The word "unlimited" has never once meant unlimited in an AI vendor contract
# The word “unlimited” has never once meant unlimited in an AI vendor contract
Pull up the Wayback Machine and look at the Cursor Pro pricing page from May 2024. Right there, top of the tier, bold text: “Unlimited fast requests.” Now pull up the same page from September 2025. The word is gone. Quietly removed in an early-evening edit nobody announced. Between those two snapshots, the included fast request count had been cut from 500 to 400 to 250, and “unlimited” had stopped being a claim Cursor was willing to defend.
This is the cleanest single example of a pattern that runs through almost every AI vendor pricing page in the gotnerfed index. The word “unlimited” gets used as a marketing claim during the customer acquisition phase, then gets quietly replaced with a softer phrase (“generous limits,” “high usage,” “premium throughput”) once the burn rate on heavy users hits the threshold the finance team can’t ignore.
The interesting part is the timeline. “Unlimited” in AI vendor copy has a remarkably consistent lifespan. Roughly 9 to 18 months from launch to retirement, based on the timestamps I can pull from archived versions of major pricing pages.
The four phases of an “unlimited” claim
Phase one is launch. The product is new, the user base is small, and “unlimited” is mostly aspirational marketing. The vendor knows the heavy users will burn money, but the total user count is low enough that the loss is manageable. The word goes on the pricing page because it does what marketing words are supposed to do: it removes a buying objection.
Phase two is growth. The user base scales, but the per-user cost is now visible in the unit economics. The vendor still keeps “unlimited” on the page because removing it would damage acquisition conversion rates. Internally, they start introducing soft caps. A rate limit nobody mentions publicly. A model routing change that downgrades heavy users to a cheaper backend. A “fair use” clause buried in the TOS. A queue priority drop for usage above a threshold the user can’t see. The word “unlimited” stays on the page. The product underneath it doesn’t.
Phase three is the pivot. The cost curve has crossed a threshold. The vendor introduces a higher tier with a different name (Pro Max, Ultra, Business). The original “unlimited” tier is repositioned as the entry-level option, and the explicit unlimited claim is implicitly moved to the higher tier, where the price point makes the economics work.
Phase four is removal. The word “unlimited” is removed from the page entirely, often during a broader pricing page redesign that obscures the change. The functional cap that’s been there since phase two becomes visible. Users who joined during phase one and remember the original promise are told the change is “clarification, not reduction.” The receipts say otherwise.
Cursor’s “unlimited” had a 16-month run
The Cursor receipts let you reconstruct the timeline precisely. Pro launched April 2024 with “unlimited fast requests” prominently displayed. By January 2025, fast requests had been cut to 400 from 500. By June 2025, the model router was defaulting to a cheaper variant for users above a certain usage threshold (this one wasn’t even announced; it surfaced via a Reddit thread comparing response speeds). By September 2025, the word was gone from the pricing page. Sixteen months from launch to retirement.
That’s one of the longer runs in the index, by the way. Most “unlimited” claims have shorter lifespans.
Suno V5 went from “unlimited generations” to a 35% quota cut in nine months
Suno’s V4 launch in mid-2025 came with marketing copy promising “unlimited song generation on Pro.” That copy was on the homepage, on the pricing page, and in the launch tweet from the founders. Nine months later, in April 2026, the V5 launch came paired with a 35% reduction in monthly Pro generations and a hard cap that hadn’t existed before. The gotnerfed receipt scored it a 43.
What’s notable about the Suno case is that the “unlimited” claim was repeated in customer support replies during the V4 era. Multiple users on the gotnerfed comment thread for the V5 receipt posted screenshots of support tickets from late 2025 where Suno staff had reaffirmed that Pro generations were uncapped. Those replies didn’t include any caveats about future product versions changing the deal.
