Microsoft Decision-1: the model that picks answers from options 35x faster than GPT-6 Sol

Microsoft has released Microsoft-Decision-1, a compact model built on the open Alibaba Qwen3.5-9B. The announcement came from Achint Srivastava, Vice President of Software Engineering in the Office of the CTO, in a blog post on October 9. Instead of generating new text, the model picks from a predefined set of options and returns a calibrated probability for each one.

What Decision-1 can do

The model handles yes/no questions, multiple-choice tasks, and ratings. It can also score AI responses and agent actions against an established rubric. Use cases include query routing, text classification, prioritization, verification, and workflow management. Decision-1 accepts only text input, with a 32,000-token context window. An important caveat from the documentation: don't use it as the sole basis for decisions about people — things like credit, employment, or personnel evaluation.

Benchmark performance

Decision-1 delivered the best accuracy across 36 benchmarks covering nearly 150,000 questions the model hadn't seen during training. Speed results: • 2.5x faster than H2O-Lightning-4B v1.1 (its closest competitor) • 35x faster than OpenAI's GPT-6 Sol • When a prompt was rephrased or reordered, the model changed its decision in just 1.3% of cases on average

The model was also tested for safety: 5,250 queries across 11 benchmarks, including checks for harmful content, jailbreak attempts, and prompt injections.

Deployment and pricing

Xbox Research used Decision-1 to sort more than 10,000 player reviews by topic. Quality results were competitive with GPT-6 Sol, but Decision-1 ran 14x faster and cost 200x less.

The Copilot team compared Decision-1 with GPT5.6 Luna: comparable quality, but 100x faster.

Pricing: $0.042 per million input tokens, with output tokens free. It's available on Microsoft Foundry and through OpenRouter — the model-routing platform that Stripe acquired in August.

A fast-growing category

Decision models are rapidly emerging as a new AI category. Decision-1's main competitors include: • OpenAI's Decisions API (public beta) • Jev from TypeSafe AI — launched three weeks before Decision-1 and became the fastest-adopted model on Vercel in 2025; Decision-1 is 2.8x faster in latency tests • Cloudflare • Strands • Liquid AI • Perplexity • Snowflake

According to The Register, more than 100 decision models are now competing in this market. Microsoft plans to soon port Decision-1 onto other base models, including OpenAI's and its own MAI models, though the reasons behind this move weren't disclosed in the announcement.