From Aryan Vatsa | Product & Market Analysis
Unlimited AI Plans Survive. The Fair Use Clause Decides What You Bought
On this page
4 of the 8 AI vendors we checked on 21 October 2026 still use the word unlimited somewhere on a current plan. None of those 4 publishes the number at which unlimited stops. An unlimited AI plan is real, but its ceiling lives in a fair use clause, an abuse guardrail or a discretion clause, and those are the words your contract actually buys.
Key takeaways
- Unlimited survives in AI pricing, but only for narrowed scopes. GitHub applies it to code completions, OpenAI to eligible base models on Business, Perplexity to basic searches and Midjourney to its slower Relax queue.
- No vendor in the sample publishes where its unlimited feature stops. The only hard numbers sit on plans that do not say unlimited, such as Perplexity's Enterprise Max at 4,000 Pro Searches a week.
- Regulators already defined an honest unlimited claim, for telecoms. The UK ad regulator accepts the word only if a legitimate heavy user faces no extra charge or suspension, and any limits are moderate and clearly explained.
- The clause worth negotiating is the restriction remedy, not the price. OpenAI's help text says access is restored if no policy violation is found. Most vendors in the sample state no remedy at all.
Short answer
Some AI vendors still sell unlimited plans, but none of the 8 we checked publishes the ceiling. The real limit is set by fair use, abuse and capacity clauses that let the vendor slow, pause or restrict heavy users at its discretion. Before you sign, get the restriction trigger, the notice period and the remedy in writing.
What an unlimited AI plan means in 2026
Flat rate pricing means one fixed fee for a period, regardless of how much you use. Unlimited is the strongest version of that promise. It says the fee is fixed and the usage is too large to matter.
In software that was mostly true for twenty years. Serving one more spreadsheet or one more ticket cost the vendor close to nothing. In AI it is not true, because every prompt consumes compute the vendor pays for by the token.
So the word now carries a hidden denominator. The vendor still sells a flat fee. It also keeps a clause that decides when your usage stops being covered by it. That clause is the product, and most buyers never read it.
This post does not repeat the general case about shrinking allowances, which is covered in the analysis of AI shrinkflation and quiet usage cuts. It goes one level down, to the exact sentences behind the word unlimited, so your procurement team can quote them back to a vendor.
How we built the sample
We read the public plan page, help centre article and terms of service for 8 vendors on 21 October 2026. They were OpenAI, GitHub, Perplexity, Midjourney, Anthropic, Google, Microsoft and Cursor. The test for counting a vendor as unlimited was simple. Does the word appear, applied to any current feature, on any plan?
We excluded API rate cards, because they are metered by design, and private enterprise order forms, because we cannot see them. Two vendors, OpenAI and Midjourney, blocked automated page retrieval. For those we used the indexed text of their own help pages, which is noted against each quote.
Which AI vendors still sell unlimited plans
Half the sample still uses the word. None uses it for the whole product. Each vendor has drawn a boundary around the part it can afford to leave uncounted, and the boundary is where the useful reading starts.
| Vendor | What is called unlimited | The governing wording |
|---|---|---|
| OpenAI, ChatGPT Business | "Virtually unlimited messages for eligible base models" | Usage "must adhere to our Services Agreement". Guardrails "may occasionally involve a temporary restriction on your usage." |
| GitHub Copilot, paid plans | Inline suggestions, listed as "Unlimited" on Pro, Pro+ and Max. | "Code completions and next edit suggestions don't use credits." Rate limits exist "to protect against abuse". |
| Perplexity | "Practically unlimited basic searches" on Free, and "unlimited Pro Searches" on Education Pro | Pro and Max show "Weekly limits" with no number. Enterprise plans show 400 and 4,000 a week. |
| Midjourney | Unlimited image generation in Relax mode on Standard, Pro and Mega | Relax jobs queue behind others based on how much you have used Relax mode. |
| Anthropic, Claude Max | Not used. "5x or 20x the Pro plan's usage allowance" | "We may limit your usage in other ways, such as weekly and monthly caps or model and feature usage, at our discretion." |
| Google, Gemini Apps | Not used. AI Pro is "4x higher than standard limits" | "Limits may change without notice, including due to capacity constraints." |
| Microsoft 365, consumer plans | Not used. Chat is listed as "Extensive usage" | "Limits may vary based on the feature, entry point, and system conditions." |
| Cursor | Not used on its pricing page. | Anysphere "may modify or discontinue all or any portion of the Service at any time". |
Quotes are taken from each vendor's public plan, help or terms page. OpenAI and Midjourney wording is from indexed copies of their own help pages, because both sites refused automated retrieval. Wording on these pages changes without a changelog, so check the live page before you quote it in a negotiation.
