Bringing the Frontier to Critique
GPT-6 Astra and Claude Fable 5.1 are now available for the review runs where context, judgment, and a clean merge matter more than bargain-basement inference.
The frontier is now a routing choice inside Critique. GPT-6 Astra and Claude Fable 5.1 both sit at a 60-credit shelf floor, both cost $50 per million output tokens, and the input split is deliberate: Astra at $5/M, Fable at $10/M.
A model catalog is only useful when it changes what the system can do. Adding two impressive names to a dropdown is not the point. The point is giving a review system a credible escalation path when the change is large, ambiguous, or expensive to get wrong — without turning every routine pull request into a frontier-model invoice.
Two frontier routes, one clean contract
GPT-6 Astra and Claude Fable 5.1 are available anywhere Critique exposes its managed runtime catalog: lead review, specialist review, and Remedy. They are first-class routes, not one-off provider exceptions. That means the same receipt can show the selected model, token usage, model cost, and Critique execution cost together.
The expensive part is visible on purpose
Published managed rates per one million tokens. Credit floors cover the review slice; token usage remains visible as its own line item.
Lower input rate for context-heavy review and Remedy plans.
Deep Anthropic escalation when the review deserves the premium lane.
Rates are USD per 1M input/output tokens. The 60-credit floor is the managed catalog floor, not a promise that every review consumes exactly 60 credits.
- GPT-6 Astra5$ / 1M
- Claude Fable 5.110$ / 1M
Astra’s input rate is half Fable’s. That matters when a review spends more tokens reading a repository than writing a verdict.
- GPT-6 Astra50$ / 1M
- Claude Fable 5.150$ / 1M
The output rate is intentionally simple: both frontier routes are $50 per million output tokens.
When should the router reach for them?
Not for every diff. Flash and mid-tier models remain the right answer for volume, straightforward changes, and specialist checks. The frontier is for the moments where a missed dependency, incorrect security conclusion, or fragile repair plan costs more than the inference.
That is the shape we want teams to adopt: cost-aware escalation, not model worship. Start economically. Escalate when the risk profile earns it. Keep the provider, model, tokens, and execution charges visible so the routing policy can improve over time.
The frontier belongs behind the merge gate
The best place for a frontier model is not necessarily the first draft. It is the final judgment: the pass where the system has already gathered the diff, repository context, specialist findings, test output, and the facts a lead reviewer needs to decide. Astra and Fable give Critique two more serious options for that job.