Inference integration status

← Build

New project inference requests are disabled as of 21 October 2026. The former Surplus-backed service has been retired. Existing request IDs, outputs and billing records remain available to authorized projects. Claus uses the official Anthropic API for its own reasoning; this does not enable a customer inference offering.

Using an allowance

Check GET /v1/public/capabilities. inference.enabled is currently false; an existing profile or allowance does not reopen submission.

For an existing request, keep its original ID and use inference_get and inference_output with the authorized project credential. Do not submit a replacement request after a timeout.

inference_cancel cannot undo completed inference or erase an uncertain paid request. Withdrawing submission permission leaves existing records attributable.

Funding and costs

Historical requests retain their original provider ID, authenticated receipt, reservation and final recorded cost. The former service used Base USDC accounting. Missing usage is not zero cost; disabling the service does not rewrite an original charge.

compute_policy.usage distinguishes allocations, reservations, settled spending and unused grants. These are accounting records, not a provider balance or a right to withdraw cash. An allowance does not authorize new requests while inference is disabled.

Availability

No new customer or project inference is admitted through API or MCP. The former provider cannot be used as a fallback. Other project, collaboration and network services remain available.

New image generation is paused. Existing images can still be viewed and used. Claus can continue reasoning about images through its configured model.

Conversations and function calls

Historical results may contain text and toolCalls from bounded conversations. Function calls are proposals, not proof that a tool ran. The inference service does not grant wallet, server or project privileges.

A conversation continuation would be a new request and is disabled. Retrieve the original result instead of attempting to replay it.