One drift is worth recording. OpenAI's help article for ChatGPT Pro once described "unlimited access" to its flagship models. The current version describes "access" and adds that "some models have separate usage allowances". The price did not change with the wording, which is the pattern described in the 5-test standard for honest rate limit design.
The fair use clauses that set the real limit
Strip the marketing away and the 8 vendors use three kinds of clause to bound a flat fee. Each one works differently in a dispute, so your team should know which one it is reading.
Abuse and automation clauses
These define a class of illegitimate use and exclude it from the promise. OpenAI's Business help text lists "automatically or programmatically extracting data", sharing account credentials, and "reselling access or using ChatGPT to power third-party services". Anthropic's consumer terms bar access "through automated or non-human means, whether through a bot, script, or otherwise," except through an API key.
These are the most defensible clauses in the sample. A flat fee cannot survive a customer who wires it into a production pipeline. The problem is enforcement, because a heavy human user and a light script can look alike to an automated guardrail.
Capacity and discretion clauses
These let the vendor reduce what you get without any misuse on your part. Anthropic's consumer terms, effective 8 October 2025, say it may "increase or decrease capacity limits". Its Max plan page adds caps "at our discretion". Google's Gemini limits page says limits "may change without notice, including due to capacity constraints".
A discretion clause is the opposite of a flat rate promise. It means the fee is fixed and the quantity is not. I would treat any plan carrying one as a usage plan with an undisclosed meter, and price it that way internally.
Sharing and seat clauses
These tie the allowance to one person. Microsoft's consumer page says "AI benefits are only available to the subscription owner and cannot be shared with others". Anthropic says you "may not make your Account available to anyone else". For a team, this is the clause that quietly converts an unlimited personal plan into a per-seat cost. That is why the bundling versus metering decision matters more than the sticker price.
Four ways flat rate pricing survives in AI
Every unlimited claim in the sample survives by shrinking what the word covers. There are four scoping moves. A vendor can use more than one at once, and each narrows the promise further.
The first is scoping to a cheap feature. GitHub's code completions run on small, fast models, so leaving them uncounted is affordable. Everything expensive moved onto metered AI credits, a shift mapped in the breakdown of Copilot's credit billing.
The second is scoping to a model tier. OpenAI's Business promise covers "eligible base models." The reasoning and research models that cost more to serve sit outside the word, under their own allowances.
The third is scoping to a legitimate user. The abuse clauses above remove the use cases that would break the economics. The promise holds only for the customer the vendor had in mind when it set the price.
The fourth is scoping to speed. Midjourney does not cap Relax mode. It moves heavy users to the back of a queue. Its own documentation says your wait depends on how much you have used Relax compared with other users. It gives waits of up to 30 minutes on one page and 1 to 10 minutes on another.
The fourth move is the most honest of the four, and it is the one I would copy if I sold an AI product. Midjourney never takes access away. It changes performance and says so in its own docs. That is close to the line regulators drew years ago for broadband.
The test regulators already wrote for unlimited claims
None of this is new. Telecoms spent a decade arguing about what unlimited can mean when the service has a real marginal cost. Two outcomes from that argument translate directly to AI subscription limits.
In the United States, the FTC sued AT&T over unlimited data plans that were slowed after heavy use. The FTC said AT&T began throttling in 2011 after customers used as little as 2 gigabytes in a billing period, affecting more than 3.5 million customers. The case settled in November 2019 for $60 million.
The order matters more than the money. AT&T is barred from calling data unlimited "without disclosing any material restrictions on the speed or amount of data". Those disclosures must be "prominent, not buried in fine print or hidden behind hyperlinks". Andrew Smith, then director of the FTC's consumer protection bureau, said providers "must tell people about any restrictions on the speed or amount of data promised".
In the UK, the Advertising Standards Authority set a two-part test. An unlimited claim is likely acceptable if a legitimate user faces no additional charge or suspension for exceeding a fair use threshold. Any limits on speed or usage must be "moderate only" and clearly explained. The ASA defines a legitimate user plainly: "a customer who abides by the service's T&Cs."
