Unlimited tokens or per-token pricing: how to choose
Most LLM APIs charge for every token. That makes every agent loop, retry and experiment a cost decision. A flat, time-based plan charges for a window of access instead. This article explains how both models work, when each one wins, and how MIRASTACK LABS sells flat capacity to individuals (endlesstokens) and to teams (unlimitedtokens).
Why per-token pricing exists
A token is a small piece of text, often part of a word. A language model reads tokens in and writes tokens out. Metered APIs count both and bill for the total. The model is easy to understand and it matches how the provider's costs behave at small scale: more tokens mean more GPU work, so more cost.
For a person who sends a few prompts a day, this is fair. The bill is small, and it follows use.
The trouble starts when software, not a person, decides how many tokens to use.
Where metered billing hurts
Agents re-send their context
A coding agent does not send one prompt. It reads files, calls tools, reads the results, and tries again. On every turn it sends the conversation so far, plus new material. Long sessions carry tens of thousands of tokens of context on each call. In our own capacity tests we replay real coding-agent sessions with contexts of 50,000 to 130,000 tokens, with users running several sessions at once. Under a per-token meter, each of those turns is billed.
Loops make cost hard to forecast
An agent that gets stuck can loop. An evaluation run that is repeated after a prompt change doubles its cost. A team that adds a second assistant to an internal tool adds a second stream of tokens. None of these are bugs in the billing. They are normal behaviour, and they make the monthly total hard to predict before the month starts.
People start rationing
The worst effect is behavioural. When every call has a price, people shorten prompts, skip the extra verification pass, avoid running the experiment a third time, and keep agents on a tight leash. The tool was bought to make people faster, and the meter makes them careful in the wrong places.
Finance feels it as well. A budget that is a range, not a number, is hard to approve and hard to defend at review.
What a flat, time-based plan is
A flat-rate plan changes the thing you buy. Instead of buying tokens, you buy access to a pool of model capacity for a window of time. Inside the window, tokens are not counted. The price is agreed before you start.
This is how broadband is sold. You do not pay per megabyte; you pay for a connection, and the provider sizes the network so that the connection works.
Flat does not mean without limits. Someone still has to size the hardware, and capacity is finite. A well-run flat plan puts the limit where it can be explained and planned for:
- Time. The window has a start and an end.
- Parallel sessions. How many requests a key can have running at once.
- Keys and seats. How many people or systems have access.
- Fair queueing. When demand peaks, requests are queued in a fair order and not dropped.
These are the levers. Token count is not one of them.
Metered or flat: a side-by-side view
| Per-token (metered) | Flat, time-based | |
|---|---|---|
| What you buy | Tokens used | Access for a window of time |
| What drives cost | Prompt length, output length, retries, loops | Seats, sessions and term, agreed up front |
| Forecasting | A range, known after the fact | A number, known before you start |
| Effect on behaviour | Encourages short prompts and fewer attempts | Encourages using the capacity you bought |
| Where limits sit | On spend | On concurrency and time |
| Best for | Low, occasional or spiky use; models only one lab offers | Sustained, high-volume or automated use |
When per-token pricing is still the right choice
A flat plan is not always better. Metered billing wins when:
- Use is low or occasional. If you send a handful of requests a week, you would pay for a window you do not fill.
- You need a specific frontier proprietary model. Those models are offered by their labs, usually by the token. We do not resell them. If your task needs one, use it from its lab.
- Use is very spiky. If you need a burst of capacity once a quarter, a metered API scales from nothing and back to nothing.
If your use is steady, automated or shared by many people, the arithmetic usually turns the other way. Do the sum with your own numbers: take a representative week of agent traffic, price it per token, and compare it with a flat quote.
How we sell flat capacity
MIRASTACK LABS hosts and runs open-weight models on GPUs it operates. We do not proxy your requests to another provider and we do not resell anyone else's capacity. That is what makes a flat price possible: we know our own costs and we size our own capacity.
For individuals: endlesstokens
endlesstokens is a self-serve time pass for developers and their coding agents. You buy a window of time and use open models as hard as you like inside it. Tokens are not counted. It works with any OpenAI- or Anthropic-compatible client and coding agent: you point the client at the endpoint and use your key. You sign in with email, GitHub or Google, and passes are bought on endlesstokens.ai, where the options and their terms are shown at checkout.
This is the "unlimited tokens" product. The limit is the time on the pass and the number of requests your key can run at the same time, not a token counter.
For teams: unlimitedtokens
unlimitedtokens is the same idea for an organisation. Per-seat access to open models for a whole workforce, billed by time and never per token, for organisations from 60 seats, on quarterly or yearly agreements. Each person gets their own keys. Administrators can see how capacity is being used. Seats, parallel sessions per key and keys per user are set in the agreement, so the plan fits the way your teams actually work. Invoices go to your company, with GST.
Your tools stay the same. The API is OpenAI- and Anthropic-compatible, so editors, agents and internal tools keep working after a change of base URL and key.
For bounded projects: infinitetokens
Some work is not a steady stream of seats. It is a scoped project that needs a large amount of capacity for a fixed time. That is infinitetokens: training clusters and inference reserved for you, on your own endpoint, for 3, 6, 9 or 12 months. We cover it in Infinite tokens for a finite time.
How to decide
Ask four questions.
- Who uses it? One developer: a pass. Sixty or more people: a team agreement. A model team with a project: a dedicated endpoint.
- How steady is the use? Steady and automated use favours flat. Rare use favours metered.
- Which models do you need? If open-weight models do the job, flat capacity is available. If you need a specific proprietary model, buy that from its lab.
- Does the budget need a fixed number? If finance wants one, a time-based agreement gives it to you.
Many organisations end up with both: a flat plan for the daily volume and a metered account for the occasional proprietary model.
How a team plan is set up
You tell us the shape: how many people, which tools, and what they do with AI. We agree the seats, sessions, keys and term, and write them down once. Keys are issued per person and connected to the tools they already use. At renewal you add seats and revisit the terms. There is no token meter at any step.
What to ask any flat-rate provider
If you compare plans, ask each provider the same questions. The answers matter more than the headline.
- Where do the models run, and are requests passed to another provider?
- What are the limits on parallel sessions, and what happens when demand peaks?
- Are prompts stored, or only metadata?
- How is capacity measured and sized before it is sold?
- How does the plan treat the heaviest users?
We answer these on our technology page: capacity is measured with real agent sessions at rising concurrency, and sized per GPU, before it is put on sale.
Talk to us
If your people or your systems use models all day, send us three facts: the number of people or systems, the tools they use, and where the work must run. We will come back with an agreement shaped to that workload.
Talk to us or write to hello@mirastacklabs.ai. For a team plan, start on the unlimitedtokens page. For one developer, go to endlesstokens.ai.