Microsoft-Decision-1: $0.042/M Tokens, 35x Faster

Microsoft unveiled Microsoft-Decision-1 on October 9, 2026 — a decision-scoring model in Microsoft Foundry priced at $0.042 per million input tokens, with output tokens free — and CEO Satya Nadella framed it as a tool for fast, structured choices rather than open-ended chat. In a Command Line post by Achint Srivastava, VP of Software Engineering in Microsoft's Office of the CTO, the company said the model targets routing, classification, prioritization, verification, and workflow control, returning calibrated probabilities that software can act on immediately.

What happened

On Friday, October 9, 2026, Satya Nadella posted on X introducing Microsoft-Decision-1 for fast decision-making, saying it delivers top performance on structured decision tasks versus LLMs and other decision models, and that Microsoft is already testing it internally for uses such as incident response. Microsoft Developer amplified the same launch and linked the Command Line article.

Unlike general-purpose language models that generate prose, Microsoft describes Decision-1 as purpose-built for fixed sets of answer options. Given those options, it returns a calibrated probability score for each one through a structured API — including yes/no, multiple-choice, rating scales, and rubric-based grading of AI responses or agent actions. RuntimeWire reported that Foundry's listing positions the model for text inputs up to 32,768 tokens with JSON outputs, and notes it does not generate open-ended explanations or accept images, audio, or video.

Key details

Base model, API shape, and price

According to Microsoft's Command Line post, Microsoft-Decision-1 was built by post-training Qwen3.5-9B for fast, single-pass decision scoring. Microsoft said it will soon rebase the model on other foundations, including Microsoft AI (MAI) and OpenAI models. Developers can start today in Microsoft Foundry; OpenRouter access is "coming soon," with no date given.

Pricing, per Microsoft: $0.042 USD per million input tokens, and output tokens are free. That figure is the clearest public cost signal for treating decision checks as a cheap, repeatable service call inside agent workflows.

Microsoft's benchmark and speed claims

Microsoft said Decision-1 achieved the highest accuracy in its own 36-benchmark comparison spanning nearly 150,000 questions, with those benchmarks kept blind from training. In the same Microsoft benchmarking, it reported the model as the fastest measured: about 4.5 times quicker than Quyet-1.0-Large (the runner-up in that write-up) and about 35 times quicker than GPT-6 Sol, with P50 latency described as roughly 35x faster than GPT-6 Sol. Those results are Microsoft's own tests; the announcement does not publish full datasets or latency conditions for independent reproduction.

On robustness, Microsoft said it perturbs the same request in eight ways and measures decision flips. It reported that Microsoft-Decision-1 changes its decision on 1.3% of perturbations on average, with zero flips when option descriptions are paraphrased or when options are reversed or shuffled. Separately, Microsoft said a safety evaluation of 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts, and prompt injection found the model refused harmful behavior while retaining utility — again, Microsoft's framing of its own tests.

Internal Microsoft pilots (company-reported)

Xbox Research: Sorted more than 10,000 open-ended feedback and review items from surveys, Steam, and Twitter/X into fixed themes. Microsoft said quality was competitive with GPT-6 Sol while running over 14 times faster and 200 times less expensive.
Copilot quality control: Used to measure chat and agentic response quality. Microsoft said the model was competitive with GPT-5.6 Luna and 100 times faster.
Incident response: On-call engineers used it for knowledge retrieval across logs, tickets, calls, and messages; Microsoft said it performed better and faster than an LLM for that retrieval role.
Microsoft Discovery: In adaptive replanning, Microsoft reported Decision-1 as 46 times more consistent than an LLM-based score at three times the speed, and nearly four times the speed on the adaptive-replanning loop overall.

Microsoft also listed example use cases such as agent step control (continue, stop, retry, or hand off), model routing, data labeling, AI judging against rubrics, intent matching, search relevance labels, content filtering, and computer/UI action selection among fixed options.

Category context and guardrails

Decision models are becoming a named product category beyond Microsoft. Cloudflare's Clef models and OpenAI's Decisions API sit in a similar peer lane — bounded, structured choices rather than free-form generation — though each vendor's packaging differs. RuntimeWire noted that Foundry's catalog warns Decision-1 should not be the sole automated decision-maker for consequential matters involving employment, credit, housing, healthcare, or legal rights, so thresholds, human review, and application-level safeguards still matter.

Why it matters

Agents need a cheap "check" step: Multi-step agent workflows accumulate delay when every verification runs through a large chat model. Microsoft's own latency math — 100 ms added to each of 20 sequential decisions equals two seconds — explains why a small scoring model is being sold as infrastructure.
Probability is the product: Returning calibrated per-option probabilities lets apps decide when to act, defer, or escalate, instead of treating a free-text reply as an action plan.
Price is part of the pitch: At $0.042 per million input tokens with free outputs, Microsoft is positioning decision scoring as something teams can call often without the token bills of flagship LLMs.
Claims still need external proof: Accuracy, 35x speed vs GPT-6 Sol, and internal Xbox/Copilot gains are Microsoft-reported. Buyers will want their own task suites before trusting escalation thresholds in production.
Control layer for agentic software: As agents propose tool calls and next steps, a dedicated scorer can sit between "reason" and "act" — a role Srivastava's announcement frames as increasingly central once agents operate in the real world.

What's next

Microsoft said it will continue releasing updates, including by incorporating evaluations and data to further optimize quality, confidence, and cost, and that it plans to rebase Decision-1 on MAI and OpenAI models. OpenRouter availability remains "coming soon" without a public date. Open questions for developers include how well scores calibrate on customer-specific rubrics, whether Microsoft publishes more reproducible benchmark materials, and how Foundry customers combine Decision-1 with human review for high-stakes workflows. For now, the concrete path is the Foundry catalog listing and the Command Line getting-started guidance published October 9, 2026.

FAQ

What is Microsoft-Decision-1?

Microsoft-Decision-1 is Microsoft's decision-scoring model announced October 9, 2026. Per Microsoft, it scores fixed answer options — for routing, classification, prioritization, verification, and workflow control — and returns calibrated probabilities through a structured API, rather than generating open-ended text.

How much does Microsoft-Decision-1 cost?

Microsoft lists input tokens at $0.042 USD per million tokens, with output tokens free. The model is available now in Microsoft Foundry, with OpenRouter access described as coming soon.

How is Microsoft-Decision-1 different from an LLM?

Microsoft positions it as a decision model, not a chatbot: you supply a fixed set of options, and it returns probability scores software can act on. Secondary coverage notes Foundry's listing excludes open-ended generation and multimodal inputs (images, audio, video). Microsoft says it was post-trained from Qwen3.5-9B for single-pass decision scoring.

Search This Blog

The Frontier Brief