Neither rule was written for AI. Both describe the exact gap in the sample. The word unlimited is fine. The missing disclosure is not.
Each AI fair use policy scored against the telecom test
Here are the 4 vendors that use the word, scored against the ASA's two criteria. This is our reading of public text, not a legal opinion. It is meant to give your team a structured way to argue, not a verdict a regulator has issued.
| Vendor and feature | Test 1: no charge or suspension for legitimate heavy use | Test 2: limits moderate and clearly explained | Our reading |
|---|---|---|---|
| Midjourney, Relax mode | Pass. Heavy use moves you back in a queue. No extra charge. | Partial. Queue logic and wait ranges are published. No threshold is. | Closest to passing |
| GitHub, code completions | Likely pass. Completions use no credits and rate limits are stated to be temporary. | Partial. The reason for limits is given. The limit is not. | Passes on access, silent on speed |
| OpenAI, Business base models | Partial. A temporary restriction can apply, aimed at misuse, with access restored if none is found. | Partial. Prohibited behaviour is listed. Thresholds are not. | "Virtually" carries the weight |
| Perplexity, Education Pro | Unknown. No enforcement wording found on the plan page. | Fail. The word is used and no limit is described. | Cannot be scored |
The OpenAI row deserves a fair hearing. A temporary restriction is a suspension in the ASA's sense. But OpenAI's help text ties it to misuse and states that if "policy-violating behavior is not found, your access will be restored". That is a remedy, and it is more than most of the sample offers.
The weakest row is the one with the strongest word. A plan that says unlimited Pro Searches, with nothing written underneath it, gives your team nothing to hold the vendor to. I would rather buy Perplexity's Enterprise Pro at a stated 400 a week than any plan whose ceiling is a guess.
Where this argument is weakest
The telecom comparison is useful and it is also imperfect. Three objections are strong enough that you should hear them before using this in a negotiation.
Compute is not bandwidth
A gigabyte costs a carrier roughly the same whoever sends it. A prompt to a reasoning model can cost a vendor many times what a short chat costs. Variance that wide makes a published ceiling genuinely hard to set, because one number either strands light users or bankrupts the plan on heavy ones.
Seen that way, discretion clauses are not bad faith. They are how a vendor offers a flat fee at all on a cost base it cannot predict. Remove the discretion and the likely outcome is not a better unlimited plan. It is no unlimited plan, and a metered bill, which is the trade covered in the guide to buyer concessions on usage based pricing.
The rules do not bind B2B contracts, and the sample is small
The FTC order binds one carrier. The ASA guidance covers UK advertising, mostly to consumers. Neither governs an enterprise order form your legal team signs, where the negotiated terms override the marketing page.
The sample is also 8 vendors, read on one day, from public pages only. Two of those pages could not be retrieved directly. Several of these pages change without a dated changelog, so a quote that is accurate today may not be accurate at your renewal. Treat the matrix as a dated snapshot, not a standing ranking.
Flat rate vs usage pricing: the clauses to negotiate
If you buy at team scale, the marketing word matters less than four contract terms. Here is how the public wording reads, and what your team can reasonably ask for instead. These are our suggested asks, not terms any vendor has agreed to.
| Term | Typical public wording | What to ask for |
|---|---|---|
| Capacity changes | Limits "may change without notice" | 30 days written notice of any reduction to included usage, with a right to exit pro rata. |
| Restriction trigger | Guardrails "may occasionally involve a temporary restriction" | The categories of behaviour that trigger a restriction, listed in the order form. |
| Restriction remedy | Often silent | A named support path and a target time to restore access when no violation is found. |
| Ceiling disclosure | "Extensive usage" or "weekly limits" | A usage figure, or a meter visible to your admins, for every feature called unlimited. |
| Refunds on termination | "All payments are non-refundable" | Pro rata refund if the vendor reduces capacity mid-term, not only if it terminates. |
The restriction remedy is the row I would fight for first. A heavy legitimate user blocked by an automated guardrail is the most likely failure your team will actually hit. Without a written remedy, the restoration timeline is whatever support decides that week.
The refund row is where most vendors will refuse, and you should expect that. Anthropic's consumer terms already concede a pro rata refund if it terminates without cause. Asking for the same when capacity drops is a small step from wording the vendor has already written. More clauses of this kind are collected in the CFO's checklist of AI contract clauses.
