Guides

    Claude Token Costs: How Pricing per Token Works

    How Claude token cost works in 2026: price per million tokens per model, input vs output, caching, batch and four worked examples with dollars per month.

    November 16, 202615 min read
    Share:

    Published 16-11-2026. Last verified on 16-11-2026 against Anthropic's pricing page and token counting documentation. Prices change with almost every model launch, so we re-check this page monthly.

    Quick answer: what does a Claude token cost?

    A Claude token costs between $0.0000001 and $0.00004, depending on the model and on whether it is input or output. Anthropic prices per million tokens (MTok): on 16-11-2026, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, Claude Opus 5.5 $4 and $20, Claude Haiku 5.5 from $0.10 and $0.50 and Claude Fable 5.2 $8 and $40. This guide is for developers, product owners and finance leads who have to put a Claude API budget on paper for a team, and it ends with four worked examples you can redo with your own numbers. Our short version: the model you pick changes the bill by up to 80 times, caching and batching change it by another factor of two to ten, and the prompt you send matters more than the model for most teams.

    How we worked out the numbers

    Our method in three lines: every price comes from Anthropic's pricing page, read on 16-11-2026; every cost is our own arithmetic on those list prices, in US dollars, without taxes or negotiated discounts; and every example shows its assumptions so you can swap in your own.

    Our gate for this page: an example only counts if it states the model, the input tokens, the output tokens and the resulting cost. If a reader cannot redo the sum, we cut it. Two things fell out because of that rule. First, a "cost of a Claude Pro conversation" example: Pro has no published token allowance, so any number would be invented. Second, a cost per word or per page conversion: we found no Anthropic source for a words-per-token ratio on the current tokenizer, and the tokenizer changed in 2026, so we count in tokens only.

    The token counts in our examples are assumptions for typical jobs, not measurements from our own runs. We had no API access in this production run, so the counts are not tested by us. The section on counting tokens shows how to replace them with your real numbers in a few minutes.

    How does Claude token pricing work?

    We start with the unit, because everything else follows from it. Claude bills every API request in tokens, the small text pieces a model reads and writes. You pay for input tokens (everything you send: system prompt, tools, conversation history, documents) and for output tokens (everything Claude writes back, including thinking on models where thinking is on). The price per token is fixed per model and listed per million tokens.

    Three things make the bill bigger than most teams expect, and we would check all three before blaming the model:

    1. The whole conversation is input on every turn. First because it is the most common surprise: in a chat or an agent loop, each new request re-sends the history, so turn 20 pays for turns 1 to 19 again unless you cache.
    2. Tools and documents count as input. Second: tool definitions, file contents and images are all tokens. One MCP server with many tools can add thousands of tokens to every request.
    3. The tokenizer changed. Last, but it catches teams that compare prices across generations: Anthropic states that Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text. Sonnet 4.6 at $3 per million and Sonnet 5.5 at $2 per million are therefore closer in practice than the list prices suggest.

    Anthropic adds a few modifiers on top of the base price: prompt caching, the Batch API, a 1.1x multiplier for US-only inference on Claude 4.6 and later models, and a premium fast mode on some models. We cover caching and batching below because they are the two you control on every request. For the LLM ops collection we picked skills that track exactly these levers in production.

    Claude cost per token for every current model

    Here are the four current models on Anthropic's pricing page on 16-11-2026, per million tokens, ordered from the cheapest to the most expensive per output token:

    ModelInput per MTokOutput per MTokCache read per MTokBatch input and output per MTok
    Claude Haiku 5.5 (prompts up to 100K)$0.10$0.50$0.01$0.05 and $0.25
    Claude Haiku 5.5 (prompts over 100K)$0.50$2.50$0.05$0.25 and $1.25
    Claude Sonnet 5.5$2$10$0.10$1 and $5
    Claude Opus 5.5$4$20$0.20$2 and $10
    Claude Fable 5.2$8$40$0.20$4 and $20

    Why each sits where it does: Haiku 5.5 is first because, for prompts up to 100,000 tokens, it is twenty times cheaper than Sonnet 5.5 on both input and output; it is still a quarter of the Sonnet price once a prompt goes over 100,000 tokens, which is why it has two rows. Sonnet 5.5 is next and is the price we would use for most budgets. Opus 5.5 is exactly double Sonnet. Fable 5.2 is double Opus and only worth it for the hardest work. The cheapest per output token is not always the cheapest per job: a model that needs three attempts costs three times its list price. For independent cost-per-task data next to Anthropic's list, your agent can pull model cost per task from Artificial Analysis through Agensi Tools, from $0.13 per call (starting price read on 16-11-2026), with no Artificial Analysis account or API key.

