Budgeting Claude API costs for a small team: model rates, caching, batch discounts, and per-person spend tracking

A common first budget is a rate per million tokens, multiplied by the tokens you expect to send and receive. That gives a rough number. A small team's actual bill also depends on cache writes, batch routing, US-only inference, and a tokenizer change that makes the same text cost more tokens. Budget for those, and your estimate is more likely to match the invoice.

Base rates per MTok

Anthropic's pricing page lists these base prices per million tokens (MTok):

The pricing page positions Haiku 5.5 for high-volume, latency-sensitive work such as classification, extraction, and routing. Its output rate for prompts over 100,000 tokens is not clearly labelled on the pricing page, so confirm it on the live table before you budget long prompts with Haiku 5.5.

Split every workload into four counts: input tokens, output tokens, cache writes, and cache reads. A single blended price hides the parts that move the total.

Caching pays only when the prefix repeats

A 5-minute cache write costs 1.25x the base input price. A 1-hour write costs 2x. A cache read costs 0.1x the base input price, or 0.05x for Opus 5.5 and Sonnet 5.5, and 0.025x for Fable 5.1 and Claude Mythos 5.1 (pricing page).

The break-even math uses the generic 0.1x read rate:

A 1-hour cache with only one read costs 2.1 units, which is more than the 2.0 units an uncached prompt used twice would cost. Caching helps when many people or many calls hit the same system prompt or reference document inside the cache window. It does little for a one-off prompt. Before you pick the 1-hour cache, check how often the same prefix recurs within an hour in your logs.

Batch and fast mode

The Batch API gives a 50% discount on input and output tokens (pricing page). Fable 5.1 batch prices are listed at $5 input and $25 output per MTok. Batch suits work nobody is waiting on, such as overnight summaries, bulk tagging, or backfilling a content library.

Fast mode is a research preview for Opus 5.5, Opus 5, and Opus 4.8 at premium prices. Opus 5.5 fast mode is $8 input and $40 output per MTok, which is twice the standard Opus 5.5 rate. Fast mode is not available with the Batch API. Because it is a preview, put it on its own budget line so you can remove it without rebuilding the plan.

US-only inference and combined multipliers

For Claude 4.6 and later models, setting inference_geo to 'us' adds a 1.1x multiplier to all token pricing categories. The default 'global' routing uses standard pricing. The pricing page says caching multipliers stack with the Batch discount and with data residency, and that fast mode stacks with caching and residency. Fast mode is not available with Batch, so those two never apply to the same request.

The Sonnet 5.5 example below uses 1 million input tokens and 200,000 output tokens a month:

The table leaves out caching because the cache hit rate depends on the workload. The caching break-even table above covers that case.

The last two rows use a planning buffer, not a rate from the pricing page. The Batch-plus-US-only row assumes the Batch discount and residency multiplier apply together, which the page does not state directly. Neither the stacked rows nor the buffer rows have been checked against a real invoice, so run one small month and compare the result with the invoice before you rely on them.

Budget for the tokenizer change

Claude 4.7 and later models, and Claude Mythos Preview, use a newer tokenizer. It produces approximately 30% more tokens for the same text. The exact increase depends on content and workload (pricing page).

If your baseline came from earlier models, it will undercount. As a planning assumption, multiply the token count by 1.3 for 4.7-and-later models until you have measured your own ratio. Replace that 1.3 with your measured figure as soon as you have one.

A monthly worksheet

Build the estimate one workflow at a time. Record these fields for each row:

• Model and workflow
• Calls per month, from logs
• Average input and output tokens per call, from logs
• Monthly input and output tokens, calculated as calls multiplied by the per-call averages
• Tokenizer buffer, if the model is 4.7 or later
• Share of calls that use caching, and the cache window
• Share of work that can go through Batch
• Residency setting
• Whether fast mode is in use

The row below is a placeholder. Replace its numbers with your own logs:

For a team of 5 to 50 people, group rows by workflow first, then by person. Add a contingency line for work that does not fit a known workflow. Choose the contingency size yourself and state it as an assumption in the budget document. For the wider budget-setting steps, see how to set an AI budget for a small company.

Long prompts over 100,000 tokens

Haiku 5.5 changes tier at 100,000 prompt tokens. Input is $0.10 per MTok up to that size and $0.50 above it. Output is $0.50 per MTok up to that size. The output rate above that size is not clearly labelled on the pricing page, so treat that tier's output cost as unknown until you have checked it. If your workload regularly sends prompts longer than 100,000 tokens, model that tier as its own line.

