
MAI Decision 1 Is Really Decision 1.1 — and the 200x Headline Isn't Microsoft's
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 124 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 130 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 135 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 182 tok/s
Two numbers about MAI Decision 1 went around on 17 November 2026: that it is 10x faster and 200x cheaper than a general model. Neither figure appears anywhere in Microsoft's own announcement. The model Microsoft actually shipped at Ignite is Microsoft-Decision-1.1, and the company's published claims are 50x faster and 79x cheaper than GPT-Sol 6.1 — against a specific named model, in a specific harness, not against "a general model". That gap is worth leading with, because it is the difference between a marketing ratio and a number you can check. Here is what shipped, what Microsoft documented, and where an independent board puts it.
First, the name
Microsoft's materials never use the string "MAI Decision 1". The Ignite lineup post calls the model Decision 1.1, the Foundry blog calls it Microsoft-Decision-1.1, and the model card in the Foundry catalogue carries the same name. The "MAI Decision 1" label comes from third-party coverage of the keynote, along with "MAI Code 2", "MAI Code 2 Flash" and "MAI Cyber 2 Flash" — the same family, renamed by observers rather than by the vendor.
That matters for one practical reason: Microsoft-Decision-1.1 is not the first decision model Microsoft has shipped. Microsoft-Decision-1 reached general availability on Foundry on 8 October 2026, built on Alibaba's Qwen3.5-9B, text-only, with no published benchmark figures at launch. What arrived on 17 November is version 1.1 of the same product line, with a different base model, multimodal input and a much fuller performance story. Treat anything you read about "MAI Decision 1" as being about 1.1 unless it says otherwise.
What is actually in the box
The contract is unchanged in shape and considerably wider in reach. You send a state — text or JSON — plus one or more typed questions, each with a fixed option set, and you get probabilities back instead of prose. Three question primitives are documented: a yes/no ("noul" in the API), a choice from a labelled set, and a score on ordered levels whose return value is the probability-weighted average of the level indexes. Everything in one call, one pass, no decoding loop.
What is new in 1.1:
• Base model — Gemma 4 26B-A4B, Google's open-weight mixture-of-experts, post-trained by Microsoft. Version 1 was post-trained on Qwen3.5-9B.
• Input — text and images. Version 1 was text-only.
• Image handling — PNG or JPEG, 280 to 1,568 pixels on each side, up to 8 images per question including any referenced from the state, capped at 645,120 processed pixels in total. The service resizes to a fixed working size, so a larger upload does not consume more tokens.
• Context — 32,768 input tokens for the state-plus-question combination, the same ceiling as version 1.
• Languages — trained and evaluated in 25+, against 35+ supported outright by the base model. Microsoft is explicit that base-model coverage does not establish validated calibration per language.
• Serving — a decision endpoint rather than a chat completion: requests go to /providers/microsoft/v1/systemone on your Foundry resource, with a deployment name in the model field. Microsoft says it runs a custom inference engine written for decision workloads rather than a text-generation stack.
• Price — $0.042 per million tokens on Global Standard, $0.05 in the EU and APAC data zones. Same as version 1, and decision calls are billed on input tokens, because there is effectively no output to bill for.
• Training window — October to November 2026, per the model card.
• Weights — none. This is a hosted model in Foundry's "Direct from Azure" portfolio. There is no download and no fine-tuning path.
The exclusions are stated as plainly as the features, and they still define the product. Microsoft-Decision-1.1 is not for text generation, open-ended question answering, conversation, translation or summarization, and not for tasks that require knowledge absent from the input. It produces no rationale. The response contains probabilities and nothing you could show a user as an explanation, which means the explanation is your job and the threshold is your decision: 0.7 escalates, 0.95 does not, and that line is set by the application, not the vendor.
The two speed-and-cost claims, and which one survives scrutiny
The circulating version — "10x faster and 200x cheaper than a general model" — has no referent. A ratio against an unnamed model cannot be reproduced, and the two published Microsoft comparisons that do name something are narrower and larger at once.
The Ignite post says Decision 1.1 is "50x faster and 79x cheaper than GPT-Sol 6.1". The Foundry launch post says it is "more than twice as fast as our first model, at the same cost", and gives a Copilot example: routing between Copilot and Cowork at the same accuracy as a large language model, 2.2x faster and 4.4x cheaper. Those are vendor measurements on Microsoft's own traffic and Microsoft's own harness. They are not wrong; they are scoped. The 50x and 79x figures are about one competitor on one kind of routing decision, not about decision models in general.
One of the claims is checkable against a third party, and it holds. Artificial Analysis measures Decision 1.1 at 123.25 milliseconds per task on its Decision Index v2.1, against 274.25 ms for Microsoft-Decision-1. That is 2.23x, which is what "more than twice as fast" means. The cost claim also survives, in a weakened form: AA computes $139.64 per million tasks for 1.1 versus $142.56 for version 1 — about 2% cheaper, not the same price, because the two versions use different token counts per task even at an identical per-token rate. "At the same cost" is accurate about the price list and slightly generous about the bill.
Where it ranks, and the shape of the claim Microsoft made
Microsoft's launch post says Decision 1.1 "scores 68, up six points from our first model and ahead of Jev 1.14, Google's Gemini Decision 1, and OpenAI's GPT-6 Luna (Decisions)". On the Artificial Analysis Decision Index v2.1 read on 18 November 2026, every part of that is true. Decision 1.1 is at 67.78. Jev 1.14 is at 66.23, Gemini Decision 1 at 65.98, GPT-6 Luna at 60.25. Version 1 is at 62.07, so the six-point gain is real as well.
What the sentence does not say is where the model sits in the full field. It sits eighth on the board, seven points behind an outright leader and one and a half ahead of the model Microsoft named:
• 74.85 — d1.2-ultra (Liquid AI)
• 72.62 — Drex Max v1.1 (Nace.AI)
• 72.58 — SubQ Snap 1.2 (Subquadratic)
• 72.52 — GLiDE v1.2 (Fastino)
• 71.36 — Mammoth 27B (Base Labs)
• 70.41 — Decider v1.3 (Perplexity)
• 68.69 — d1.2 (Liquid AI)
• 67.78 — Microsoft-Decision-1.1

