	Go, then $20 Plus	$100 Pro	$200 Pro	None, any tier
Claude	Yes	$20 Pro	$100 Max	$200 Max	$17/mo at Pro only
Gemini	Yes	Google One AI	Higher One tiers	Higher One tiers	Via Google One
Perplexity	Yes	$20 Pro	—	$200 Max	Annual-equivalent shown
Grok	Yes	$30 SuperGrok	—	$30 SuperGrok	$300/yr

Three things fall out of that table. ChatGPT and Claude price their $20, $100 and $200 rungs identically, which makes the choice between them entirely about the product rather than the money — ChatGPT just adds a cheaper Go tier underneath, the one that carries ads. Grok is the only one with a single paid rung, so there is no upgrade decision but also nowhere to go. And ChatGPT is the only product here that offers no annual discount at any tier, which costs a year-long Plus subscriber $36 against Claude’s equivalent.

Gemini is the odd one out because its assistant is sold inside a storage plan, which makes a like-for-like row genuinely hard to write rather than an omission.

What changed in this category recently

A pillar is only as good as its freshness, so here is what moved in the months before this was written, all of it checked at source rather than repeated.

What moved	Why it matters here
ChatGPT Pro split into $100 and $200	Most comparisons still describe one tier
Claude became five named products	”Claude alternatives” is now five questions
Grok’s free tier gained a five-mode picker	The free-tier table above inverts
Consumer Sora was discontinued	ChatGPT no longer covers video

ChatGPT Pro became two prices. It is now $100 and $200 rather than a single $200 plan, and a great deal of the writing about it has not caught up. If you read a comparison that describes Pro as one tier, it predates the change.

Claude became five products. Code, Cowork, Design and Science on Anthropic’s pricing page, plus Security on Enterprise. That is why shopping for a Claude replacement returns such a strange set of results, and it is recent enough that most listicles still treat Claude as one thing.

Grok’s free tier got a picker. Five selectable modes on a free account is a more generous offer than either famous rival makes, and it is the single biggest reason the free-tier table at the top of this post looks upside down.

ChatGPT lost consumer video. The standalone Sora app was discontinued, and sora.com now redirects to a sunset page. If you were counting video generation as part of a ChatGPT subscription, that is no longer part of the deal, and Gemini is the one in this list that still reaches a video model.

None of that is stable. Two of the five changed their model line-ups within the last quarter, which is the honest reason to distrust any ranking of this category more than a few months old, this one included.

Five questions, in order

Are you going to pay? If no, go to Gemini or Grok, because they are the only two that give a free account real model choice. If yes, keep reading.

Does your output carry your name? If yes, Claude, and accept that you will need something else for images.

Do you need to check the answer? If yes, Perplexity, and expect to keep a second tool for drafting.

Do you want one subscription to cover everything? ChatGPT, which is the answer for most people and the reason it is our category pick.

Do you actually want something that acts rather than answers? None of these five. Take the four questions from the agent section above to whichever tool you shortlist, because they are the ones a ranking will not answer for you.

Final word

The interesting finding from a week of running the same prompts through these products is how close they have become where everyone is looking, and how far apart they still are where nobody is.

At $20 the differences are real but narrow, and mostly about which ecosystem you are already inside. On the free tier the gaps are wide, and they run against the reputations: the two best-known assistants give a non-paying user the least model choice of the five, and the least-known gives the most.

For most people paying for one: ChatGPT, on breadth. For anyone not paying: Gemini or Grok. For prose with your name on it: Claude, whatever else you also run.

And if what you meant by “assistant” was something that acts on your calendar rather than answers your questions, this was the wrong list — which is worth knowing before you spend $20 finding out, and worth taking the four questions above with you when you go looking.