ENPT

Microsoft-Decision-1

Last updated:

GA—$0.042/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity6.4/10

    GA in Foundry and listed on OpenRouter. Microsoft Learn has a public deployment/API guide and REST examples; no model-specific SDK found. One live integration. Closed weights; no model-specific SLA or status page found.

  • Capability7.9/10

    Three decision types; JevBench reports 84.4 calibration, 459 ms median latency and composite 69.1 (#6 hosted API). Its 21 context-limit failures count in 1,500 items. Strong third-party evidence, though Jev leads overall.

  • Adoption24/100

    Announced by Satya Nadella (~4.7k likes on day one); now listed on OpenRouter. Microsoft reports internal tests at Xbox Research, Copilot, on-call incident response and Microsoft Discovery; no named external production customer verified.

Vendor claims

  • Highest accuracy in a 36-benchmark comparison (~150,000 blind questions); fastest measured at launch, 2.5× faster than H2O-Lightning-4B v1.1 and ~35× faster than GPT-6 Sol at p50; 1.3% decision flips across 8 perturbations[Vendor claim — not independently verified]

News covering this model

Read this first

  1. GA, closed API. Available in Microsoft Foundry and OpenRouter. No weights.
  2. Separate vendor claims from external measurements. Microsoft’s 36-benchmark set is not public; JevBench provides the independent comparison below.

Specs
Base Qwen3.5-9B, post-trained for single-pass decision scoring. Microsoft says it will rebase on MAI and OpenAI models
Types Yes/no, multiple choice, rating; rubric grading of AI responses and agent actions
Price $0.042 per 1M input tokens; output free
Access Foundry · OpenRouter · Microsoft Learn API guide

Numbers to read carefully

The current launch article reports the highest accuracy in a 36-benchmark comparison and says the model was 2.5× faster than H2O-Lightning-4B v1.1 and ~35× faster than GPT-6 Sol at p50. It also reports 1.3% average decision flips across eight perturbations, 5,250 safety requests across 11 benchmarks, and internal Xbox/Copilot results. These remain vendor claims: the 36-benchmark set and per-task results are not public. The article was updated to add a Jev comparison; the independent JevBench results below provide a separate peer measurement.

JevBench result (10 Oct 2026)

Benchmark Heaven’s hosted API board, which says it measures models independently of TypeSafe AI, places Microsoft-Decision-1 #6 of 27 API offerings on the JevBench v1.6.1 scale: composite 69.1, capability 70.8, intelligence 57.2, and calibration 84.4. The full run covered 1,500 decisions. 21 context-limit failures remain in the scored denominator. Median endpoint latency was 459 ms on a 124-request speed subset; board-calculated cost was $0.019 per 1,000 decisions.

Jev ranks #4 at 71.5 and Sage #1 at 74.0 on the same API board, so Microsoft does not lead the overall comparison. Its category radar is stronger than Jev’s on several named use cases. Benchmark Heaven cautions that category pools and supported types can differ; treat these as category signals, not a controlled head-to-head:

Use case Microsoft-Decision-1 Jev 1.13.0
Recruiting 29.7 0.0
Risk assessment 51.2 32.8
Gaming 25.1 5.1
Advertising 37.0 16.4
Scientific discovery 36.9 21.2
Moderation and trust & safety 67.2 51.6

The mixed-language radar reports 88.32 across 79 completed items. The launch post calls this the strongest API model in its mixed-language top ten; the sample is useful but small. The board’s score, calibration and speed are independent measurements; the Microsoft launch claims above remain vendor claims.

Clinical System One preview (10 Oct 2026)

The EdgeEvals System One board lists Microsoft-Decision-1 tied at #5 of 38, with 611/669 credited decisions. It records 152/153 clinical-note decisions, 224/240 triage, 136/156 criteria and 99/120 coding decisions. The board also records 2 severe triage errors. Median latency is 417 ms; cost is $0.03129 per 1,000 decisions, or about $0.021 for this 669-case run. Its irrelevant-text check held 130/130 cases that were correct before adding irrelevant text, with no new severe errors.

These are preview-board results, not clinical validation: the board says its fixtures are synthetic and have not been reviewed by clinicians or certified coders. No patient data is used. This is an additional task-specific result; catalog scores stay unchanged. Data · methodology · post.

Score update (10 Oct 2026)

Capability rose from 5.8 to 7.9 on public third-party evidence. On 10 Oct, maturity also rises from 6.0 to 6.4: Foundry lists the model as GA, Microsoft Learn publishes an API guide, and OpenRouter lists it. Maturity working: availability 10 × 0.30 + docs 7 × 0.25 + integrations 5 × 0.25 + support 2 × 0.20 = 6.40. The guide has REST and Azure CLI examples plus a Python quickstart, but no model-specific SDK or public model SLA was found. The single OpenRouter integration gets a middle-band score. Adoption rises from 20 to 24: engagement 30 × 0.35 + ecosystem 20 × 0.35 + production 20 × 0.30 = 23.5 → 24. OpenRouter is one live integration; no named external production customer was verified. JevBench’s 459 ms median is an independent measurement; its board does not publish p99 here.

Score working (9 Oct 2026)

  • Maturity = 7 × 0.30 + 5 × 0.25 + 4 × 0.25 + 8 × 0.20 = 2.10 + 1.25 + 1.00 + 1.60 = 5.95 → 6.0
  • Capability = 9 × 0.30 + 5 × 0.25 + 4 × 0.25 + 4 × 0.20 = 2.70 + 1.25 + 1.00 + 0.80 = 5.75 → 5.8
  • Adoption = 30 × 0.35 + 10 × 0.35 + 20 × 0.30 = 10.5 + 3.5 + 6.0 = 20.0 → 20

Latency and accuracy at 4 under the “no absolute numbers” precedent; calibration 5 because it is claimed with no method. Re-score when an independent board measures it.

Artificial Analysis Decision Index v1.0 (13 Oct 2026)

Per the AA launch article, Microsoft-Decision-1 scores 54 (tied 2nd with GLiDE), is on the cost Pareto frontier at $31 per 1M tasks, and has the best non-hallucination rate among decision models (40% “I don’t know” when not right). Scores unchanged: accuracy was already 8. AA is a separate, independent board (four evaluations: AA-Omniscience-Decision, HLE-MultiChoice, When2Call, WildGuardTest). Its scores and ranks are not comparable with Decision Index 0.3 or JevBench. AA finds every decision model overconfident overall, mainly on the knowledge and reasoning tasks; most are fairly well calibrated on When2Call and WildGuardTest.

Launch-day observations (10 Oct 2026)

Yum describes Microsoft-Decision-1 as a fast, structured decision model rather than a general reasoning model. This matches Microsoft’s stated scope: fixed options and typed outputs. The Qwen3.5-9B lineage is also confirmed by Microsoft’s Foundry catalog.

The post’s 200–600 ms range is a personal usage report without workload, region, or measurement protocol. It is plausible and consistent with JevBench’s 459 ms median on a 124-request speed subset, but the two figures are not directly comparable. Microsoft Learn’s public API guide is marked last updated 9 Oct 2026 and describes the /providers/microsoft/v1/systemone endpoint and request shape as similar to a System One API. So the post’s “no documentation” observation appears to have missed an already published guide; “nearly identical” is stronger than the documentation supports.

The author’s view that comparing against GPT-6 Sol is odd, and that Microsoft is rushing into the category, is opinion. Microsoft does highlight GPT-6 Sol in its launch comparison; that contextualizes speed against a general LLM but does not replace comparison with decision-model peers. JevBench supplies that independent peer comparison.