This is a familiar pattern. “Unlimited” claims tend to be stated by support, by marketing, by founders on podcasts, by the homepage, and on the pricing page. Every channel says the same thing during the customer acquisition phase. Then the cap arrives, the documentation gets updated, and the prior statements get treated as obsolete rather than as commitments that were broken.
Midjourney Basic’s “unlimited relaxed” got specific because the term was already unstable
Midjourney’s approach is worth a close look because they tried to thread a needle. The Basic tier originally offered “200 fast generations per month plus unlimited relaxed generations.” The qualifier “relaxed” was supposed to make the cap-free claim economically viable, since “relaxed” generations went through a slower, cheaper queue.
It worked for a while. Until February 2026, when the Basic tier was removed entirely and the minimum subscription jumped from $10 to $30. The gotnerfed receipt on this one scored a 73, the same band as the Claude Code Pro removal. The “unlimited relaxed” promise had effectively been honored for the duration of the tier’s existence, but the tier itself was retired, which made the promise moot.
This is the failure mode of “unlimited” claims that aren’t outright broken. The vendor doesn’t have to break the claim. They can just retire the tier the claim was attached to. The user’s recourse is the same either way: pay more, or leave.
Anthropic’s “unlimited Pro” lasted six months
Anthropic Pro launched in late 2024 with marketing copy that called the chat product’s usage “essentially unlimited” for most users. By mid-2025, that copy had been softened to “extensive usage” and a message cap had been introduced. The cap was generous enough that most users didn’t hit it. The change still mattered because the copy revision was the first step in a sequence that ended with the April 2026 Claude Code tier removal.
This is what makes “unlimited” copy retirement a leading indicator in the gotnerfed framework. The vendor doesn’t change the wording without a reason. The reason is almost always that the finance team has flagged a cost trajectory the marketing claim is now exposed to. Once the wording changes, the structural moves follow within 12 to 18 months. The pattern is consistent enough that you can use the copy change as a planning signal even before any feature change ships.
What the TOS says when the pricing page says “unlimited”
This is the part most users don’t read, and it’s where the legal reality diverges from the marketing claim.
Almost every AI vendor’s TOS contains a “fair use” or “reasonable use” clause that explicitly preserves the vendor’s right to throttle, cap, or terminate accounts whose usage falls outside what the vendor considers reasonable. The clause is broadly worded by design. The vendor defines what “reasonable” means, and the definition is allowed to change without notice in most cases.
Cursor’s TOS during the “unlimited fast requests” era included a fair-use clause that gave Cursor the right to “implement rate limits or other usage controls to ensure platform stability and fair access for all users.” That clause was the legal cover for every subsequent cap reduction. The word “unlimited” on the pricing page wasn’t legally binding because the TOS preserved the vendor’s discretion.
This isn’t an accident or a poorly drafted contract. It’s the standard structure. OpenAI’s terms include similar language. Anthropic’s do. Suno, Midjourney, GitHub, Runway, and Replit all have variants of the same clause. The pattern is that the pricing page says “unlimited” and the TOS says “subject to fair use as we define it.” Marketing wins the customer; the contract preserves the option to change the deal.
If you want to know what a vendor’s “unlimited” claim is worth in practice, the test isn’t to read the pricing page. The test is to read section 4 or 5 of the TOS (usage limits, fair use, suspension of service). If the clause includes any version of “we may modify usage limits at our discretion,” then “unlimited” means “unlimited until it isn’t.”
Five questions to ask before paying for an “unlimited” plan
Worth running through these before you commit to a vendor whose marketing leans heavily on the word.
What’s the average lifespan of “unlimited” claims in this product category? For AI tools specifically, the gotnerfed index says 9 to 18 months. For SaaS broadly, it tends to be longer, but AI tools have shorter cycles because the underlying compute economics shift faster.
Has the vendor already softened the wording on their own page in the last 12 months? Wayback Machine is free. Run a diff. If “unlimited” became “extensive” or “generous” or “high usage” recently, the structural cap is already in motion.
