Overview
Microsoft-Decision-1.1 is the second model in Microsoft's decision-scoring line. Achint Srivastava of Microsoft's Office of the CTO announced it as generally available on November 17, 2026, after a public preview the month before. Like Microsoft-Decision-1, it does not generate free text: given a fixed set of answer options — a yes/no, a choice or a score on a scale — it returns a calibrated probability for each option, for routing, classification, verification and agent guardrails.
Where the first model was post-trained from Qwen3.5-9B, version 1.1 is built on Gemma 4 26B A4B, a mid-sized mixture-of-experts model, and runs on a custom inference engine. Microsoft says it is more intelligent, better calibrated and more than twice as fast as Microsoft-Decision-1 at the same cost, and it adds image input: a request can now include photos or screenshots. The Microsoft Foundry catalog lists text and image input (up to 8 images per call), JSON output and a 32,767-token context window.
On the Artificial Analysis Decision Index v2.1, Microsoft reports a score of 68, six points above the first model, and says that places it ahead of Jev 1.14, Google's Gemini Decision 1 and OpenAI's GPT-6 Luna (Decisions); the post gives no competitor scores. In Microsoft's own evaluation of 102 benchmarks and more than 385,000 held-out questions it had the highest average accuracy of eight decision models and led seven of nine capability categories. For Copilot request routing, Microsoft says it matched an LLM's accuracy at 2.2x the speed and 4.4x lower cost.
Microsoft says more than 4,000 customers adopted the decision model in its first five weeks, naming Epic, Equinor, Atos Group and Raiffeisen Bank. It is served in Microsoft Foundry, on OpenRouter and on Vercel AI Gateway, and can act as a judge in Foundry Evaluations.
| Released | 2026-11-17 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Context | 32,767 tokens |
| Architecture | Gemma 4 26B A4B (mixture-of-experts) post-trained by Microsoft for single-pass decision scoring; returns per-option probabilities as JSON instead of generated text |
| Modalities | Text, Vision |
| Status | Generally available |
Strengths
- Image input — photos and screenshots, up to 8 per call — on top of text, a first for the Microsoft-Decision line
- More than twice as fast as Microsoft-Decision-1 at the same cost, per Microsoft
- Score of 68 on the Artificial Analysis Decision Index v2.1, six points above the first model, as reported by Microsoft
- Highest average accuracy of eight decision models across Microsoft's 102-benchmark, 385,000-question evaluation
- Calibrated probability for every answer option, returned as JSON from one structured call
Best for
- Reach for it to route requests to the right queue, team or model, so larger, pricier models are only called when needed.
- Reach for it to grade AI responses against a rubric or check that an answer is grounded in supplied evidence, including as a judge in Foundry Evaluations.
- Reach for it to screen content or screenshots for safety and to automate high-confidence decisions while escalating uncertain ones.
- Look elsewhere for text generation, open-ended question answering, audio or video input, or explanations — the Foundry catalog rules these out — and never use it as the sole decision-maker in consequential decisions about people.
How to access
| Provider | Model ID |
|---|---|
| Microsoft Foundry ↗ | Microsoft-Decision-1.1 |
| Vercel AI Gateway ↗ | microsoft/decision-1.1 |
Microsoft-Decision (decision models) — every version
The full lineage of the Microsoft-Decision (decision models) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Microsoft-Decision-1.1current | 2026-11-17 | 32K | Proprietary |
| Microsoft-Decision-1 | 2026-10-09 | 32K | Proprietary |
FAQ
What changed in Microsoft-Decision-1.1?
It moves from Qwen3.5-9B to Gemma 4 26B A4B as its base, adds image input, and runs on a custom inference engine. Microsoft says it is better calibrated, more accurate and more than twice as fast as Microsoft-Decision-1 at the same cost.
Where can I use Microsoft-Decision-1.1?
It is generally available in Microsoft Foundry and is also served through OpenRouter and Vercel AI Gateway (model id microsoft/decision-1.1 on Vercel).
How large is Microsoft-Decision-1.1's context window?
The Microsoft Foundry catalog lists a 32,767-token context window, with text and up to 8 images per API call.
Is Microsoft-Decision-1.1 open-weight?
No. Microsoft offers Microsoft-Decision-1.1 as a hosted model on Foundry, OpenRouter and Vercel AI Gateway, and its launch post links no downloadable weights for it.
