How AI subscription limits actually work: the 4 metering models (2026)

Every AI subscription page promises "more usage" on the paid tier. What none of them say plainly is that "usage" means four genuinely different things depending on the vendor. Buying Claude Pro, Cursor Pro, and Lovable Pro looks like buying the same thing — a $20-ish monthly plan with "more" of something — but the meter underneath works in completely different ways. After tracking a dozen vendors daily, the limits sort cleanly into four models. Knowing which one you're buying is the difference between "I have a clear monthly budget" and "I'll find the ceiling when I hit it."

The four models at a glance

Model 1: the vendor-relative usage bucket

This is the most common model and the least informative. Claude tells you Max gives "5x or 20x more usage than Pro." ChatGPT now mirrors it exactly with 5x and 20x Pro tiers at $100 and $200. Perplexity, Mistral's rebranded "Vibe," and Grok all describe their paid limits the same way: more than the tier below, by an unstated amount. There is no published number anywhere on these pages — the meter is real, but it's the vendor's to see, not yours. The honest read is that you're buying a relative upgrade and trusting that "5x" is measured against something stable. QuotaLedger records the exact relative wording each vendor uses, because even that quietly changes.

Model 2: the time-windowed refresh cap

A handful of vendors tie the limit to a clock instead of a monthly total. Google's Gemini is the clearest: limits are compute-based, refresh every 5 hours, and run up to a weekly cap, and Gemini attaches concrete Flow-credit numbers (200 / 1,000 / 10,000+) on top. Devin (the rebranded Windsurf) publishes a daily-and-weekly refresh. This model is more honest about shape — you know it's a rolling bucket with a backstop, not an unlimited firehose — even when the exact compute-per-window stays vague. If you work in bursts, a 5-hour refresh behaves very differently from a flat monthly quota, which is why the window matters as much as the size.

Model 3: the per-user credit budget

Here the price literally buys a dollar amount. GitHub Copilot moved to a credit model of $15 / $70 / $200 in monthly credits; Replit and v0 include a fixed budget of Agent/credit spend and then bill per use; Cursor includes "a set amount of model usage" with overage at on-demand pricing. Bolt meters the same shape but in raw tokens rather than dollars — a per-user monthly token allotment (1M free, "starts at" 10M on Pro) that rolls over for one extra month, with higher token tiers instead of pay-per-use overage. This is the most transparent model about the amount — you know exactly how many dollars or credits you start the month with. The catch is that the budget is denominated in model spend, so the same $20 buys far fewer requests on an expensive frontier model than on a cheap one. The "limit" is a wallet, not a request count, and how long it lasts is on you.

Model 4: workspace-pooled credits

Lovable is the lone example among the thirteen, and it's the most counter-intuitive. Its monthly credits are pooled across the whole workspace, not handed out per seat — so adding teammates doesn't add capacity, it splits it. The sharpest tell: Lovable's Pro ($25) and Business ($50) tiers both include the same 100 monthly credits. Paying double doesn't buy more usage; it buys governance — roles, controls, SSO-grade workspace features. If you read it as a capacity upgrade you'll overpay. This is the clearest case of why "what model is this?" matters more than the price.

So how do you compare plans?

You can compare price cleanly — everything here is canonical USD, and the market clusters hard at $20 for the mainstream tier and $100–$200 for power users. You cannot compare raw "usage" across vendors, because a usage bucket, a 5-hour window, a dollar budget, and a shared credit pool aren't the same unit. The useful question isn't "who gives the most usage" — nobody publishes enough to answer that — but "which kind of limit fits how I work": a predictable budget (model 3), a burst-friendly refresh (model 2), a team pool (model 4), or a simple relative step up (model 1). For the specifics on any one vendor, the per-vendor reference pages quote each plan verbatim with the date checked.

FAQ

How do AI subscription limits work?

Four ways in 2026: vendor-relative usage buckets ("more"/"5x"/"20x", no number), time-windowed refresh caps (a clock-based reset up to a weekly ceiling), per-user credit budgets (a fixed dollar/credit allowance, then overage), and workspace-pooled credits (one shared pool for a whole workspace).

Why don't AI companies publish exact limits?

Compute-based limits let them adjust capacity without changing the advertised plan, and a vague "more usage" is easier to maintain than a hard number customers will hold them to. Gemini is the main exception, publishing a refresh window, a weekly cap, and Flow-credit counts.

What's the difference between a usage bucket and a credit budget?

A usage bucket is a relative, invisible allowance you find by hitting it. A credit budget is an explicit dollar/credit amount you start each month with, then pay per use — transparent about the amount, but it buys less on pricier models.

What are workspace-pooled credits?

A single monthly credit allowance shared across all users in a workspace instead of per seat. Lovable uses it; its Pro and Business tiers include the same 100 credits, so the upgrade is governance, not capacity.

Which plans actually tell you the limit?

Almost none give one hard general-usage number. Gemini publishes the most (refresh window, weekly cap, Flow credits); the credit-budget vendors publish the dollar/credit amount. Everyone else stays relative. QuotaLedger records the exact wording and the date checked.