Finally, decide internally whether you want a flat rate at all. If your heaviest 10 users drive most of the value, a capped flat plan will throttle exactly the people you bought it for. A metered plan with a budget alert may serve them better. Compare both against the published rates in the comparison of Claude plan and API pricing before you commit.
Frequently asked questions
Which AI tools still have unlimited plans?
In our review of 8 vendors on 21 October 2026, four still used the word. GitHub Copilot paid plans list unlimited code completions, ChatGPT Business offers virtually unlimited messages on eligible base models, Perplexity offers practically unlimited basic searches and unlimited Pro Searches on Education Pro, and Midjourney offers unlimited Relax mode generation. None publishes the point where the unlimited feature stops.
What does fair use policy mean for an AI subscription?
A fair use policy defines the usage a flat fee is meant to cover and lets the vendor restrict use outside it. In AI plans it usually bans automated or scripted access, account sharing and reselling, and it reserves the right to slow or pause heavy users. The policy, not the plan name, sets your real limit, so read it before you sign.
Can an AI company limit an unlimited plan?
Yes, in almost every case we read. OpenAI says its guardrails may cause a temporary restriction. Anthropic reserves the right to decrease capacity limits. Google says Gemini limits may change without notice. Telecoms regulators have pushed back on similar practices, but those rules cover consumer advertising, not your negotiated enterprise agreement, so the contract is where to set limits on the vendor.
Is ChatGPT Business really unlimited?
Not without conditions. OpenAI's help text describes virtually unlimited messages for eligible base models, which excludes some more expensive models with their own allowances. Usage must follow the Services Agreement, which bans programmatic data extraction, credential sharing and reselling. Guardrails may impose a temporary restriction, and OpenAI says access is restored if no policy violation is found.
Is flat rate or usage based pricing better for AI tools?
It depends on how your usage is distributed. Flat rate pricing suits a team with steady, moderate use, because the bill is predictable and the fair use ceiling is rarely reached. Usage based pricing suits a team whose heaviest users drive most of the value, because a flat plan will throttle exactly those people. Model both against your last 90 days of usage before deciding.
What should I ask an AI vendor about usage limits before signing?
Ask four questions in writing. What behaviour triggers a restriction? How much notice do you give before reducing included usage? How is access restored if a restriction was a mistake, and how fast? Can our admins see a usage meter for every feature you call unlimited? A vendor that cannot answer these has not defined what it is selling.
Where to start
Pull the plan page and terms for your two largest AI subscriptions this week. Find the one sentence in each that lets the vendor change what you get, and paste both into your renewal file with today's date. If the wording changes before renewal, you will be able to prove it.
Then send your account manager the four questions from the FAQ above. Their answers, or their silence, will tell you whether you are on a flat rate plan or a metered plan with the meter hidden.
References
- Federal Trade Commission, AT&T to Pay $60 Million to Resolve FTC Allegations It Misled Consumers with 'Unlimited Data' Promises, 5 November 2019. Used for the settlement, throttling facts, disclosure order and the Smith quote.
- Advertising Standards Authority, Broadband and telecoms: Unlimited claims, 21 May 2025. Used for the two-part unlimited test and the legitimate user definition.
- OpenAI Help Center, ChatGPT Business models and limits and What is ChatGPT Pro, versions indexed October 2026. Read from indexed copies because the site refused automated retrieval.
- GitHub, Copilot plans and Rate limits for GitHub Copilot, read 21 October 2026. Used for unlimited completions and rate limit wording.
- Perplexity Help Center, Which Perplexity subscription plan is right for you, read 21 October 2026. Used for unlimited wording and weekly Pro Search limits.
- Midjourney, Comparing Midjourney Plans and GPU speed documentation, indexed October 2026. Used for unlimited Relax and queue wording. Read from indexed copies.
- Anthropic, Consumer Terms of Service, effective 8 October 2025, and What is the Max plan, read 21 October 2026.
- Google, Gemini Apps limits and upgrades; Microsoft, How AI usage works across Microsoft 365 subscriptions, updated October 2026; Cursor, Terms of Service, updated 3 September 2026. All read 21 October 2026.
The weakest part of this source base is that vendor help pages are revised without dated changelogs, and two of them could only be read through indexed copies. Every quote above is accurate to the version we read on 21 October 2026 and should be re-checked at your renewal.
Related reading