Anthropic brings per-message effort control to Claude on Google Cloud Vertex AI

Anthropic extended a beta feature that lets developers change a Claude model's reasoning effort mid-conversation to Google Cloud's Vertex AI platform on September 3, 2026.

What's new

Per the Claude Platform release notes: "Per-message effort changes, in beta, are also available on Google Cloud for Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5, with the same mid-conversation-output-config-2026-07-01 beta header."

The notable details:

• Three models covered: Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5 all get the capability on Vertex AI, matching the lineup already supported on Anthropic's own API.
• Same beta header, same behavior: Google Cloud customers use the identical mid-conversation-output-config-2026-07-01 beta header as the first-party Claude API, meaning code written against Anthropic's direct API for this feature should port to Vertex AI with minimal changes.
• Mid-conversation, not just per-request: the feature lets a caller change reasoning effort partway through an ongoing conversation, rather than only being able to set it once at the start of a session.

Context

Per-message effort control first shipped on Anthropic's own API; extending it to Vertex AI closes a capability gap for the large share of enterprise customers who consume Claude exclusively through Google Cloud for procurement, billing, or data-residency reasons rather than calling Anthropic directly. Keeping the same beta header and behavior across both surfaces is a deliberate choice — it avoids Vertex AI customers needing a separate integration path or waiting for a translated feature, which has been a recurring friction point for cloud-marketplace-hosted models generally.

Why it matters

Enterprises that standardized on Vertex AI for compliance or procurement reasons often trail direct-API customers in feature availability by weeks or months; matching feature parity this quickly signals Anthropic treats its cloud-marketplace listings as first-class deployment targets rather than a secondary channel. For developers building agentic workflows where reasoning cost should flex with task difficulty — cheap effort for simple steps, higher effort for hard ones — being able to make that adjustment mid-conversation, and now do it identically whether calling Anthropic directly or through Vertex, removes a reason to pick one distribution channel over the other purely for feature reasons.