Anthropic launches Inference Hooks beta for Claude Enterprise

Anthropic has opened a beta of Inference Hooks for Claude Enterprise organizations, a new governance layer that lets companies route every prompt through their own security infrastructure before Claude answers it.

What's new

According to Anthropic's Claude Platform release notes, published August 5, 2026: "Inference hooks are now in beta for Claude Enterprise organizations. Point Claude at your organization's AI security server, and each governed prompt across claude.ai, Cowork, and Claude Code is held for the server's allow or deny verdict before inference proceeds."

The mechanism works as a checkpoint: an enterprise customer designates an internal AI security server, and Claude pauses each request from claude.ai, the Cowork agent product, or Claude Code until that server returns a verdict. Anthropic says requests sent to the hook endpoint are signed, failure handling is configurable (so admins can decide whether an unreachable security server blocks or allows requests by default), and every denial is logged to the compliance Activity Feed that Anthropic ships for enterprise admins.

The same release notes entry confirms Anthropic has retired the Claude Opus 4.1 model (claude-opus-4-1-20250805); calls to it now return an error, with Anthropic pointing customers to Claude Opus 5 and noting that researchers can request continued access through its External Researcher Access Program.

Context

Inference Hooks extends a pattern Anthropic has been building out over the past several months as it courts large regulated enterprises: giving IT and security teams a way to sit between end users and the model rather than trusting client-side controls alone. It follows Anthropic's compliance Activity Feed and sits alongside other admin-facing controls Anthropic has rolled into the Claude Platform for Enterprise tier customers, extending oversight to Claude's newer surfaces, including Cowork, Anthropic's agentic-work product, and Claude Code.

The timing also lines up with Anthropic's broader enterprise governance push this cycle: the company separately named its first Chief Global Affairs Officer this week and has been publishing more detail on its own security practices, including a source-code security playbook.

Why it matters

For large enterprise buyers, the absence of a server-side policy gate has been one of the standard objections to deploying LLM agents broadly — security teams want a kill switch that doesn't depend on the model behaving itself. By making the security server, not the client, the final arbiter of whether inference proceeds, Anthropic is directly targeting that objection, and doing so across its full product surface (chat, the Cowork agent, and Claude Code) rather than just the API. Whether this becomes a durable differentiator will depend on how quickly OpenAI, Google, and other enterprise-focused vendors ship comparable server-side gating, and on how much latency the extra round-trip to a customer's security server adds to real-world usage.