For months, FinOps teams managing Claude at scale have been asking the same question: "Who in our organization is actually driving this bill?"
Anthropic just answered it.
The new Claude Enterprise Analytics API gives Enterprise plan admins programmatic access to per-user cost, usage, and engagement data across every Claude surface — chat, Claude Code, Cowork, and Office agents. Nine endpoints. Named users. Dollar amounts. Broken down by model, context window, inference region, and speed.
This is the visibility layer that's been missing from enterprise AI adoption. And it's about time.
What is Anthropic's Enterprise Analytics API & Why This Matters
Let's put this in context. Anthropic has had a billing API since mid-2025 — the Admin Usage & Cost API. It tracks token consumption grouped by workspace, API key, model, and service tier. It's useful for developers and platform teams who manage API integrations.
But it doesn't answer the business question.
When your Anthropic invoice jumps 30% in a month, the Admin API can tell you it came from Workspace X using Opus 4.6. What it can't tell you is which human beings caused it, what they were doing, or whether it was worth it.
The Enterprise Analytics API changes that. It returns named users with email addresses, their individual token usage, their USD spend, and their engagement patterns — conversations, Claude Code sessions, commits, pull requests, lines of code, Cowork actions, skills used, connectors invoked. The full picture.
This is the difference between "our Claude bill went up" and "12 engineers in Platform drove 68% of the increase, primarily through Claude Code sessions in the 200K-1M context window."
One is a budget problem. The other is an actionable conversation.
What the API Actually Gives You
Nine endpoints split across two categories.
Engagement and adoption — 5 endpoints:
These return aggregated metrics per organization, per day, with up to 90 days of history (from January 1, 2026 onward):
User activity is the most detailed endpoint. For each user, you get conversation counts, messages sent, projects created, files uploaded, artifacts created, skills and connectors used. For Claude Code users, you get commits, pull requests, lines of code added and removed, and session counts. For Cowork, you get sessions, tool actions, dispatch turns, and skill/connector invocations. Even Office agent usage in Excel and PowerPoint is tracked.
Activity summary gives organization-wide DAU, WAU, and MAU counts across Claude, Claude Code, and Cowork — plus seat utilization and pending invite counts. This is your adoption scorecard in a single API call.
Chat project usage breaks down conversations and users by project. If you've organized your enterprise around projects (which you should), this tells you which ones are actually getting used.
Skill usage and Connector usage show which skills and MCP connectors your organization uses, with per-surface breakdowns. If you deployed custom skills or connected Jira, Slack, GitHub, or other tools — now you can measure whether anyone is actually using them.
Usage and cost — 4 endpoints (beta)
This is where the FinOps value gets serious.
Per-user token usage ranks users by token consumption over a date range. You can break it down by product (Claude vs Claude Code vs Cowork), model (Opus vs Sonnet vs Haiku), context window (0-200K vs 200K-1M), inference region, or speed (standard vs fast).
Per-user cost does the same, but in USD. Break down by product, model, cost type (tokens, web search, code execution), or token type.
Token usage over time gives you time-series data bucketed by minute, hour, or day — with all the same breakdowns. This is your trending dashboard.
Cost over time mirrors token usage over time, but in dollars. Add cost type and token type groupings for full granularity.
The Context Window Dimension Is the Sleeper Hit
Every usage and cost endpoint supports breaking down by context_window, a binary split between 0-200K and 200K-1M.
Here's why this matters more than it looks.
In March 2026, Anthropic removed the 2x pricing premium that previously applied when requests crossed 200K input tokens. Everyone celebrated. But the removal of that pricing cliff also removed the last natural guardrail on context size. The per-token rate is now flat across the full million — but more tokens still means a bigger bill.
With the Enterprise Analytics API, you can now track which users are regularly operating in the 200K-1M window. Not at the API key level — at the person level. If a developer's Claude Code sessions consistently hit the high-context tier, you can have a targeted conversation about whether that context is necessary, or whether prompt optimization could reduce costs without impacting productivity.
The pricing guardrail is gone. The visibility guardrail just arrived.
The Claude Code Blind Spot — Partially Solved
Claude Code has been one of the hardest AI cost drivers to attribute. It runs locally on developer machines, consumes API tokens in the background, and until now, the only way to track costs was through API key grouping — which tells you almost nothing about who did what.
