What "Unlimited" Means Here — tokens, pacing, and the exact second
Every flat-rate AI plan on the market has limits — Claude Pro and Max have session and weekly caps, ChatGPT Plus has message windows, Cursor and GitHub Copilot meter “premium requests”. The difference between those plans is not whether limits exist; it is whether the vendor tells you what they are and when they apply. This page is our answer, written to be linked from the product itself: if a 429 response pointed you here, everything below applies to you.
What is genuinely unlimited
Section titled “What is genuinely unlimited”Tokens. Nothing you send or receive during your reserved hours is metered, counted against a quota, or billed. There are no overage charges, no per-token line items, no end-of-month surprise. Your invoice is decided the day you subscribe and never moves with usage. This is not marketing rounding: the billing system has no code path that can charge you for a token.
That is the whole point of the product, and it is why the economics work the way they do: past the break-even volume, every extra token widens your savings instead of your bill. The arithmetic — at what monthly volume the flat block overtakes per-token list prices, pool by pool — lives in its own living report: Flat Rate vs Per-Token. The short version: most people who run a coding agent daily pass break-even in the first week of the month.
What is finite
Section titled “What is finite”Three things, all stated plainly:
- Your reserved hours. A subscription is a daily 8-hour block (or several). Outside it, the key is closed — that is the product you chose at checkout, not a surprise.
- One request at a time per key. Concurrency is bought, not assumed: each subscription serves one request at a time, and combined keys stack parallel capacity where subscriptions overlap.
- Sustained delivery rate. Within your hours, delivery is paced. Your block includes a sustained throughput — the same way an unlimited broadband line includes a speed, not an infinite one. Tokens are never the unit; time is.
How pacing behaves
Section titled “How pacing behaves”This is the part most vendors keep vague, so here it is exactly. The first thing to know is when pacing engages at all: it is not part of normal operation. Bursts, heavy sessions, a busy day after a quiet one — all of that rides the accumulated headroom and serves at full speed. A key meets the pacer only when its sustained consumption runs notably and continuously above what its block sustainably delivers, and keeps running there. When it does engage:
- Short waits are invisible. When your next request only needs a few seconds of headroom, we hold it server-side and serve it the moment it is ready. You see a slightly slower first byte, never an error. Most pacing events are this kind.
- Longer waits are a 429 with the exact second. When the wait is real, the response is a 429 whose
Retry-Afterheader carries the true clock wait — not a vague “try again later”, not a flat number that hides the real one. Retry at that second and the request serves. Standard SDKs (OpenAI, Anthropic, Vercel AI SDK) honor this header automatically for waits up to a minute. - Beyond a minute, the error message itself states the real wait at your current sustained usage, and the way to remove it immediately (more subscriptions in a combined key). No hidden countdown.
- Pacing refills continuously. There is no weekly cliff, no “locked out until Thursday”, no monthly reset to wait for. The clock in the header is the whole story: when it elapses, you are served.
- It never bills and it never escalates. A paced request costs you nothing and flags nothing. There are no strikes, no account reviews triggered by usage, no retroactive charges. The pace is the same rule for every key in the pool, applied by arithmetic, not by a human looking at your account.
What a request draws from that sustained budget is not a token count. It emerges from many interacting factors — the size and shape of the request, the model’s context handling, the hardware it is served on at that moment, how much of the input is served from cache, the load on the pool at that hour, and more. The same nominal request can weigh very differently depending on that mix, which is why we publish the rule and the behavior rather than a formula. The rule itself is the fair-use clause in our Terms, §1.2; this page is what it means in practice.
The floor and the ceiling
Section titled “The floor and the ceiling”Because tokens are free and time is finite, the value of a block is a multiple of what the same tokens would cost per token at list prices. That multiple has a floor and a ceiling, and we state both honestly:
- The floor is arithmetic. From twice the break-even volume, the flat block is at least 2× cheaper than per-token list prices — by definition, not by promise. Pacing does not stop a normal workload from getting there: with the cache-hit rates agent workloads actually show (70–95% of input served from cache, which we optimize for), the sustained budget of one key sits comfortably above twice break-even.
- The ceiling is not a fixed number, because what a request really costs to serve is not a token count: it moves with your load shape, your context size, your cache-hit rate, and the pool’s capacity at that hour. The same million tokens can be several times cheaper or dearer to serve depending on that mix. What we can report is what we observe: as of November 2026, the heaviest keys we serve sustain a list-price equivalent of roughly 5–15× their subscription price before pacing holds them there. If you want one number as a mental model, think “around 10×” — as a measured observation, not a guarantee.
- Above the ceiling, the product is a different one. A workload that needs 30× sustained is not a subscriber’s workload; it is a fleet’s. The honest answers there are more subscriptions in a combined key (each adds its own budget and its own concurrency) or per-token pricing, which exists precisely for unbounded batch consumption.
”But you said unlimited”
Section titled “”But you said unlimited””If you hit a 429 and feel the word “unlimited” was the lie, here is the claim and the test, stated so you can check both:
- The claim is unlimited tokens on a flat bill — and that part is absolute. Count what your usage this month would have cost at any published per-token list price and compare it with your invoice. For most paced keys that ratio is the strongest evidence on the page: pacing engages because the key is extracting several times its price in list-equivalent value, every month, with no meter running.
- What paces is delivery, in time, with the exact wait disclosed in the response that paces you. A plan whose limits are published to the second, refill continuously, never bill and never lock you out for days is — as far as we can tell — the most explicit limit disclosure of any flat-rate AI plan on the market. We would genuinely welcome the comparison: put this page and any competitor’s fair-use page side by side, or hand them both to an LLM and ask which plan tells you more about what you actually bought.
When you should not subscribe
Section titled “When you should not subscribe”The same arithmetic says when the flat block is wrong for you, and we would rather say it here than in a refund thread: sporadic use (below break-even volume, per-token is cheaper — see the curve), round-the-clock traffic on one block (buy the blocks you use, or spread subscriptions across blocks with a combined key), and unbounded batch jobs that need everything now and nothing tomorrow (that is per-token’s product, not ours).
Common questions
Section titled “Common questions”Is pacing a penalty or a flag on my account? No. It is the same arithmetic applied to every key, it costs nothing, it records nothing against you, and it clears by itself at the second stated in the header.
How do I get more throughput today? Add subscriptions and fold them into a combined key: budgets and concurrency stack where hours overlap, and coverage extends where they do not.
Is there an SLA? No. We do not guarantee throughput, latency or a savings multiple; the Terms govern. What this page adds is precision about how the product behaves, not a service-level commitment.
Where is the break-even math? In the Flat Rate vs Per-Token report, recomputed from live prices on every build.