Tokens, Explained for the People Who Pay the AI Bill

Most software still reaches the budget as a count: users, devices, cores. Multiply the count by a unit price and you know the bill before it arrives. AI tools are still sold that way at the front door. Claude Enterprise is $20 per seat per month on an annual plan, and Microsoft 365 Copilot is $30 per user per month paid yearly (list prices, accessed 23 October 2026).

Behind the seat sits a second unit. Anthropic bills Enterprise usage at API rates on top of the seat. GitHub moved Copilot to credits converted from tokens in June, which we covered in “Copilot Seat Prices Held. The Credit Meter Did Not.” Microsoft's own Copilot pricing page warns that some AI experiences are billed on usage, beyond the license fee. The unit behind all of these is the token.

This is the first post in a short series on the AI bill that SAM, procurement and finance teams are now asked to reconcile: a seat plus a meter, bought through many channels, and increasingly tied to API keys and agents rather than to people. It starts with the unit.

What a token is, in billing terms

A token is a fragment of text, or of an image or audio clip, that a model reads or writes. Vendors price tokens per million (often written MTok). How many tokens a page of text becomes depends on the model's tokenizer, the language and the content, so the same document can cost a different number of tokens on two models.

Every request has two sides. Input tokens are everything sent to the model: the instructions, the attached documents, the conversation so far, the definitions of any tools it may call. Output tokens are everything the model writes back. The two are priced separately, and that is where the first surprise sits.

Output costs about five times input

On the main current models of the three large API vendors, an output token costs five times an input token. List prices per million tokens, standard tier, accessed 23 October 2026:

The ratio is not universal (OpenAI's gpt-5.6-terra lists $2 in and $12 out), but it is the pattern. The practical consequence: two workloads on the same model with the same users can cost very differently. Extraction and classification read a lot and write a little. Drafting, code generation and report writing write a lot. If your forecast multiplies requests by an average price, it hides the mix that decides the bill.

Thinking tokens you pay for but never see

Reasoning models work through a problem before they answer. That work is made of tokens too. OpenAI's reasoning guide states that reasoning tokens are billed as output tokens, are not visible through the API, and are counted in the usage data of each response. It also warns that a request can stop at its output limit before any visible answer appears, so you pay for the input and the reasoning and get nothing back. OpenAI suggests reserving at least 25,000 tokens for reasoning and output when starting out. Google's Gemini price list labels its output rates as including thinking tokens.

So the length of the answer a user sees is a poor guide to cost. A two-line reply can carry thousands of billed output tokens. Usage data that separates reasoning tokens from visible output is the only way to see it.

Context: every turn re-sends the conversation

A chat feels like a conversation, but the model has no memory between requests. Each new turn sends the earlier turns again as input, plus any documents attached along the way. A long thread, or an assistant that pastes a 200-page contract into every question, pays for that text on every call.

Some price lists also change rate with prompt length. Google lists Gemini 3.1 Pro Preview at $2 per million input tokens up to 200,000 tokens and $4 above. OpenAI prices prompts above 272,000 tokens as long context, at $20 instead of $10 for gpt-6-astra input. Anthropic prices Claude Haiku 5.5 at $0.10 up to 100,000 prompt tokens and $0.50 above, while its other current models charge the standard rate across the full context window (list prices, accessed 23 October 2026). The same question can cost a different amount depending on what was attached to it.

Caching: cheaper reads, paid writes, sometimes rent

Prompt caching stores a repeated block of input, such as a long system prompt or a reference document, so that later requests read it at a discount. The discount is large and the details differ by vendor:

• Anthropic: cache reads at 0.1 times the input price, or 0.05 times on Opus 5.5 and Sonnet 5.5. Writing to the cache costs extra: 1.25 times input for a five-minute cache, 2 times for one hour.
• OpenAI: a separate cached-input rate, for example $0.10 per million on gpt-6.1-sol against $2 for fresh input.
• Google: a cached-input rate plus storage charged per million tokens per hour, $0.50 an hour on Gemini 3.8 Flash until 31 December 2026.

All are list prices from the vendors' pages above, accessed 23 October 2026. The cache hit rate is a cost driver in its own right. A workload that writes to the cache and rarely reads from it pays more than one that never caches at all.

Batch: half price for waiting

Anthropic, OpenAI and Google all take 50% off input and output for work submitted in batches that finish within hours rather than seconds. OpenAI's Flex tier is priced the same way. Overnight summarization, document classification and evaluation runs rarely need an answer in two seconds. If they run at the interactive rate, that is a choice someone made, often by default.

Tools, search and agents add their own meters

Models that call tools pay for the tool as well as the tokens around it. Anthropic and OpenAI list web search at $10 per 1,000 searches or calls, plus the tokens the results add to the prompt. Google gives 5,000 grounded searches a month across Gemini 3 and newer models, then charges $14 per 1,000. OpenAI's file search costs $2.50 per 1,000 calls plus $0.10 per GB per day of storage after the first GB. Anthropic's Managed Agents add $0.08 per session-hour of runtime (list prices, accessed 23 October 2026).

Agents multiply all of this. One task becomes a loop: plan, call a tool, read the result, try again. Each step is a new request whose input includes everything before it, and failed attempts are billed like successful ones. Peter Walker of OpenRouter, writing in OnlyCFO, calls the effective price after caching, waiting and retries the landed cost. For a SAM or FinOps team the useful unit is cost per completed task, not cost per request.

Speed, location and tokenizer change the rate

Three more multipliers sit on top of the list price:

• Speed. OpenAI's Fast tier (formerly Priority) is 2 times the standard rate. Google's Priority tier for Gemini 4 Argon is $3.60 in and $18 out against $2 and $10. Anthropic's fast mode for Opus 5.5 is $8 and $40 against $4 and $20.
• Location. Anthropic charges 1.1 times on every token category for US-only inference.
• Tokenizer. Anthropic states that the tokenizer in Claude 4.7 and later models produces about 30% more tokens for the same text. A lower price per token on a newer model does not guarantee a lower price per document.

Five questions to ask about any AI line item

1. What is the unit, and who converts it? Tokens, credits, requests or session-hours, and at what published rate.
2. Which model and which tier produced it? Standard, batch, priority or fast, and in which region.
3. How much was input, output and reasoning, and how much was cached? Without that split you cannot tell a wasteful prompt from a heavy workload.
4. Who or what generated it? A named user, an API key, a service account or an agent running on its own schedule.
5. What is committed and what is on demand? The seat floor, any prepaid credits, and what happens when they run out.

The question this leaves open

A token price is precise. The bill is not, because it combines seats, meters and keys bought by different teams through different channels. Where do the seats, the meters and the keys sit side by side, and who reconciles them? The next post in this series puts the four largest vendors' published prices next to each other.

Software Inventory Manager (SIM) inventories your software. For more on the AI bill, read If You Cannot Read the AI Bill, You Have No Position and Anthropic Split the Enterprise Seat From the Meter.