ReasonBlocks

ReasonBlocks is a hosted service for turning existing AI workflow traces into a specialized candidate model. It captures supported model requests through a gateway or accepts uploaded datasets, prepares an evaluation for review, and trains and evaluates a candidate model, letting users inspect comparison outputs before deciding whether to use it.

The workflow covers capture, training and evaluation, and testing. Users can test a private candidate endpoint and, when eligible, explicitly enable a limited traffic canary with a chosen percentage of requests, which can be disabled or rolled back. The capture gateway forwards supported traffic to OpenAI or Anthropic using the user's own provider key. Dataset preparation pauses for evaluation criteria and approval notes before paid baseline, training, and evaluation work, with run budgets and credit reservations constraining that work. Optional local filtering can remove detected personal details or replace them with source-specific aliases on saved training copies.

Access is requested through a conversation or access request, with sign-in available through Google. The service is hosted; customer-owned GPU and on-premises deployment are not available in this release.