    Older models such as Opus 4.8 ($5 and $25) and Sonnet 4.6 ($3 and $15) are still listed. Our Claude API pricing guide has the full table including fast mode and data residency.

    Why output tokens cost five times more

    Every current Claude model prices output at five times its input, according to the pricing page, and in our view this ratio explains most budget surprises. Writing a token takes more compute than reading one, and Anthropic passes that on. The practical consequence: for jobs that read a lot and write a little (classification, extraction, summaries), input decides the bill; for jobs that write a lot (drafting, code generation, long reasoning), output does.

    We would check one setting before anything else: thinking. On models where adaptive thinking is on, thinking tokens are output tokens. A higher effort level often means more thinking and a higher bill, so for routine jobs we would test the lowest effort that still passes your quality check.

    How many tokens do you get in Claude Pro?

    None that you can count: Claude Pro has no published token allowance. Anthropic's Help Center describes usage limits instead: paid plans have a session limit that resets every five hours and weekly limits that reset once a week. How far that goes depends on the length of your conversations, the model, the features you use and the effort level, and usage in claude.ai, Claude Code and Claude Desktop all counts toward the same limit.

    So "claude pro token limit" has no number, and we would be wary of any page that gives one. What you can see is your own usage under Settings, then Usage. If you hit the limits in Claude Code every week, the decision is between a bigger plan and the API, where every token is billed and visible. Our guide to Claude Code usage limits walks through that trade-off.

    How do you count tokens before you send?

    Use Anthropic's token counting endpoint; it is the first tool we would reach for. It accepts the same input as a normal message (system prompt, tools, images and PDFs) and returns the number of input tokens without running the model, according to the token counting docs. Two caveats from the same page: the count is an estimate that can differ by a small amount, and server tools, the MCP connector and files sent by URL are not supported, so for those you read the usage object in the real response.

    Recount when you change models. The docs say the newer tokenizer on Claude 4.7 and later produces about 30% more tokens for the same text, so a count measured on an older model is too low for today's lineup.

    Three skills make this routine, and we have ordered them by when you would use them:

    1. Prompt Preflight, free, checks a Claude prompt before it ships: input tokens, position in the context window and cost per call and per 1,000 calls. First because it runs before you spend anything.
    2. Claude API X-Ray, free, splits a real API response into cache read, cache write, uncached input and output and prices each part. Second because it checks what actually happened.
    3. MCP Token Budget, free, estimates how much context your MCP tool definitions take on every request. Third because it only matters once you use several MCP servers, but then it often explains a bill nobody can account for.

    Four worked examples

    Each example states our assumptions, so you can see exactly where we guessed. Prices are list prices on 16-11-2026, standard requests unless stated. These are our calculations on assumed token counts, not measured runs.

    1. A working chat session on Sonnet 5.5. Assumption: 20 turns in which the history is re-sent, adding up to 150,000 input tokens and 10,000 output tokens. Cost: 150,000 x $2 per million is $0.30, plus 10,000 x $10 per million is $0.10, so $0.40 per session. At five sessions a day for 22 working days (110 sessions), that is $44.00 per person per month. With prompt caching on the history (20,000 tokens written to the 5-minute cache at $2.50 per million, 130,000 read from cache at $0.10 per million, output unchanged), the session costs $0.163 and the month $17.93.

    2. A 50-file code review on Opus 5.5. Assumption: 50 files of 3,000 tokens plus 5,000 tokens of instructions, so 155,000 input tokens and 20,000 output tokens for the review. Cost: $0.62 input plus $0.40 output, $1.02 per review; 20 reviews a month is $20.40. On Sonnet 5.5 the same review costs $0.51. We would keep large or security-sensitive reviews on Opus and send routine pull requests to Sonnet. Skills for the review itself are in the code quality and review collection.

    3. Summarising a long document on Sonnet 5.5 and Haiku 5.5. Assumption: a report of 60,000 input tokens and a summary of 1,500 output tokens. Cost on Sonnet 5.5: $0.12 plus $0.015, so $0.135 per document and $27.00 for 200 documents a month. On Haiku 5.5: $0.006 plus $0.00075, so $0.00675 per document and $1.35 for 200. Whether Haiku is good enough is a quality question, not a price question; check 20 summaries by hand before you switch. More summarisation skills are in the research and analysis collection.

