Microsoft-Decision-1: An 85 ms Model That Answers for $0.042 a Million Tokens, and Why Your Ticket Queue Should Care

Ask a general-purpose LLM to decide which team a support ticket belongs to and you wait for it to write the answer back, then pay for every token of that answer. Microsoft's new model does the same bounded job in a median 85 milliseconds, charges $0.042 per million input tokens, and charges nothing at all for output[1] — because it never generates prose in the first place.

What actually happened

On 9 October 2026 Microsoft announced Microsoft-Decision-1, a decision-scoring model available in Microsoft Foundry and through OpenRouter, in a post by Achint Srivastava, VP of Software Engineering in the Office of the CTO[1].

It is not a chat model, and that is the point. You hand it input text or JSON plus a fixed set of options; it returns a calibrated probability for each option through a structured API call — yes/no, multiple choice, rating scales, and rubric-based grading of AI output and agent actions[1]. Microsoft's deployment guide describes the difference plainly: unlike a generative LLM, the model "doesn't generate a free-form response or written rationale. It makes a focused judgment about text or JSON and returns a typed, numerical decision"[2].

Under the hood it is a small, repurposed open model. Microsoft post-trained Qwen3.5-9B for fast, single-pass decision scoring, and says it plans to rebase the service on other models, including its own MAI family and OpenAI's[1].

The numbers Microsoft published are its own, and they are specific. Across a 36-benchmark comparison covering 147,137 questions that were kept blind from training, Microsoft-Decision-1 posted an average accuracy of 83.5, ahead of Jev 1.13.0 at 82.3, Quyet-1.0-Large at 81.9, GPT-6 Luna Decisions at 79.4 and H2O-Lightning-4B at 77.2[1]. On speed it reports a median (p50) latency of 85 ms and a p95 of 125 ms — 2.5 times quicker than the runner-up, H2O-Lightning-4B v1.1, and 35 times quicker than GPT-6 Sol — measured against JevBench v1.6.1 adjusted median latency checked on 7 October, with Decision-1 measured through Foundry in the same region[1]. The post carries an editor's note saying it was later updated to add Jev's accuracy and calibration numbers[1].

Two properties matter more to builders than the leaderboard. First, stability: Microsoft perturbs the same request eight ways and counts how often the decision flips, reporting flips on just 1.3% of perturbations on average, and zero flips when option descriptions are paraphrased or options are reversed or shuffled[1]. Second, calibration: the probability is part of the API response, not an internal ranking, so Microsoft's stated contract is that a 90% prediction should be right about nine times out of ten on representative cases[1]. Safety testing covered 5,250 requests across 11 benchmarks spanning harmful content, jailbreak attempts and prompt injection[1].

Microsoft's internal pilots are where the cost claim comes from. Xbox Research sorted more than 10,000 open-ended feedback items into fixed research themes and found Decision-1 "competitive on quality with GPT-6 Sol while running over 14 times faster and 200 times less expensive"[1]. The Copilot team's quality-control testing found it competitive with GPT5.6 Luna and 100 times faster, and in Microsoft Discovery's adaptive replanning loop it was 46 times more consistent than the LLM-based score at three times the speed[1].

Pricing is the headline: $0.042 per million input tokens, output tokens free[1]. For comparison, OpenAI's Decisions API — in public beta, currently serving only gpt-6-luna, and pitched as about 10x faster than its Responses API — charges $0.10 per million input tokens with no output or cache charges[3]. Both companies are converging on the same shape: pay for evidence in, not prose out.

Why it matters for Indian SMBs and mid-market teams

• The volume work in an SMB is decisions, not essays. Which queue does this ticket go to? Is this invoice line valid? Is this lead sales or spam? Does this support reply meet our quality rubric? None of it needs eloquence, and all of it runs thousands of times a day — which is exactly the workload where a per-token prose model quietly becomes the biggest line item in your AI bill.
• Free output tokens change the arithmetic more than the input price. In classification and routing tasks the model's answer usually dwarfs the input. Microsoft's own Xbox figure — 200x cheaper than GPT-6 Sol at comparable quality on 10,000+ real feedback items[1] — is a vendor-reported result, but the direction is structural: you stop paying for tokens you never needed.
• 85 ms puts the check inside the request, not in a nightly batch. At that latency you can afford a decision gate on every incoming message: auto-act above a probability threshold, send the middle band to a person, auto-reject the bottom. That threshold policy is also the answer to the audit question every Indian firm eventually faces — who decided this, and on what confidence?
• Set thresholds from your own labelled sample, not from a vendor table. OpenAI's own Decisions guidance tells developers to choose routing and filter thresholds from labelled examples using the cost of false positives versus false negatives[3] — the right advice regardless of whose model you buy. A benchmark average of 83.5 says nothing about your Hindi-language invoice forms.
• Treat every number in the announcement as vendor-reported. The benchmarks were run by Microsoft, blind from training but not blind to Microsoft, and independent third-party runs had not appeared in the days after launch[1]. Procurement should mirror Microsoft's own framing: pilot it on your documents, measure precision and recall at your threshold, and keep the human-review band wide until the evidence narrows.

Where Guruji Corporation fits in

Most teams do not need a new model; they need to find the dozen bounded decisions buried inside a workflow that is currently being run by a general-purpose LLM — or by an intern. We map those decision points, pull a labelled sample from your own history, run it against the candidate models, and hand back the threshold table: at what confidence you automate, at what confidence a person looks, and what each choice costs you per thousand tickets.

If the answer turns out to be "your current setup is fine at your volume", we will tell you that too — with the numbers. Talk to us and we will run the evaluation on your data before you commit to anything.

Sources - Introducing Microsoft-Decision-1, our model for fast decision-making — Microsoft - Deploy and use Microsoft-Decision-1 in Microsoft Foundry — Microsoft Learn - Decisions API — OpenAI