What does the fair-use clause in the TOS say? Specifically: who defines “fair,” what’s the notice period for changes, and what’s the remedy if you exceed the undisclosed cap (throttling, suspension, forced upgrade, or quiet model downgrade)?
What’s the vendor’s funding stage? Pre-Series-C vendors are still in the customer acquisition phase, which is when “unlimited” claims are most aggressive and least defended. Post-Series-C vendors are under pressure to show revenue per user, which is when caps appear.
What’s the cost per heavy user vs. the subscription price? If your usage costs the vendor more in inference than they’re charging you in subscription, you’re being subsidized, and the subsidy will end. The only question is when.
OpenAI’s “unlimited GPT-4o” was the most visible breakage of all
The case worth dissecting in some depth is OpenAI’s “unlimited GPT-4o” for ChatGPT Plus, because it was the most public version of the pattern and because the wording shifts are still visible in the archived pricing pages.
When GPT-4o launched in May 2024, OpenAI’s marketing positioned it as a major upgrade in capability that was now included on the $20 Plus tier with what they called “much higher usage limits than GPT-4.” The CEO described the rollout in a public tweet as making “the best model available to everyone.” The pricing page didn’t use the literal word “unlimited,” but the framing across launch coverage, support documentation, and product copy consistently implied that the cap was high enough that most users would never encounter it.
By mid-2024, ChatGPT Plus users were reporting hard message caps on GPT-4o that hadn’t been disclosed in the launch. The published number floated between 40 and 80 messages per 3 hour window, depending on which support article you read on which day. The number was adjusted without announcement multiple times during the second half of 2024. By 2025, the cap had been formalized at 80 messages per 3 hours with an additional “model-aware” throttle that further reduced the cap during peak load.
The thing that’s interesting about this case is that the breakage wasn’t a single event. It was a series of small downward adjustments, each one too minor to trigger a press response, that cumulatively rewrote what the Plus tier offered. Users who measured their usage in late 2024 and compared it to late 2025 found the effective cap had dropped by about 30%. The pricing page had been quietly updated to remove the “much higher” language by spring 2025. No receipt, no announcement, just a slow walk-back of the original claim.
This is the harder version of “unlimited” to track because there’s no single event to score. Each individual adjustment is a 16-point price-increase receipt or a 20-point model-swap receipt at most. The cumulative effect across 18 months is closer to a 60-point structural change. The index can capture the steps; it can’t easily capture the trajectory.
If you were on ChatGPT Plus from launch and still pay for it today, you’re paying the same $20 for a meaningfully smaller product than the one that was sold to you. That’s the realization that catches users by surprise when they sit down to compare year-over-year. The price held. The product shrank.
What “unlimited” should mean
Worth being specific about the version of “unlimited” that would hold up under scrutiny.
An “unlimited” claim is legitimate if the vendor publishes a hard cap above which throttling kicks in, names the cap in plain language on the pricing page, and contractually commits to 90 days of notice before changing it. That’s it. That’s the version of the claim that survives contact with reality.
Almost no AI vendor offers this. The exceptions are a small number of enterprise contracts where price-lock and usage-floor clauses are negotiated individually. For consumer and prosumer tiers, the word “unlimited” should be read as “unlimited for now.”
You’d think enough of this would have happened by now that customers would have learned to discount the claim. The signups suggest otherwise. Every new AI tool launch in 2026 that’s used the word in its marketing has seen the same adoption curve as the launches that used it in 2024 and 2025. The pattern repeats because it works. The receipts are public. The word still moves customers.
The next time you see “unlimited” in an AI tool’s pricing, run two searches. First on Wayback Machine for the prior version of that page. Second on gotnerfed for any receipts on the vendor. The combination tells you, within a few minutes, where in the four-phase cycle that vendor currently sits. And whether your subscription is buying you a service or buying you a position in someone else’s growth funnel.