Is there an AI coding plan with no token limits? What the catch is

Updated 13 November 2026

Yes, but read "no token limits" carefully, because every plan limits something. The useful question is which limit you'll run into and whether you can plan around it. Here's what each kind of plan limits, and what NeverMetered limits instead.

Three ways AI coding is limited

What runs out under each kind of plan.
Plan styleWhat runs outWhat it feels like
Consumer subscription (Claude Pro/Max, ChatGPT/Codex plans)A shared usage allowance in rolling windows, such as a 5-hour session and a weekly capFine until a heavy day, then you wait for a reset or pay more. See Claude limits and Codex limits.
Per-token APIYour balance or budgetNothing is blocked, but every token costs money, so heavy agent use costs the most. See why the bill is high.
Flat plan priced by parallel streamsHow many requests run at the same timeNo token quota and no reset window. When more requests arrive than you have streams, extras wait briefly, then get a 429.

What "no token limits" means here

On NeverMetered, "unlimited tokens" means we don't meter or charge per token, and a paid plan has no token quota, no 5-hour window and no weekly cap. Our terms make that a commitment: we won't add a token quota, a 5-hour window or a weekly cap to a paid plan.

The catch

It isn't unlimited capacity, so while there is no token quota there are other limits. They're written in the same terms.

  • Parallel streams. Your plan sets how many requests can generate at once. A few extra requests wait in a queue and anything past that gets a 429. One person on one agent rarely notices. Several agents at once is where you need to choose the number carefully: see how parallel streams work.
  • Context window and output size. Each request has a maximum context (input plus output) and a maximum number of output tokens, set per plan.
  • Request length. One request can run for a limited time. That's long enough for very long generations, but it isn't forever.
  • Fair-share queueing. When the service is busy, waiting requests are ordered by recent use, so a heavy user can wait longer than a light one.
  • One model. You get one open-weight model, Qwen3.8-27B, not a choice of frontier models. It's a strong open coding model but not the same as the newest Claude or GPT, and for some tasks you'll want one of those. That's why many people use it as a second lane next to their main tool.
  • Speed. Generation is steady but not the fastest around. How it runs publishes what we measured.

Who this suits

It suits you if you run agents for hours a day or several at once, and a per-token bill or a weekly cap keeps getting in the way. It also suits you if you want to leave a job running overnight with no meter running, or you want to know next month's cost in advance.

It suits you less if you only use an agent in short bursts, because a per-token API or a cheaper subscription may cost less. The same goes if every task needs a specific frontier model. The cost calculator shows where the break-even sits for your usage.

Is there really no weekly limit?
On a paid NeverMetered plan there is no token quota, no 5-hour window and no weekly cap, and our terms commit to never adding one. Parallel streams, context size, request length and fair queueing still apply.
What happens when I use all my streams?
Extra requests wait in a short queue and start when a stream frees up. If the queue is full too, the request gets a 429 with a retry hint. Most coding tools retry on their own.
Can I try it before paying?
Yes. A free plan with a small daily allowance of requests lets you point your own agent at it. See get started.
Sources