- Claude auto-fallback now tries AWS Bedrock before internal relay paths for better reliability.
- Guardrail policies now also apply to Anthropic-compatible /v1/messages, including streaming responses.
- Credential health is now model-aware, so one unavailable model no longer degrades the full account path.
- Subscription usage now resets on each customer’s actual Stripe billing period, improving post-renewal accuracy.
- Video generation requests missing required resolution now return a clear 400 error.
- Claude token counting now uses provider-accurate accounting instead of approximation.
- Fixed IAM rule evaluation edge cases.
- Fixed /v1/responses URL derivation for account-scoped compatible providers.
Back to changelog
Safer routing, better guardrails coverage, and request-handling fixes
Get Started
Cut your AI bill, not your usage
Route every request to the right model. Track every dollar you spend. Cut your LLM costs by up to 60%.
