Tokens, cost and caching
Send a message and look underneath the reply. Player One tells you what that exchange cost: how much text went in, how much came back, and how much of the input was served from cache. In a chat you have been in a while it often reads something like "94% cached".
This page is about those numbers: why the same question costs more in an old chat than a new one, and what to do about it.
What a token is
A token is the unit AI providers count and charge in: a chunk of text, roughly three quarters of a word. Everything is priced this way, what you send and what the model writes back, and writing costs more per token than reading. You never have to count them: cost follows the amount of text moving, and most of it is not the sentence you typed.
Why a long chat costs more with every message
A model has no memory of its own between messages, so every time you send one the whole conversation goes back to the provider: the instructions your assistant runs on, the skills it has loaded, everything either of you has said, and every document it has opened. Your new sentence might be twenty words, while what actually gets sent is twenty thousand.
That is why a thread you have been in all afternoon costs more per message than a fresh one, and answers more slowly, even when the questions are the same size. The gauge beside the model name in the message box shows how full it has become, and Conversations covers what happens when it fills.
What caching does about it
Caching is the providers' answer to that repetition. Because the start of the conversation is identical every time, the provider can keep it ready to reuse rather than read it in from scratch, and charges that part of your message at a much lower rate.
| What is being paid for | Roughly what it costs |
|---|---|
| New text | the full rate |
| Text read back from the cache | a fraction of the full rate |
| Putting text into the cache | slightly above the full rate, once |
The catch is that a cache is short-lived: typically minutes, stretching to about an hour where the provider offers a longer one, and it expires quietly if nothing uses it. So a conversation you keep moving stays cheap, while one you abandon for two hours pays full price again on the next message before settling back down.
What breaks a cache
Four things, in rough order of how often they catch people out.
A long gap. The entry expires and the next message rebuilds it.
A change at the beginning. The cache holds only while the start of the conversation stays exactly the same. Editing your project's standing instructions, changing what an agent is told to be, or turning a skill on or off rewrites that opening, and it is paid for again. Not a reason to avoid changing them, just a reason not to fiddle mid-session.
Switching model or provider part way through. A cache belongs to one model at one provider, so moving starts a new one. Player One marks the switch in the thread.
Compaction. When a conversation outgrows the model's context window, Player One summarises the older part of it in place. That rewrites the history, so the cache starts again.
Start a new chat for a new topic
The cheapest habit here is the simplest: when the subject changes, start a new chat. A new topic gains nothing from three hours of unrelated history, and pays for all of it on every message.
If the old thread taught your assistant something worth keeping, ask it to write that up as a skill before you leave. The lesson outlives the conversation and the new chat starts with room to work. See Teaching it.
Keep side work out of the main chat
When part of a task will be long, noisy or slow, ask for a helper or a specialist to do it. A helper runs that slice in its own conversation and only its conclusion comes back into yours, so the working-out never joins what you re-send later.
It is the biggest lever most people have; Agents, delegation and team chat explains how to ask.
Choosing what runs where
Two things decide how expensive a piece of work is: which model serves it, and how hard it thinks before answering. Both live on Admin > Agent Models (/admin/models), and you can change them for one chat from the model chip in the message box.
The real saving is keeping routine, mechanical work on a cheaper model: naming chats, keeping the search index up to date, background traffic nobody reads. Player One's built-in rules already send that work to your cheapest model. They are listed under Default rules on the Advanced Routing tab (/admin/models/routing), and a rule of your own for the same job overrides them, so that is the place to look if a bill looks wrong.
The false saving is turning the thinking down on work that matters. A model given too little room to think makes mistakes, and the retries cost more than the thinking would have. Ask your assistant where the spend is going before you change anything: it can read your usage and tell you.
The two ways of paying for AI
Player One works with a wide range of AI providers, connected on Admin > Providers (/admin/providers). There are two ways to pay, and they behave very differently.
A consumer subscription is a plan you already hold, such as a ChatGPT plan, connected by signing in. It is a flat monthly fee, usually subsidised, with the provider's own limits. You cannot overspend on it: at the limit you wait rather than pay more. That suits interactive, human-paced work and is the cheaper option for most people, though it is finite and heavy automated work will exhaust it.
An API key is billed per token: you pay for what you use, there is no ceiling, and the bill follows the usage. That suits work which has to run whatever the hour, jobs and scheduled tasks that must not be blocked by a plan limit. It needs watching, because nothing stops it.
Most people end up with both. Connect several accounts and Player One tries them in the order you set, falling over to the next when one is rate-limited. It prefers your subscriptions and asks before putting work onto anything that costs per token. Nothing runs until you have connected at least one account of your own, which is the first thing a new installation asks for. Connect your AI accounts walks through it, and Models and providers covers switching.
Where to see what you are spending
Admin > Usage (/admin/usage) is the honest picture: totals and a chart over time, broken down by model and by agent, over today, the last 7, 30 or 90 days, or all time. Its API calls tab (/admin/usage/calls) lists what outside apps have spent through your installation, one row per call.
Closer in, the reply footer covers one exchange, /usage and /context cover the conversation, and the provider button in the top bar shows how much of each connected plan you have used. If something looks wrong, just ask: "where is my spend going?" gets a plain answer rather than a spreadsheet.