AI Subscription Jargon, Translated Into Plain English (2026)

TL;DR. AI pricing pages are written in a dialect. Here's the dictionary: a rolling window is a bucket that slowly refills (no midnight reset); a token is the AI's word-piece unit (~1,000 tokens ≈ 750 words); compute is the actual work behind a message, which is what you're really buying; a weekly cap stops round-the-clock heavy use on top of the hourly window; and reasoning modes burn allowance much faster because the AI is doing more work per message. Once you know these five, every AI pricing page reads the same.

"I still don't understand the AI jargon." A reader said exactly that to our site guide while standing on a page full of numbers — and they were right to. The words on AI pricing pages ("rolling window," "compute-based limits," "token allowances") aren't hard because the ideas are hard. They're hard because nobody translates them. Here's the translation, one term at a time, each with a single metaphor you can keep.

"Rolling window" — a bucket with a slow drip

Your message allowance doesn't reset at midnight. It refills continuously on a timer — usually five hours. Picture a bucket with a steady drip refilling it: every message you send takes a scoop out; wait a while and the level climbs back on its own.

Two consequences people miss: there is no magic reset moment you can wait for (the refill is gradual, tied to when you used the messages), and unused allowance never banks — a quiet morning doesn't earn you a bigger afternoon. Both Claude and ChatGPT work this way in 2026.

"Token" — the kilowatt-hour of AI

A token is the AI's unit of text — a word, or a piece of one. Rule of thumb: 1,000 tokens ≈ 750 words. You never buy tokens directly on a consumer plan, but they're the meter running behind everything: what you type, what you upload, and what the AI writes back all get counted in tokens, the way your electric company counts kilowatt-hours. When a vendor says limits are "token-based," it means the meter measures the amount of text processed, not the number of times you hit send. (Deeper dive: tokens and context windows, explained.)

"Compute" — what you're actually buying

Here's the concept that makes everything else click: vendors don't sell messages, they sell work. A two-line question makes the AI do a sliver of work. Uploading a 200-page PDF and asking for analysis makes it do orders of magnitude more. Both count as "one message."

That's why no honest pricing page promises a fixed message count anymore — two messages can differ in cost by 100×. When you see "limits vary with usage," it isn't evasion; it's the vendor saying we meter the work, not the sends. The practical move: short chats and small attachments stretch your plan; giant uploads and marathon conversations drain it.

"Weekly cap" — the second ceiling

On top of the rolling window, most plans add a weekly ceiling. Why both? The rolling window stops short bursts; the weekly cap stops someone from sitting exactly under the hourly limit around the clock, all week. Most people never touch it. If you've ever seen "you've reached your weekly limit" while the hourly meter looked fine — that's this, and it means your overall volume, not your pace, hit the line.

"Reasoning mode" / "Deep research" — the gas-guzzler setting

These modes make the AI think longer and search deeper before answering — and they consume allowance much faster than normal chat, because each message triggers far more work (see "compute," above). They're worth it for genuinely hard tasks. The mistake is leaving them on for routine questions, which is the single most common way people burn a day's allowance by lunch.

"Model tiers" — light, medium, and heavy machinery

Every vendor now offers a lineup of models — lighter ones that are fast and cheap to run, heavier ones that are slower, smarter, and hungrier. Your allowance stretches or shrinks depending on which one you pick: the same plan might give you hundreds of light-model messages or a few dozen heavy-model ones. (OpenAI even publishes its message estimates as ranges per model now.) The strategy is the same everywhere: default to the lighter model, save the heavy one for work that needs it. If you're not sure which Claude model fits a task, this picker answers it in two questions.

The limits of the metaphors

One honest caveat: all of these systems flex. Vendors adjust limits with demand, publish ranges instead of numbers, and change the rules more often than the pricing pages get rewritten. Treat any specific figure as a dated snapshot — including ours — and treat the concepts above as the stable part. The concepts haven't changed in two years; the numbers change monthly.

That's the whole dialect. Rolling window: refilling bucket. Token: kilowatt-hour. Compute: the work, which is what you're buying. Weekly cap: the second ceiling. Reasoning modes: gas-guzzlers. Model tiers: light vs heavy machinery. Now go re-read the actual numbers — or the ChatGPT and Claude plan breakdowns — and notice they've stopped being confusing.

More terms, one paragraph each, live in the AI glossary.