Decision-1.1 now available on AI Gateway
Decision-1.1 adds structured Boolean, category, and score outputs through Vercel AI Gateway, with probabilities that can inform thresholds and escalation. It keeps Decision-1's request format, so switching is a model-ID change.
**Decision-1.1** is available on Vercel AI Gateway for structured Boolean, choice, and score answers. Multiple questions can share one input, and its calibrated probabilities can support thresholds for acting, review, or escalation. It accepts Decision-1's request format; change the model ID to `microsoft/decision-1.1` to switch. Use it for bounded classification or routing steps where your application needs a typed result and can define what confidence triggers action or human review. Requests appear in logs, count toward budgets, and support fallbacks. The source describes the model as more accurate, better calibrated, and faster than Decision-1 but gives no comparative measurements. Examples require AI SDK 7.0.128 or later for TypeScript, or OpenAI Python SDK 3.26.0 or later.
No separate implication has been recorded yet.
**Decision-1.1** is available on Vercel AI Gateway for structured Boolean, choice, and score answers. Multiple questions can share one input, and its calibrated probabilities can support thresholds for acting, review, or escalation. It accepts Decision-1's request format; change the model ID to `microsoft/decision-1.1` to switch. Use it for bounded classification or routing steps where your application needs a typed result and can define what confidence triggers action or human review. Requests appear in logs, count toward budgets, and support fallbacks. The source describes the model as more accurate, better calibrated, and faster than Decision-1 but gives no comparative measurements. Examples require AI SDK 7.0.128 or later for TypeScript, or OpenAI Python SDK 3.26.0 or later.
This updates the Gateway’s bounded-decision option with calibrated probabilities, shared inputs for multiple questions, and compatibility with Decision-1’s request format. Thresholds can route cases to action or human review, but the claimed accuracy, calibration, and speed gains lack comparative measurements, so workload testing remains necessary.