33  New Analytics and Cost Controls for Claude Enterprise

33.1 Overview

Published July 2, 2026, this announcement marked the most significant expansion of Claude Enterprise’s administration tooling since the product’s launch. The update introduced richer admin analytics, model-level entitlements, spend-threshold alerts, and a programmatic Admin API — giving organizations the granular visibility and control they needed to scale Claude deployments across thousands of users with confidence.

The core problem the update addressed was straightforward: as Claude adoption grew inside enterprises, administrators needed to answer two questions that legacy dashboards could not handle well. Where is the money going? And are we getting value for it? A Fortune 500 company might have 15,000 employees using Claude — some writing marketing copy, some running Claude Code for eight hours a day, some doing one-off research. Without fine-grained analytics, cost allocation was guesswork, and proving ROI to a CFO was nearly impossible.

The July update solved this with three pillars: deeper analytics (usage and cost broken down by group and user, plus dedicated Claude Code insights), intelligent cost controls (model defaults, entitlements, and multi-tier spend alerts), and programmatic access (an Analytics API for exporting data into existing FinOps tools, and an Admin API for scripting cost-control workflows).

33.2 The Analytics Overhaul

33.2.1 Usage and Cost by Group and User

The new analytics dashboard moved beyond aggregate token counts. Administrators could now see usage and cost broken down along two axes simultaneously: by organizational group (team, department, or any custom grouping defined in the identity provider) and by individual user.

Dimension	What Admins See	Use Case
Organization (aggregate)	Total spend, token volume, trend lines	Budget tracking, forecasting
Group	Spend per team, model mix, session counts	Chargeback, capacity planning
User	Individual usage, model usage, active days	Identify power users, anomalies

This granularity solved a common pain point: the ability to do chargeback — billing individual departments for their actual Claude usage rather than splitting a single invoice evenly. A company could now tell its engineering org, “You consumed 62% of the Claude budget this month at $X,” with the data to back it up.

The dashboards also surfaced trend data. Rather than a single point-in-time snapshot, admins could view 7-day, 30-day, and 90-day trends, making it possible to spot anomalies — a sudden spike in Opus usage from a team that had previously only used Sonnet, for example — before they became budget problems.

33.2.2 Claude Code Insights: Two Dedicated Tabs

Claude Code, as the most token-intensive surface, received its own dedicated analytics with two tabs: Usage and Value.

The Usage tab answered the operational questions:

• How many developers are actively using Claude Code?
• How many sessions are running per day?
• What commands are being executed most frequently?
• What is the average session length?

These metrics let engineering managers understand adoption patterns. If 80% of sessions came from 20% of developers, that signaled a training gap. If session lengths were consistently short, it might indicate that developers were giving up too early rather than letting the agent complete complex tasks.

The Value tab was the more strategically important addition. It attempted to quantify the return on Claude Code spend:

Metric	What It Measures	Why It Matters
Productivity lift	Estimated time saved vs. baseline	ROI for leadership
Cost per commit	Token cost divided by commits accepted	Unit economics
Annual value	Extrapolated yearly savings	Budget justification

The “annual value” figure was particularly powerful for the conversations administrators needed to have with finance teams. Instead of arguing abstractly about productivity, an admin could point to a concrete number: “Claude Code is delivering an estimated $4.2M in annual value against $180K in spend.”

Note

The Value tab metrics were estimates based on models of developer productivity, not direct measurement of individual output. Anthropic was transparent that these were directional figures intended for internal planning, not for performance reviews of individual developers.

33.3 Analytics Chat: Natural Language Queries

Perhaps the most novel feature was Analytics Chat — the ability to ask questions about usage and cost in natural language and receive charts and tables as answers.

An administrator could type:

Show me Opus 4.7 spend by team for the last 30 days, sorted by cost.

or

Which users had the highest cost per commit last week?

The system would parse the query, execute it against the underlying analytics data, and render the result as an interactive chart or sortable table. This was a deliberate bet on lowering the barrier to insight. Not every administrator wants to learn a query language or navigate a complex dashboard. Natural language access meant that a finance analyst could pull the data they needed without filing a ticket with IT.

Common query patterns the system was optimized for included:

• Time-bounded spend: “How much did we spend on Haiku last month?”
• Comparative analysis: “Compare Engineering vs. Product usage.”
• Anomaly detection: “Show me any users whose spend jumped more than 50% week-over-week.”
• Forecasting: “Project our spend for the next quarter based on current trends.”

Tip

Analytics Chat was not a general-purpose AI assistant — it was constrained to querying your organization’s own analytics data. This scoping made it reliable for the specific domain it covered, and avoided the hallucination risks of open-ended chat.

33.4 The Analytics API

For organizations that wanted to integrate Claude analytics into their existing FinOps or observability stacks, Anthropic introduced the Analytics API. This REST API exposed the same data available in the dashboards, formatted for machine consumption.

The API was designed with first-class integrations for popular FinOps platforms:

Platform	Integration Type	What It Provides
Datadog	Native integration	Dashboards, alerts, anomaly detection
CloudZero	Native integration	Cost allocation, unit economics
Custom	REST API	Pull data into any BI tool

A typical integration with Datadog, for example, would let a DevOps team create a dashboard that showed Claude spend alongside AWS infrastructure costs, application performance metrics, and incident timelines — providing a unified operational view.