Usage attribution is not the same as cost

Budget reviews often mix two questions. Usage attribution answers who or what consumed tokens: which person, project, or model. Monetary cost reporting answers what those tokens are worth at the applicable rates. Keep them in separate columns. A team can know who used the most tokens and still be unable to state the cost if some calls have no price.

Spendly describes views by provider, model, and person, along with budgets, alerts, and forecasts for projects (Spendly). Test those views against your own data before you rely on them. Spendly also says it shows pricing coverage, labels estimates, counts unpriced operations separately, and that the provider invoice remains the final billing source. Treat any dollar figure in a dashboard as an estimate until it matches the invoice, and track the share of calls that have pricing next to the spend total. To set spend warnings, see AI cost alerts: thresholds that actually work.

Log six fields for each call: provider, model, team, project, usage, and spend.

Per-person example

The figures below are hypothetical. They use the confirmed Sonnet 5.5 and Opus 5.5 base rates, with no multipliers applied. The example shows three people. A ten-person team would add one row per person in the same format:

To adapt this example, replace the person and project rows with your team's names and logged usage. To roll those rows up by project, see how to allocate AI costs by project.

Measuring a change

When you change a model, turn on caching, or move a workload to Batch, compare before and after on real runs. Spendly saves a baseline, waits for 10 completed post-change results with at least 80% pricing coverage, and then reports the result. Spendly says it never switches a provider or model for you, so your team makes the change and the tool measures the outcome (Spendly). A spreadsheet can follow the same discipline: do not judge a change on two or three runs.

Is a tool worth it at your size?

Spendly lists team plans by tracked employee count (Spendly). Starter is €29 a month for up to 10 tracked employees, Growth is €79 for up to 50, Business is €199 for up to 200, and Enterprise starts at €499 for 201 or more. Prices exclude applicable taxes. Admins and view-only users do not count toward the limit.

Compare the plan price with the Claude spend it would help you control. If the monthly bill is small, the worksheet above may be enough on its own. If several people across several projects share the budget, per-person tracking is harder to keep by hand.

FAQ

How much does the Claude API cost per million tokens?

It depends on the model. Sonnet 5.5 is $2 input and $10 output per MTok. Opus 5.5 is $4 and $20. Fable 5.1 is $10 and $50. Haiku 5.5 is $0.10 input and $0.50 output up to 100,000 prompt tokens.

Does prompt caching save money on the Claude API?

Only when the same prefix is reused. A 5-minute cache write pays off after one read, and a 1-hour write pays off after two reads.

How much does the Batch API discount Claude?

It gives 50% off input and output tokens. Fast mode is not available with Batch.

What does inference_geo set to 'us' cost?

For Claude 4.6 and later models, it adds a 1.1x multiplier to all token pricing categories. The default 'global' routing uses standard pricing.

Do Claude 4.7 and later models cost more for the same text?

The pricing page lists Opus 4.7 and Opus 4.6 at the same per-token rate ($5 input, $25 output). Claude 4.7 and later models use a tokenizer that produces approximately 30% more tokens for the same text, so the cost for the same text can rise.

How do I budget Claude API spend per person on a small team?

Record usage for each call with its person, project, and model. Build the estimate per model first, then roll it up by person and project.

Do the caching, Batch, residency, and fast mode multipliers combine?

Not all four at once. The pricing page says caching multipliers stack with Batch and with data residency, and that fast mode stacks with caching and residency. Fast mode is not available with Batch. The page does not state Batch and residency together, so check that combination against one month of real spend.

Is the provider invoice the final billing source for Claude API costs?

Spendly says the provider invoice remains the final billing source (Spendly). That is Spendly's statement, and this article does not extend it to Claude API charges. For Claude API spend, check your usage against the invoice you receive from the provider, and treat any dashboard figure as an estimate until the two match.

Keep the numbers current

Rates change, and the lineup of models changes too. Re-check every rate on the live pricing page before each budget cycle. Fast mode may end or change price. This article should be reviewed by 2026-11-15. The rates here were checked on 2026-10-15.

If you want Claude spend by model and by person against your own budget, Spendly's 7-day trial needs no card to start (Spendly). Run the worksheet first, then check whether the numbers you collected match what the tool shows.