Seven models score higher, and the leader is 7.1 points clear. Microsoft's separate claim — "the highest average accuracy of eight decision models" across its own 102-benchmark, 385,000-question internal suite — is a statement about a field Microsoft chose, which is a normal thing for a vendor to publish and a bad thing to read as a ranking. The independent field runs to eighteen ranked models, and Microsoft is in its top half.
The sub-scores are more informative than the composite, because they show what the post-training bought. On AA's six constituent evaluations, Decision 1.1 scores 95.1 on When2Call tool calling, 88.8 on EvalSafe-O*NET, 75.7 on symmetry, 58.7 on AA-Paperclip document decisions, 60.5 on omniscience and 46.1 on HLE multiple-choice. Its accuracy index is 48.45 against 40.59 for version 1, and its Brier score — the calibration measure, lower is better — improves from 0.1965 to 0.1604. That Brier movement is the most meaningful number in the release: it says the probabilities got closer to being true, which is the only property that makes thresholding a decision model a sane engineering choice.

The base-model switch is the real story
Version 1 was a post-train of Qwen3.5-9B, a dense 9B model. Version 1.1 is a post-train of Gemma 4 26B-A4B, a mixture-of-experts with roughly 4B active parameters. Microsoft's own explanation is about serving economics rather than quality: enough active parameters for judgement, small enough to run across A100, H100/200 and GB200/300 fleets, with a custom engine that routes requests across the GPU pool for cost rather than latency alone.
Three consequences follow, and they are the ones a platform team should care about. First, a MoE with 4B active is cheap to serve per call precisely because most of the model is idle — that is where the 123 ms comes from. Second, the switch resets the language and calibration story: the base changed, the 25-language claim is about the new base, and any calibration you measured on version 1 does not transfer. Third, licensing on the base is Google's, while the post-trained model is Microsoft's and hosted only — so the open-weight lineage buys you nothing operationally. You cannot download this model, and the base being open does not change that.
What a call costs, in the units you actually pay
At $0.042 per million input tokens, a routing call carrying a 1,200-token ticket and a three-option question costs about $0.00005 — five hundredths of a cent. A million such calls cost roughly $50 at the price list; AA's measured $139.64 per million tasks is higher because its tasks are longer than a support ticket, which is the useful reminder that per-token prices and per-task costs are different units and the second one is the one that shows up in a budget review.
The comparison that makes the economics legible is against a general model doing the same job in JSON mode. You pay for its output tokens, which for a structured answer is tens to hundreds of tokens per call rather than zero, and you pay for the reasoning it does before answering. That is the gap Microsoft is describing with 79x, and the gap is real even when the exact multiple is a vendor's choice of baseline.
Where this gets awkward is the same place it always does with a single-vendor hosted scorer: there is one price list, one rate limit, and one region story. Foundry offers Global Standard and Data Zone Standard deployments, and the Data Zone variant costs 19% more at $0.05 per million tokens. If your workload is latency-sensitive, the docs recommend co-locating the Foundry resource with your application, which is honest advice that also tells you the model is not magic — it is a small MoE behind a queue.
Getting to it without rebuilding your stack
The decision endpoint itself is not on OrcaRouter's catalogue, and its predecessor is not either — Microsoft ships this model through Foundry and nowhere else. What is on the catalogue, at provider list price with 0% markup passed through, are the two objects you need to price the decision before you buy it. Gemma 4 26B A4B is the post-train's base and routes at $0.25 per million input tokens and $2.90 per million output; it is the base rather than the post-train, so it will not return Decision 1.1's calibrated probabilities, but it is the same architecture and the honest floor for what a general model of that shape costs. Jev 1.14, which Microsoft names and beats by 1.55 points on the independent board, routes at $0.042 per million tokens in and out.
That combination is useful for the evaluation phase rather than the production one. If you are deciding whether a decision model belongs in a routing path at all, you can benchmark the two routable endpoints on your own labelled tickets through one key, with automatic failover between providers, before you open an Azure deployment at all. If the answer is yes and the numbers favour Decision 1.1, you deploy it in Foundry and keep the router in front — one endpoint for decision calls and generative calls alike, so a decision model's rate limit does not become your application's outage.

Who should move now, and who should wait a fortnight
Move now if you are already paying a general model to classify, grade or gate, and your volume is high enough that per-call cost is visible on a chart. The 123 ms median and $0.042 price are a straightforward win over a frontier model doing the same judgement, and the multimodal input is the difference between a scorer you can point at screenshots and one you cannot.
Wait if your decision touches people — credit, employment, housing, insurance, healthcare, legal — because Microsoft's own card says not to use it as the sole decision-maker there, and nothing in 1.1 changes that. Wait also if you need calibration you can defend: 1.1's Brier of 0.160 is an improvement, not a guarantee, and the docs are frank that scores shift with wording and option order. Validate on your own labels, randomise option order, and set thresholds against the cost of the two error types rather than against a vendor chart.
The honest summary of 17 November is smaller and more useful than the 200x headline: Microsoft shipped a faster, multimodal, better-calibrated decision scorer at a price that makes per-request judgement affordable, and it is good enough to be in the top half of an independent field rather than at the top of it. The model is real. The ratio in the tweet is not.
