Claude's new cost tools make room for better AI experiments

A clearer picture of AI spending

On October 29, Anthropic announced an early test version of its Usage and Cost Analytics API for Claude Console and Claude Enterprise organisations. An API is a connection that lets software share information. This one lets a team's tools read its AI usage and spending reports.

The reports can break costs down by the AI system used, team workspace, connection key, and the person or software account behind a request. Anthropic says its earlier cost-reporting service and the listed older Enterprise reporting connections remain supported. Check the documentation and your account's access before planning around the new version.

Our view is that clearer spending reports can help you choose the next experiment. Pair them with a measure of what the feature actually improves for someone.

Read the announcement: Claude Platform release notes: October 29, 2026.

Give the experiment a useful job

Begin with work that repeats, such as copying information between tools or answering the same question. You might want a draft a colleague can check quickly, or a summary that helps them see which request needs attention.

First record how the current process works and how long it takes. Then try AI on a few examples that reflect the task. Count the time spent checking and correcting answers, along with any extra attempts the AI needs. You are learning about the whole job.

Use invented or approved information while practising. Before using real customer records, confirm the service settings, who can access them, and how the information will be handled.

Understand what starts a charge

If you are learning to build with AI, ask which actions use a paid AI service and when. Pressing a button, running a task in the background, and automatically trying again after a failure can each create a separate request. Ask for an explanation you can follow.

Choose a trial budget you are comfortable spending. Check the provider's spending controls and review actual use during the trial. A report showing what you spent is different from a limit that stops spending. Find out how failed requests, repeated attempts, and unusually large amounts of information are handled before inviting more people to use the feature.

1. Choose one task and describe a good result.
2. Set a trial budget and check the spending controls.
3. Compare a few examples using the same checks each time.
4. Decide whether to keep, change, or stop the trial based on cost and usefulness.

Let a small trial guide the next decision

A contained trial can show where AI saves effort, where people should take the lead, and what still needs a better answer. Finding a poor fit can also be useful: it helps you make the next decision sooner.

We expect measuring real work to remain a helpful skill as AI changes. That is our outlook. Connecting a bill to a task, and a task to an outcome, gives you a clearer starting point for training, choosing tools, or discussing a build with a team.