Microsoft Decision-1 Is 35x Faster Than GPT-6 Sol on Structured Decision Tasks
Microsoft has released Decision-1, a model purpose-built for structured decision-scoring, available now in Microsoft Azure AI Foundry and on OpenRouter. The model targets routing, classification, prioritization, verification, and workflow control — tasks that require a calibrated probability output rather than generated text.
Decision-1 was post-trained on Qwen3.5-9B for single-pass decision scoring. Microsoft says it will rebase future versions on its own MAI models and on OpenAI’s stack.
Benchmark Claims
Microsoft evaluated Decision-1 across 36 benchmarks spanning roughly 150,000 questions, all held blind from training. Across that set, Decision-1 scored highest on accuracy.
On latency, the comparison is stark: P50 latency is approximately 35 times lower than GPT-6 Sol, and 4.5 times lower than Quyet-1.0-Large, the closest competitor on Microsoft’s internal leaderboard. For sequential decision pipelines, where each step depends on the output of the prior one, that latency gap compounds — every 100ms added across 20 chained decisions adds two seconds to overall workflow time.
How It Works
Decision-1 accepts a structured prompt with a fixed set of answer options and returns a calibrated probability for each. Supported modes:
- Yes/No
- Multiple-choice
- Rating scales
- Rubric-based grading of AI responses or agent actions
The model does not generate free text. It scores options. That makes it directly compatible with downstream software without post-processing or parsing.
Robustness Under Production Conditions
Microsoft says Decision-1 was tested specifically for consistency under real-world noise: paraphrased instructions, reordered answer choices, changed option keys, and formatting variation. The goal is deterministic behavior when inputs are semantically equivalent but syntactically different — a known failure mode for general-purpose LLMs used as classifiers.
Market Context
Decision models as a category are gaining fast. TypeSafe AI’s Jev, which also produces calibrated decision probabilities, launched September 15 and today raised $870 million at a $7.5 billion valuation, with self-reported adoption at one in three Fortune 500 companies. Microsoft’s entry into the same segment, using a JevBench-adjacent evaluation framework and identical use-case framing, positions Decision-1 directly against TypeSafe on enterprise routing and classification workloads.
Liquid AI’s d1-3B, released this week, targets the same category at the edge — 8ms latency on an RTX 4090 for on-device decision inference.
Where to Access
- Azure AI Foundry:
microsoft/Microsoft-Decision-1 - OpenRouter: Available via standard API