Microsoft Decision 1 added to BenchGecko · decision scoring at $0.042 per M input
Microsoft released Microsoft-Decision-1 on 2026-10-09 for structured decision scoring. BenchGecko lists it at $0.042 per M input tokens, free output, and 32,768 context.
What landed
Microsoft Decision 1 is a decision-scoring model from Microsoft, released 2026-10-09 and available in Microsoft Foundry and on OpenRouter. Unlike chat LLMs, it returns calibrated probabilities over fixed options for routing, classification, prioritization, verification, and workflow control.
Microsoft says the model was post-trained from Qwen3.5-9B for single-pass decision scoring. Supported question types include yes/no, multiple choice, and rating or rubric scores.
Pricing and context
Microsoft's announcement lists:
- Input · $0.042 USD per million tokens
- Output · free
OpenRouter lists the same $0.042 / $0 per 1M pricing and a 32,768 token context window. OpenRouter's measured P50 latency for the Azure route is about 0.25 s.
For price context on BenchGecko, GPT-6 Luna is $0.10/$0.50 per M tokens and GPT-6 Sol is $2/$10, both with 1,050,000 context. Decision 1 is not a drop-in chat replacement for those models.
Vendor-reported evaluation notes
From Microsoft's 2026-10-09 post (vendor-reported, not BenchGecko tests):
- Highest accuracy in a 36-benchmark comparison spanning nearly 150,000 questions kept blind from training
- Fastest measured in that suite · 2.5 times quicker than H2O-Lightning-4B v1.1 and 35 times quicker than GPT-6 Sol
- Robustness · decision flips on 1.3% of eight-way input perturbations on average, with zero flips when option descriptions are paraphrased or options are reversed or shuffled
- Safety screen · 5,250 requests across 11 benchmarks covering harmful content, jailbreaks, and prompt injection
Internal use cases Microsoft describes include Xbox Research labeling more than 10,000 open-ended feedback items (competitive with GPT-6 Sol quality while over 14 times faster and 200 times less expensive), Copilot quality control (competitive with GPT-6 Luna and 100 times faster), and Microsoft Discovery adaptive replanning (46 times more consistent than an LLM score at 3 times the speed, nearly 4 times faster on adaptive replanning).
Limits
Microsoft positions the model for structured decisions, not open-ended generation, conversation, translation, or summarization. It is not intended as the sole system for high-stakes personal decisions.
Sources
- Microsoft-Decision-1 announcement
- OpenRouter · Microsoft-Decision-1
- BenchGecko model page: /model/microsoft-decision-1
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.