    4. 1,000 classifications on Haiku 5.5. Assumption: each item is 800 input tokens (instructions plus one support ticket) and 20 output tokens (a label), so 800,000 input and 20,000 output tokens per 1,000 items. Cost: $0.08 plus $0.01, so $0.09 per 1,000 items. Through the Batch API at half price: $0.045. At 30,000 items a month that is $2.70 standard or $1.35 in batch. The same job on Sonnet 5.5 costs $1.80 per 1,000 and $54.00 a month. Classification and tagging skills sit in the data and databases collection.

    What the four examples show, in our reading: the model choice moved the monthly cost by a factor of 20 (examples 3 and 4), caching by a factor of 2.5 (example 1) and batching by a factor of 2 (example 4). One trap for example 2: on Haiku 5.5 a 155,000-token prompt falls in the over-100K tier at $0.50 and $2.50, so long prompts cost five times more per token than short ones.

    To turn these examples into a budget for your own team, LLM Cost and Token Economics Modeler, $14, projects token costs at scale and builds an optimization plan. Budget owners will find more in the finance and accounting collection.

    How caching and batching change the bill

    Prompt caching cuts the cost of repeated input; the Batch API halves everything for work that can wait. We treat both as defaults, not as optimizations. Both stack, according to the pricing page.

    Caching works like this: writing to the 5-minute cache costs 1.25 times the base input price and pays for itself after one read; the 1-hour cache costs 2 times and pays for itself after two reads. Reads are cheap: 0.05 times the base input price on Opus 5.5 and Sonnet 5.5 ($0.20 and $0.10 per million) and 0.1 times on most other models, including Haiku 5.5 ($0.01 per million for short prompts). Our prompt caching guide explains where to put cache breakpoints.

    The Batch API gives a 50% discount on input and output for requests processed asynchronously. Fast mode is not available with batch. We would batch every job that does not need an answer within the hour: nightly classification, bulk summaries, evaluation runs. For pipelines like that, see the workflow automation collection.

    If caching is set up but the bill does not move, the cache is probably missing. LLM Prompt-Cache Auditor, $9, audits API code for the usual reasons, such as a timestamp at the top of the system prompt that changes every request.

    What is the cheapest setup for heavy use?

    Our ranking for a team that runs Claude all day, in the order we would apply the steps:

    1. Route each job to the cheapest model that passes a quality check. First because, in our examples, it moved costs by a factor of 20, more than any other lever.
    2. Cache everything that repeats. Second: system prompts, tool definitions and long documents are re-sent constantly, and cached reads cost 2.5% to 10% of the input price, depending on the model.
    3. Batch everything that can wait. Third: a flat 50% off, but only for work that does not need an immediate answer.
    4. Trim the prompt. Last because its savings are smaller per request, but they apply to every request: unused tools, old history and verbose instructions.

    For Claude Code users the same logic applies to subscription limits. Where Do Your Claude Code Tokens Go?, free, shows where tokens go and gives rules for the biggest drains, and Claude Code Token Audit, $1.99, breaks cost down by model, context size and sub-agents. Our articles on reducing Claude Code token usage and on skills that reduce Claude Code costs go further.

    Data jobs are a separate cost line. When your agent needs live data (search results, company data, SEO metrics), Agensi Tools bills per use from one balance, shows the exact price before a paid use runs and charges nothing for a use that fails, so you do not pay a monthly subscription to a data provider you use twice a week.

    Token pricing risks to watch

    Three risks we would put in front of any budget owner, because our own examples above depend on all three staying stable. Prices change: Opus dropped from $5 and $25 (Opus 5) to $4 and $20 (Opus 5.5) in two months, and Sonnet 5.5 cache reads were halved on 07-10-2026, so date every budget line. Token counts change with the model: the 30% tokenizer difference means a migration can raise token volume while lowering the price per token. And agent loops have no natural stop: an agent that retries without a limit can spend a day's budget in an hour. Agent Cost Scan, free, reads agent logs and reports cost per run, the model that eats the budget and three ranked fixes.

    Your next step

    We sell skills and pay-per-call tools that your agent uses through one MCP connection, and we name ours in the sections above. No vendor paid to be listed.

    Take one real job from last week, count its tokens with the token counting endpoint or Prompt Preflight, and redo the matching worked example with your numbers. That one sum tells you more than any price list. To run the free cost skills and Agensi Tools from your own agent, connect it once through Agensi MCP.

    Written by Samuel Rose, Founder of Agensi. Next scheduled review: 16-12-2026.