# Example: Query organization spend via the Analytics API
curl -X GET "https://api.anthropic.com/v1/analytics/usage" \
  -H "Authorization: Bearer $ADMIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "start_date": "2026-06-01",
    "end_date": "2026-06-30",
    "group_by": ["group", "model"]
  }'

The API supported filtering by date range, group, user, model, and surface (Chat, Code, Cowork), and returned data in JSON or CSV format for easy ingestion into downstream tools.

33.5 Model Defaults and Entitlements

The second pillar of the update was model-level control. Previously, once a user had access to Claude Enterprise, they could typically use any model the organization was entitled to. That created a risk: a user who didn’t know the difference between Opus and Haiku might rack up significant costs by defaulting to the most expensive model.

The new model defaults feature let administrators set, at the org or group level, which model users would get by default. A cost-conscious organization could set Haiku as the default for general Chat use while allowing Claude Code sessions to default to Sonnet for better coding performance.

Model entitlements went further, allowing admins to restrict which models a given group could access at all. This enabled tiered access strategies:

Tier	Entitlement	Rationale
Executive	All models	Maximum capability for high-stakes work
Engineering	Opus, Sonnet	Code requires frontier models
General staff	Sonnet, Haiku	Cost-effective for everyday tasks
Contractors	Haiku only	Minimize spend on external users

# Example model entitlement policy (Admin API)
model_policy:
  defaults:
    org_wide: claude-sonnet-4-6
    groups:
      engineering: claude-opus-4-7
      customer_support: claude-haiku-4-5
  entitlements:
    executives:
      allowed: [claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5]
    contractors:
      allowed: [claude-haiku-4-5]

33.6 Spend-Threshold Alerts

The third pillar was proactive spend alerts — a multi-tier system designed to catch cost overruns before they happened, rather than after.

The alert system operated at two levels with distinct thresholds:

Level	First Alert	Second Alert	Action
Organization	75% of monthly budget	90% of monthly budget	Notify admins; optional auto-restrict
User	75% of user budget	95% of user budget	Notify user + admin

The two-tier structure reflected different failure modes. An organization hitting 90% of its monthly budget on the 15th was a systemic problem requiring admin intervention. A single user hitting 95% of their individual cap might be doing legitimate intensive work — but it warranted a check-in.

Note

The 90% organization threshold could optionally trigger automatic restrictions — for example, downgrading default models to Haiku or pausing non-essential surfaces. This “circuit breaker” behavior was opt-in, since it could disrupt active work.

Alerts could be delivered via email, Slack, Microsoft Teams, or webhook, and were configurable per admin role. A finance contact might receive all org-level alerts, while team leads received alerts only for their group.

33.7 The Admin API

The final component was the Admin API — a comprehensive REST API that exposed every administrative function available in the Claude Console, enabling scripted and automated cost-control workflows.

The Admin API supported:

• User and group management: provision, deprovision, and modify access programmatically.
• Policy configuration: set model defaults, entitlements, and spend caps via API.
• Analytics queries: pull usage and cost data (same as the Analytics API).
• Alert management: create, update, and delete spend-threshold alerts.

A common pattern was to combine the Admin API with an identity provider’s lifecycle events. When an employee left the company, the IdP would trigger a webhook that called the Admin API to immediately revoke their Claude access — closing a gap where departed employees could continue incurring costs until a quarterly audit caught it.

# Example: Automated deprovisioning via Admin API
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ADMIN_API_KEY"])

def deprovision_user(user_id: str):
    """Revoke all access for a departing employee."""
    client.admin.users.update(
        user_id=user_id,
        status="suspended",
        model_entitlements=[],  # Remove all model access
    )
    # Log the action for audit
    print(f"Deprovisioned user {user_id} at {datetime.utcnow()}")

33.8 A Practical Cost-Control Playbook

Putting all the pieces together, the announcement enabled a coherent cost-control strategy for enterprises:

1. Set model defaults at the group level so the right model is used by default for the right work.
2. Establish user budgets with 75%/95% alerts so anomalies surface early.
3. Configure org-level circuit breakers at 90% to prevent catastrophic overruns.
4. Export analytics via API into your FinOps platform for unified visibility.
5. Automate lifecycle management so access is revoked promptly when employees leave.

33.9 Key Takeaways

• Three pillars of control. The update delivered deeper analytics, intelligent cost controls (defaults, entitlements, alerts), and programmatic access (Analytics API + Admin API).
• Analytics now breaks down by group and user. Chargeback and anomaly detection became first-class capabilities, ending the era of aggregate-only dashboards.
• Claude Code gets dedicated analytics. The Usage tab tracks adoption; the Value tab quantifies ROI through productivity lift, cost per commit, and annual value estimates.
• Analytics Chat enables natural-language queries. Administrators can ask questions in plain English and receive interactive charts, lowering the barrier to insight.
• Model defaults and entitlements prevent accidental overspend. Organizations can tier access so the right model is used for the right job.
• Multi-tier spend alerts catch problems early. Organization-level (75%/90%) and user-level (75%/95%) thresholds notify the right people before budgets blow out.
• The Admin API enables automation. Provisioning, policy, analytics, and alerts are all scriptable, integrating Claude governance into existing IT and FinOps workflows.
• The design philosophy is proactive, not reactive. Every feature is oriented toward catching cost issues before they become problems — circuit breakers over post-mortems.