Bottom line: Anthropic shows in a new guide how IT admins can make Claude costs in enterprises visible and controllable through access gating, model controls, spending limits, and analytics tools.
Anthropic has published a guide for IT admins describing which built-in tools can be used to make Claude costs in enterprises visible and controllable. The article is aimed both at companies with an enterprise-wide rollout and at teams developing on the Claude platform via API.
Anthropic advises companies to use cost per outcome achieved rather than pure token consumption as the primary metric. As a decision aid, the guide proposes two questions: What would the work have cost without AI, measured in resources and time, or would the project not have been undertaken at all without AI? And is the task complex work requiring judgment and reasoning, or does it involve a high volume share of simple, repetitive activities? Anthropic points out that using a cheaper model for complex tasks can increase overall costs through retries and necessary human corrections, while using a frontier model for simple document processing pays for capabilities the task never uses.
For model selection, the Claude model family offers different deployment profiles: Opus for long-running work and coding, Sonnet for everyday tasks and analysis, and Haiku for high-volume routine tasks. Additionally, so-called Effort Controls allow the thinking effort and computational intensity of a model to be adjusted depending on the task. An Advisor tool enables smaller models to consult a frontier model when needed, as soon as they reach their limits. According to Anthropic, many organizations deploy multiple models in parallel within a single project – for example, a frontier model to assist in evaluating complex insurance claims, while Haiku pre-tags and sorts incoming documents.
The concrete control mechanisms differ depending on whether Claude runs as a product for employees or as an API foundation for custom applications: in the first case, control options lie with the admin, in the second with the development teams. For Claude Enterprise, Anthropic recommends a three-step approach. First, access to products such as Claude Code and Claude Cowork should be granted via access gating on a group or role basis and rolled out gradually rather than company-wide all at once. Next, model controls at two levels can be used to govern which models a team is allowed to use (entitlements) and which model a new conversation starts with by default (defaults) – for example, Sonnet as the default for most teams, while teams with more demanding tasks are given access to more capable models. Only afterward, once a baseline of real usage is available, should hard spending limits be set for the entire organization, individual users, or groups; for groups, the limit then applies per member. Additionally, admins can automate the review of requests to raise the limit and identify users who are approaching their limit or showing highly fluctuating usage patterns.
For ongoing monitoring of consumption, Anthropic states that three tools are available: Usage Analytics breaks down spending by person, team, and model, with the exported data intended to closely match invoices to facilitate reconciliation with billing. The Analytics API provides the same data to existing BI tools, financial systems, and internal dashboards, allowing Claude spending to be budgeted and forecast alongside other cost items. Via an Analytics Chat feature, admins can also ask usage questions in natural language.
Source: claude.com · Published August 3, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.