The Enterprise Analytics API brings Claude Code into the light. Per-user metrics include session counts, commits, pull requests, and lines of code changed. Combined with per-user cost data, you can calculate meaningful efficiency metrics: cost per commit, cost per pull request, cost per session.
This matters because Claude Code usage is inherently variable. A senior engineer debugging a complex system might burn through 500K tokens in a session and ship a critical fix. A junior developer experimenting might burn the same amount and produce nothing mergeable. Both show up the same in the Admin API. The Enterprise Analytics API lets you tell the difference.
One caveat: if your organization uses Claude Code via Amazon Bedrock, that data won't appear in the Analytics API. Anthropic calls this out explicitly in their documentation. For Bedrock-routed Claude Code, you'll still need to rely on AWS cost data.
What's Not in the API (Yet):
A few gaps worth noting.
No per-request granularity. The usage and cost endpoints aggregate across time buckets (minute/hour/day) and user-level rollups. You won't see individual API calls or specific conversations. If you need request-level detail, the Compliance API is the separate tool for that.
Data freshness has limits. Cost and usage data refreshes every four hours and may take up to 24 hours. Values can be revised for up to 30 days as late events reconcile. Anthropic recommends querying dates at least 30 days in the past for invoicing-grade accuracy.
Engagement data is delayed by three days. The activity and adoption endpoints aggregate data the day after it occurs, and it becomes queryable three days later. Real-time adoption dashboards aren't possible with this API.
Enterprise-only. This is not available on Team or Pro plans. You need an Enterprise plan and Primary Owner access to generate API keys.
Cost endpoints are in beta. Expect potential schema changes. Build defensively.
The Bigger Pattern: AI Providers Are Speed-Running the Cloud Billing Playboo
Step back and look at what's happened in just 12 months.
Anthropic launched the Admin Usage & Cost API in mid-2025 — basic token and cost tracking by workspace and API key. In March 2026, they removed the long-context pricing premium. Days later, they released the Enterprise Analytics API with per-user cost attribution, engagement metrics, and multi-surface tracking.
That's the same arc AWS billing took from 2012 to 2020, basic CUR, then tags, then linked accounts, then granular usage types, then Savings Plans visibility, compressed into about a year.
OpenAI is on a similar trajectory. Google is getting there. The direction is the same everywhere: model providers understand that enterprise adoption requires enterprise-grade cost visibility. You can't scale AI spend to millions of dollars a year and expect a monthly PDF invoice to be sufficient.
The question for FinOps teams isn't whether to integrate these APIs. It's how fast you can operationalize them into the same workflows you already run for cloud.
A Practical Playbook:
If you're a FinOps practitioner or platform engineer responsible for Claude costs, here's how to put this to work.
Week 1: Get the API key and pull per-user cost data. Use the per-user cost endpoint for the last 30 days. Sort by spend. Identify your top 15 users. This is your "who to talk to" list.
Week 2: Cross-reference with context window and model breakdowns. For your top spenders, break down by context window and model. Are they using Opus when Sonnet would suffice? Are they consistently in the 200K-1M context window?
Week 3: Build an adoption scorecard. Pull the activity summary endpoint. Map active users against seat count. Calculate utilization rate by surface (Claude, Claude Code, Cowork). Feed this into your monthly review.
Week 4: Set up weekly automated pulls. Use the time-series endpoints (cost over time, token usage over time) bucketed by day. Build a simple dashboard. Set alerts for sustained increases above your baseline.
Ongoing: Combine with cloud cost data. AI costs are infrastructure costs. They belong in the same dashboard, the same allocation framework, and the same governance process as your AWS, GCP, Azure, Kubernetes, and SaaS spend.
Finally
The Enterprise Analytics API isn't just a nice-to-have feature update. It's a signal that Anthropic takes enterprise cost governance seriously — and that the era of "check the invoice and hope for the best" is ending.
Per-user attribution. Multi-surface tracking. Context window visibility. Model and speed breakdowns. Engagement and adoption metrics alongside cost data.
This is what FinOps for AI is supposed to look like. The data is finally there. Now it's on us to use it.