Guide · AI costs
Cutting Claude and ChatGPT costs in your organisation.
Your AI bill depends mostly on the model chosen, the length of conversations and how you pay. Here are the levers, from the simplest to the most far-reaching.
Where the bill comes from
What you pay for, request by request.
Claude and ChatGPT are paid for in two ways: a per-employee subscription, or usage-based billing, when the organisation goes through the API or a consumption-based plan.
With usage-based billing, every request is charged by the number of tokens, the pieces of words the model reads and writes: those sent to it, as input, and those it produces, as output.
At Anthropic, Opus 5.5 costs twice as much as Sonnet 5.5, and a token written by the model costs five times a token sent to it. Model choice and the length of exchanges therefore make up most of the bill.
On a per-employee subscription, spend does not move with usage: what counts then is quotas, governance and footprint. The levers below act directly on the bill when usage is billed per token.
Anthropic and OpenAI public documentation as of 16 November 2026.
The levers
Six levers, from the simplest to the most far-reaching.
- 01
Choose the model for each request
The most powerful model is often used for requests that do not need it, such as summarising minutes or rewording an email. A mid-range model like Sonnet keeps, on our benchmark of 1,421 public business conversations, almost all the quality for half the price; the most powerful one stays on hand for long, difficult requests.
- 02
Set the reasoning effort
Recent models think before they answer, and that thinking is billed like written text. Claude and OpenAI’s models let you set this effort: a low effort is enough for simple requests and uses fewer tokens.
- 03
Start a fresh conversation
The model keeps nothing from one request to the next: with every message, the whole conversation is sent back to it and billed again. A conversation that drags on costs more and more; it is better to open a fresh one as soon as the topic changes.
- 04
Reuse what does not change
Instructions and documents that come back from one request to the next can be cached. At Anthropic, reading them again costs a tenth of their normal price, a twentieth on Opus and Sonnet; at OpenAI, caching is on by default.
- 05
Group what is not urgent
For bulk work that can wait, such as sorting documents or summarising archives, Anthropic and OpenAI offer batch processing at half price, with results within the day.
- 06
Give the right context first time
A poorly framed request calls for follow-ups, and every follow-up sends the whole conversation again. An AI that already knows the company reads one record instead of entire documents: it needs fewer tokens to get the right context.
Measure
Measure before you cut.
A spending cap limits the bill, but it cannot tell Opus launched for a trivial question from a legitimate heavy task: it cuts both.
First you need to know which models are used, in which workspaces, and at what cost, then follow the effect of each change month after month, against your own history.
With Sobria
These levers, without thinking about them.
Sobria Advisor, a router for claude.ai, recommends, before every request, a less costly model that, on our benchmark of 1,421 public business conversations, keeps almost all the quality, and shows the cost difference; nothing changes without the employee’s click. When a conversation grows too long, it suggests opening a fresh one.
Sobria Brain gives all your AI your company’s memory: it aims to answer most questions from its records, with far fewer tokens.
Sobria Report reconciles spend to your invoices every month and shows the estimated savings from model choice.
Frequently asked questions
The questions we are asked.
Should you always use the most powerful model?
No. For everyday requests, a mid-range model keeps, on our benchmark of 1,421 public business conversations, almost all the quality for far less; keep the most powerful one for long, difficult requests.
Is a spending cap enough to control AI costs?
No: it limits the bill without telling a trivial question from a legitimate heavy task, and it cuts both.
Why does a long conversation cost more?
Because the model rereads the whole conversation with every message: each new request is billed with everything that came before it.
Do these levers apply to a per-employee subscription?
Partly: on a subscription, the bill does not move with usage, but your AI’s quotas, governance and footprint still depend on it.