Anthropic gives API teams a lot to work with: usage reports by workspace, key and model, spend limits for each workspace, and several ways to pay less for the same work. What your team adds is the business view, meaning which team each dollar belongs to and what the work produced.
We suggest you give each dollar an owner first, then choose the levers that fit each workload, and then check each change against the bill and the result. Prices change, so the ones here are from Anthropic's pricing page in October 2026.
What Anthropic already gives you
The Claude Console has a Usage page and a Cost page. The Usage and Cost API returns the same data to an Admin API key, usually within five minutes of each request.
- Usage comes in buckets of a minute, an hour or a day. You can group and filter it by API key, workspace, model, service tier, context window, inference region and speed. It counts uncached input, cache writes, cache reads and output separately, and it counts web searches.
- Cost comes in US dollars, by day. You can group it by workspace or by a description that names the model and the region.
- Controls. Each workspace can have its own monthly spend limit with alerts, and its own rate limits. An organization can have up to 100 workspaces by default.
What the reports leave out
The cost report groups by workspace, not by API key, so a cost for each key means multiplying its tokens by each model's price. The Default Workspace shows no ID in either report, and you cannot set limits on it. Requests from the Console's playground carry no API key. And if your organization still has a Priority Tier commitment, its costs are left out of the cost report, so you read them from the usage report instead.
No report names the team behind a workspace, and none can say whether a request resolved a ticket or finished a document. That part is yours to add, and it is the part Valistry is built for: it maps each workspace and key to a team, and joins the spend to results from your own systems.
Give each dollar an owner
On the Claude API, the workspace is where an owner is recorded, so set workspaces up with care.
- One workspace for each team or product, named so that finance can read it, for example "Support assistant, production".
- Keep the Default Workspace empty. Its usage arrives with no ID, and it cannot take a spend limit.
- Separate development and testing into their own workspaces, so experiments do not mix with production spend and can run under a lower limit.
- Use service accounts for production systems. A personal key stops working when its owner leaves the workspace, and a service account key does not.
- Map each workspace to a team once, and keep that list beside your budget. Prompt caches are also kept apart by workspace, so a workspace for each product helps caching too.
The levers Anthropic offers, and what each saves
- Model choice
- Sonnet 5.5 for most production workHalf Opus 5.5's input and output price
- Haiku 4.5 for simple tasksA quarter of Opus 5.5's price
- Newer tokenizer, from Claude 4.7About 30% more tokens for the same text
- Caching
- Prompt cachingA cache read costs 10% of input, 5% on Opus 5.5
- Stable content firstNo published figure
- One-hour cache lifetimeWrite costs 2x input; pays back after two reads
- How calls run
- Batch API for work that can wait50% less than standard
- Batch and caching togetherThe two discounts stack
- Where calls run
- Global routing, the defaultUS-only inference costs 1.1x
- Tools
- Web search only where needed$10 per 1,000 searches
- Unused tool definitions removedNo published figure
- Speed
- Faster Opus 5.5 outputCosts 2x standard
Each figure is a list price or Anthropic's own guide. Your bill shows your own.
Model choice. On the pricing page, Anthropic suggests Haiku for simple tasks, Sonnet for most production work, and Opus for the most complex reasoning.
- Haiku 4.5
- $1
- $5
- $0.10
- Sonnet 5.5
- $2
- $10
- $0.20
- Opus 5.5
- $4
- $20
- $0.20
| Model | Input | Output | Cache read |
|---|---|---|---|
| Haiku 4.5 | $1 | $5 | $0.10 |
| Sonnet 5.5 | $2 | $10 | $0.20 |
| Opus 5.5 | $4 | $20 | $0.20 |
Per million tokens. The Batch API charges half of each.
There is one thing to watch when you move to a newer model. Claude 4.7 and later models use a newer tokenizer, which Anthropic says produces about 30% more tokens for the same text. So a model with a lower price per token can still cost more per task. Anthropic's token counting endpoint is free, and its own advice is to count the same request on both models before you move.
Caching. A cache read costs 10% of the input price, 5% on Opus 5.5. A five-minute cache write costs 1.25 times input and pays back after one read; a one-hour write costs twice input and pays back after two. Adding one cache_control field to a request turns on automatic caching.
- The cache is built in order: tools, then the system prompt, then the messages. A change at one level clears that level and everything after it, so stable content goes first and changing content goes last.
- Changing the tool definitions clears the whole cache.
- Each model has a minimum length, from 512 to 4,096 tokens. A shorter prompt is not cached, and the API returns no error, so check
cache_read_input_tokensin the response to see it working.
How calls run. The Batch API costs half the standard price for work that can wait. Most batches finish within an hour, and a request that has not run within 24 hours expires and is not billed. Batch and caching discounts stack, though cache hits in a batch are best effort; Anthropic reports hit rates of 30% to 98%, and suggests the one-hour cache for batches that share context.
Where calls run. The Claude API routes globally by default. Keeping inference in the US with inference_geo costs 1.1 times the standard price on Claude 4.6 and later models. On Amazon Bedrock and Google Cloud, regional endpoints cost 10% more than global ones. We suggest choosing a region only where a contract or a law requires it.
Tools. Web search costs $10 per 1,000 searches, on top of the tokens its results add. Web fetch adds no fee beyond tokens. Code execution includes 1,550 free hours a month for each organization, then $0.05 an hour for each container, and it is free when web search or web fetch is in the same request. Every tool definition is also sent as input tokens on each call, so a long list of unused tools costs a little on every request.
What costs more. Fast mode on Opus 5.5 costs $8 per million input tokens and $40 per million output tokens, twice the standard price. It cannot run in a batch, and it suits work where a person is waiting. Long context is not a cost on recent models: Claude 4.6 and later models charge the same rate per token across the whole one-million-token window.
How to know a change worked
A change has worked when the same workload produces the same results for less, with no drop in quality. You can measure that in four steps.
- Pick one workload and one change, for example moving a document summary job into the Batch API.
- Write down the baseline over 30 days: the workload's cost from its workspace in the cost report, and the number of finished tasks from your own system, such as documents processed or tickets resolved. Divide the cost by the finished tasks.
- Make the change, and leave everything else as it was.
- Measure the same workload over the same number of days, and compare the cost per finished task.
Keep your quality checks running next to the cost. If answers need more retries or more human review after a change, the extra time people spend can cancel the saving on the bill.
Compare models on what a finished task costs, not on the price per token.
For caching, the usage report gives you a direct check. Compare cache reads with cache writes for the workspace each day: writes that are seldom read cost more than they save, and the fix is usually to move the changing part of the prompt to the end.
What Valistry adds
Valistry reads Anthropic's usage and cost reports with a read-only key, and never reads a prompt or an output. It puts Claude API spend beside every other AI vendor and seat, and shows where a cheaper model or caching would save, checked on your own runs. The Anthropic page lists what it reads.
The teams that get the most from the Claude API usually know which team owns each workspace, which levers fit each workload, and what one finished task costs. The data for all three is already in Anthropic's reports and in your own systems.