By Anthropic shipped three enterprise cost controls on July 2: model-level entitlements, spend-threshold alerts, and a redesigned analytics dashboard with natural language querying. These are not nice-to-have governance features. They are the direct response to a billing crisis that has already ended Claude Code deployments at major companies. If your organization runs Claude agents or Claude Code at any scale, you need to configure these today.
The Problem These Controls Were Built to Solve
Agentic AI tasks do not behave like chat queries. According to Gartner’s March 2026 analysis, a single agentic task triggers 5 to 30 model calls. GitHub’s May 2026 research puts the worst-case token consumption of an agentic coding session at 1,000 times that of a single-turn query. A complex orchestrated workflow that cost $0.04 in 2023 now costs $1.20 per interaction — a 30x increase driven entirely by agentic loop depth, not price hikes.
The result: enterprises have been blindsided by exponential bills. Uber burned through its entire 2026 AI budget in four months. Microsoft pulled Claude Code from its enterprise stack after costs spiraled. One company reportedly spent $500M on Claude in a single month with no cap in place. Until July 2, the available Claude Enterprise spend controls were too coarse to stop this — admins could set an org-level spend cap but had no visibility into who was spending what on which model before the cap hit.
What Launched on July 2
Model-Level Entitlements
The most impactful change for most teams. Admins can now set the default Claude model for new conversations across chat, Cowork, and Claude Code — and restrict which models specific groups can access at all.
Without entitlements, every user who opens Claude Code gets the most capable model available, which is also the most expensive. If Opus 4.8 is in your Enterprise plan, that is what engineers are using by default for routine tasks that Sonnet 5 handles at a fraction of the cost. The fix: set Sonnet 5 as the default for Claude Code, reserve Opus 4.8 access for senior engineers or specific high-complexity workflows. The controls map to your existing SCIM groups, so this takes minutes to configure if your team directory is already provisioned.
Spend-Threshold Alerts
Org-level alerts fire at 75% and 90% of your set spend limit — enough lead time to raise the cap before engineers hit a wall mid-task. Individual users see alerts at 75% and 95% of their personal limits, with a direct link to request an increase from admins. The Admin API lets you automate the full workflow: flag users approaching limits, trigger notifications, run approval scripts — all without manual monitoring.
Analytics Dashboard and Analytics Chat
The admin dashboard now shows cost and usage broken down by group and by user, with outputs listed next to their dollar cost: artifacts created, files edited, skills and connectors consumed. Two new Claude Code Insight tabs surface developer-specific metrics — active developer counts, session counts, top commands — plus a value calculation layer showing productivity lift and cost per commit with visible formulas.
Analytics Chat deserves specific attention. Admins ask plain English questions — “Which teams doubled their Claude usage this month?” or “What is our cost per deploy across engineering groups?” — and get exportable charts. This is not a novelty feature. It is the FinOps conversation every engineering leader has been trying to have with a spreadsheet.
The API Layer for Teams That Build Their Own Monitoring
The Analytics API (Claude Enterprise on claude.ai) and the Usage and Cost Admin API (Claude Console API users) expose the same underlying data programmatically. The Usage API returns token breakdowns by model, workspace, API key, and service tier at up to one-minute granularity, with a typical 5-minute data lag. The Cost API returns USD costs at daily granularity, grouped by workspace or description.
curl "https://api.anthropic.com/v1/organizations/usage_report/messages?\
starting_at=2026-07-01T00:00:00Z&\
ending_at=2026-07-05T00:00:00Z&\
group_by[]=model&\
bucket_width=1d" \
--header "anthropic-version: 2023-06-01" \
--header "x-api-key: $ANTHROPIC_ADMIN_KEY"If your team already uses Datadog, CloudZero, Grafana Cloud, or Vantage, Anthropic has native integrations for all four. Datadog normalizes Claude usage in FOCUS format so it sits inside existing Cloud Cost Management dashboards. CloudZero maps token consumption to business dimensions — team, product, feature, environment — for genuine unit economics rather than aggregate spend totals.
What to Configure Right Now
If you are on Claude Enterprise, prioritize these in order:
- Set model defaults. Change the default model for Claude Code and chat from Opus 4.8 to Sonnet 5. This single change will cut costs for the majority of your engineering team’s sessions immediately.
- Configure spend alerts. Set org-level limits and enable the 75%/90% notifications. For large teams, wire the Admin API into your incident channel.
- Pull Analytics into existing tools. Connect the Analytics API to your cloud cost tracking platform. Do not build a new dashboard — attach to the one you already review.
- Audit Claude Code usage by team. Use the Code Insights tab to identify who is generating the highest token counts. Those teams are candidates for model routing — Sonnet 5 for routine sessions, Opus 4.8 only for complex work that requires it.
The billing crisis that forced these controls into existence did not happen because agentic AI is inherently expensive — it happened because organizations had no visibility into consumption until bills arrived. These controls fix that. The only remaining failure mode is not configuring them.
