InvestNot yet confirmed elsewhere1 publisher2 min readPublished
Microsoft claims its Decision-1 scoring model answers 35 times faster than GPT-6 Sol
Microsoft released Decision-1, a model that scores fixed answer options and, by its own benchmarks, runs 35 times faster than GPT-6 Sol. Whether routing work moves off general-purpose models now depends on independent tests and on pricing that confirm those self-reported figures.
The Investor · Invest desk
What happened
- Decision-1 is a post-trained version of Qwen3.5-9B, the open-weight model from Alibaba, and does not sit on a Microsoft-made base model.
- The model assigns calibrated probability scores to a fixed set of answer options, which can be yes/no questions or multiple-choice setups.
- Microsoft is running internal trials of the model in incident response, quality control and scientific discovery workflows.
- Outside users can reach Decision-1 through Microsoft Foundry now, and the model is scheduled for deployment on OpenRouter.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Teams routing tickets or alerts at volume through general models stand to save, since a 9B scorer needs less compute per answer, provided Microsoft's per-call rate follows that compute.
- decision A calibrated score forces each team to choose the confidence cutoff below which a case goes to a human, and that cutoff decides how much work the model takes off staff.
- exposure Cutoffs tuned on the Qwen-based release are exposed to change when Microsoft moves Decision-1 onto MAI or OpenAI models, because a new base can shift what a given score means.
Microsoft says Decision-1's latency is 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol [15][16]. If both comparisons came from the same test, Quyet-1.0-Large would itself be about 7.8 times faster than GPT-6 Sol, because 35 divided by 4.5 is 7.8 [13]. Most of the gap against the general model is shared with the nearest rival. A buyer choosing between Decision-1 and GPT-6 Sol is mostly choosing a category, while a buyer choosing between Decision-1 and Quyet-1.0-Large is choosing a 4.5-times difference in latency [15].
The speed has an ordinary source. The 9B in the name puts the model in the compact class, and smaller models generally need less computing power per answer [7]. It also reads the whole input, up to 32,000 tokens, and returns its scores in one shot without generating an answer word by word [3].
Microsoft did not pretrain a base model for this product [5]. The work it did was post-training, and the launch leans on distribution through Foundry and OpenRouter [9]. Future updates are planned on MAI and OpenAI models [6]. Microsoft has already said it will move the product off the Qwen base it launched on.
Independent tests could cut the accuracy lead. The suite of nearly 150,000 blind questions is Microsoft's own, and so are the speed figures [14][4]. Confidence scores that look calibrated on a benchmark could also drift on a company's live tickets. CryptoBriefing's report of the launch does not include a price [12], and a high enough per-call rate would wipe out a latency advantage of any size.
I think the demand is real, or rather it is real wherever routing and classification run at volume [1]. Liquid AI and vLLM Semantic Router have brought out similar specialised products in the same period [10]. The counter-thesis is volume. A team making few calls a day can live with a general model's latency, and a 35-times speedup on a handful of decisions saves it little.
The thesis fails if Decision-1's scores turn out no better calibrated on live data than the confidence a general model states about its own answers. According to CryptoBriefing, the point of a calibrated score is that it "lets a business decide when to trust the machine and when to escalate to a human" [11].
What to watch
- Whether Microsoft publishes results from the incident response and quality control trials, the first evidence from live workloads.
- A third-party latency test running Decision-1, Quyet-1.0-Large and GPT-6 Sol on the same setup, to confirm or break the implied 7.8-times gap between the two rivals.
- Whether the first MAI- or OpenAI-based update of Decision-1 keeps the 4.5-times latency lead the Qwen-based version claims.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption15
- Hype gap+35
- Incentives65
- Confidence35
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Microsoft describes Decision-1 as a decision-scoring model built for tasks like routing, classification, prioritization, verification, and workflow control.
- [2]
Decision-1 assigns calibrated probability scores to a set of fixed answer options, which can be simple yes/no questions or multiple-choice setups.
- [3]
Decision-1 can handle inputs of up to 32,000 tokens in a single pass, taking in the whole input and returning its scores in one shot rather than generating an answer word by word.
- [4]
The benchmark figures are Microsoft's own and self-reported; testing remains internal for now.
- [5]
Decision-1 is not built on a Microsoft-made base model; it is a post-trained version of Qwen3.5-9B, the open-weight model from Alibaba.
- [6]
Microsoft has said future updates of Decision-1 are planned on Microsoft AI (MAI) and OpenAI models.
- [7]
The 9B in the name refers to the model's size, placing it in the compact category; smaller models generally need less computing power per answer.
- [8]
Microsoft says Decision-1 is in internal trials across incident response, quality control, and scientific discovery workflows.
- [9]
Decision-1 is available to outside users through Microsoft Foundry and is scheduled for deployment on OpenRouter, which gives developers a single gateway to many different models.
- [10]
Similar specialized offerings from Liquid AI and vLLM Semantic Router have emerged in the same time frame as Decision-1.
- [11]
A calibrated confidence score lets a business decide when to trust the machine and when to escalate to a human.
- [12]
CryptoBriefing's report of the Decision-1 launch does not include pricing for the model.
- [13]
If Microsoft's two latency comparisons were measured on the same test, Quyet-1.0-Large is implied to be about 7.8 times faster than GPT-6 Sol.
- [14]
Microsoft claims Decision-1 achieved the highest accuracy across a benchmark suite of nearly 150,000 blind questions.
ReportedInsufficientSource: Microsoft, as reported by CryptoBriefing2 sources— create a free account to open themView cited source - [15]
According to Microsoft, Decision-1's latency is 4.5 times faster than Quyet-1.0-Large, which Microsoft identifies as the nearest competitor.
ReportedInsufficientSource: Microsoft, as reported by CryptoBriefing2 sources— create a free account to open themView cited source - [16]
Against GPT-6 Sol, Decision-1 is reportedly 35 times faster.
ReportedInsufficientSource: Microsoft, as reported by CryptoBriefing2 sources— create a free account to open themView cited source
Sources
1 independent publisher whose own reporting we read for this story.
- cryptobriefing.comMicrosoft launches Decision-1, an AI model built to make fast calls
1 article · October 9, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.