What does Claude cost a business to run?

Ask what does Claude cost and you get a list of token prices with nothing to compare them to. A managing director needs the price of a support assistant, a document workflow or a coding agent, and that price is set by how the system is built.

All prices are from Anthropic’s pricing page and API pricing documentation, as of November 2026, in US dollars. The worked examples are arithmetic from those prices, with the assumptions stated.

What does Claude cost: seats for people, tokens for systems

Most businesses pay in two ways at once. Employees who work in the Claude app (chat, documents, Cowork, Claude Code) need a seat: a fixed price per person per month, with usage limits per seat. Software that calls Claude on its own, such as a support tool or your own agent, uses the API and pays for every token it reads and writes. A token is a piece of a word; a page of business text is several hundred tokens. If a person types and reads, buy a seat. If a system runs without a person, use the API.

Seat types can be mixed, and Team includes up to $500 a month in pooled API credits. Enterprise works the other way round: a cheap seat plus usage at API rates, so each person’s cost follows what that person does. How Cowork fits into daily team work is in our article on Claude Cowork for business.

Output costs five times input on every model. Haiku 5.5 prompts over 100,000 tokens cost $0.50 and $2.50. Older models remain available, some at higher prices than their successors.

What drives the cost in practice

Six decisions move the bill further than the price per token does.

The model is the largest one. Haiku 5.5 and Fable 5.1 differ by a factor of 100 on input, and sorting tickets or extracting fields from a form rarely needs the strongest model. Pick the model per step and test the choice on real cases.

Context size comes next. Every request resends instructions, documents and conversation history, and you pay for all of it every time. An assistant that attaches a 20-page handbook to each question pays for 20 pages per question.

Caching serves text resent within minutes from a cache at 10% of the input price, or 5% on Opus 5.5 and Sonnet 5.5. Writing it once costs 25% more than normal input. Anything with a fixed instruction block should cache, and we turn caching on before we change anything else.

Work that can wait runs through the Batch API at half price, with results arriving later instead of within seconds.

Agents add a multiplier of their own. An agent plans, calls a tool, reads the result and calls Claude again, often dozens of times per task, and it resends a growing context on every call.

Parallel agents each carry their own context, so four agents cost roughly four times one and finish sooner. When that trade is worth it is in our guide to Claude Code agent teams.

Smaller line items: web search at $10 per 1,000 searches plus result tokens, a 1.1 factor for US-only processing, and $0.08 per running hour for agents hosted by Anthropic (Managed Agents).

Speed versus price: fast mode

Fast mode runs the same Opus model at the same quality, configured for speed. Anthropic’s fast mode documentation says it writes output up to 2.5 times faster; the wait for the first word stays the same.

Fast mode costs twice the standard price. When it launched on Opus 4.6 in February 2026, it cost six times as much: $30 and $150. It no longer runs on Opus 4.6, and on Opus 4.8 it ends on November 30, 2026. It does not exist for Sonnet or Haiku or on the Batch API, and in Claude Code subscriptions it is paid from usage credits only.

Pay for it when a person waits for each answer and their time is worth more than the surcharge. Take one Opus 5.5 turn with 100,000 cached context tokens and a 2,000-token answer: about $0.06 at standard speed, $0.12 in fast mode. If the answer takes 40 seconds at standard speed, fast mode saves up to 24 seconds, about $9 per hour of waiting removed. That is cheap during live debugging and wasted on a nightly report nobody watches.

Fast and standard speed keep separate caches, so turning fast mode on in the middle of a session pays the full fast price for the whole context once. Choose the speed when the session starts.

Worked examples for typical company uses

Each example is arithmetic from the published prices. Replace the assumptions with yours.

A support assistant

3,000 conversations a month, four turns each. Every turn sends 20,000 tokens of fixed instructions and product knowledge plus 2,000 tokens of conversation, and returns 400 tokens. On Sonnet 5.5 without caching, that is about $576 a month. With the fixed block cached, it drops to about $120, plus a few cents whenever the cache is rewritten after a quiet spell. On Haiku 5.5 with caching it is under $10, if its answers hold up on your real tickets.

A document workflow

2,000 invoices or contracts a month, each 10,000 tokens plus 1,500 tokens of instructions, returning 800 tokens of structured data, run overnight through the Batch API. Sonnet 5.5 costs about $31 a month, or $62 without batch. Opus 5.5 costs about $62. At this volume the model hardly shows on the bill. The money goes into building a workflow that reads the right fields and flags what it is unsure of.

A coding agent

One task takes 40 steps with an average context of 80,000 tokens, 72,000 of them cached, and 1,500 tokens of output per step, on Opus 5.5. That comes to about $3.40 per task with caching and $14 without. Twenty tasks a month cost about $68 per developer, or $280 uncached. Fast mode doubles it, and an agent team of four multiplies it by about four. For steady daily use, a premium Team seat at $100 a month is the more predictable choice.

How to budget and cap it

Anthropic documents its hard limits on the rate limits page. An organization sets its own monthly spend limit on the Console billing page. Once it is reached, requests are refused until the next month or until someone raises the limit. Each usage tier also has a ceiling: $500 a month on Start, $1,000 on Build, $200,000 on Scale.

Below that, each workspace gets its own spend and rate limits. Give every project its own workspace, and a runaway test cannot drain the support assistant’s budget. On Team and Enterprise, admins decide whether usage beyond the plan, fast mode included, is allowed, and Enterprise adds admin spend limits.

What we recommend: run a pilot on real data and measure what each completed task costs. Set the workspace limit at two to three times that monthly figure. Turn on caching and batch before you consider a cheaper model. Review usage monthly; when an agent’s cost per task keeps rising, its context is almost always growing.

What Claude does not cost

• You never buy the model or its weights, so there is no license and nothing to depreciate. You pay for use.
• There is no hardware to buy or run: no servers, graphics cards or power bill for computing.
• Claude 4.6 and later models include the full 1M token context window at standard prices, so long documents carry no surcharge. Haiku 5.5 above 100,000 tokens is the exception.
• API keys and projects carry no fee. You pay for tokens and listed extras such as web search.

Building the system, testing it and paying the people who check its output appear on no price list, and they belong in the budget from the start.

What this means for a business

Most surprises on a Claude bill come from tasks nobody measured. The same support volume costs under $10 or $576 a month depending on the model, the caching and how much text each request carries, and all three are decided when the system is designed.

Define the task, pilot it on real data, measure what each completed task costs, then cap it. Give seats to the people who work with Claude daily. Run systems that work on their own through the API, each in its own workspace with its own limit.

Working that out for your processes, with numbers you can check, is the first conversation in our AI consulting.