At its Ignite conference on 17 November, Microsoft AI announced a family of seven new models spanning coding, security, imaging, voice, transcription and routing. The headline figure that travelled furthest — that a new "MAI Decision 1" is 10x faster and 200x cheaper than a general model — does not appear in Microsoft's own materials. The model Microsoft actually shipped is Microsoft-Decision-1.1, and the company's published claims are 50x faster and 79x cheaper than OpenAI's GPT-Sol 6.1, measured on a specific routing benchmark. Independent benchmarks confirm a real but smaller speed gain over Microsoft's own previous version, and place the new model eighth in a field of eighteen.
The seven models #
Microsoft's Ignite post lists the following models, all now running on the company's second-generation Maia 200 accelerator, which Microsoft says delivers 1.7x more tokens per watt [1][9]:
- MAI Cyber 2 Flash — vulnerability discovery and fixes; Microsoft claims 92.3% on CyberGym-E2E vs 91.5% for GPT-6 Astra at under 2% of the cost per task (priced at list price)[3][4]
- MAI Code 2 — everyday enterprise coding; Microsoft claims 98% of GPT-6.1 Sol performance at 40% of the cost, and 96% of Claude Opus 5.5 at 5% of the cost[3][4]
- MAI Code 2 Flash — lightweight coding; claimed to match GPT-6 Luna at 65% of the price and Haiku 5.5 at 20%[3][4]
- MAI Image 2.7 — image generation and editing; ranked #1 on Artificial Analysis image editing leaderboard at about one-sixth the price of the four models behind it[4]
- MAI Realtime 1 — speech-to-speech; claimed to beat GPT-Realtime 2.1, starting at $0.01 per audio minute[3][4]
- Dragon Copilot Transcribe — built on MAI Transcribe 2, described as the world's fastest and most accurate multilingual transcription model, powering 10 million patient encounters a month with multilingual error rates halved[1][3]
- Decision 1.1 — a specialised decision/routing model; Microsoft claims 50x faster and 79x cheaper than GPT-Sol 6.1[1][3][8]
What "Decision 1.1" actually is #
Microsoft's own materials never use the string "MAI Decision 1". The Ignite post calls the model Decision 1.1, the Foundry blog calls it Microsoft-Decision-1.1, and the model card carries the same name [2]. The "MAI Decision 1" label comes from third-party coverage of the keynote [2][6][7].
This is not Microsoft's first decision model. Microsoft-Decision-1 reached general availability on Foundry on 8 October 2026, built on Alibaba's Qwen3.5-9B (text-only, no published benchmarks at launch) [2]. Version 1.1 switches the base to Google's Gemma 4 26B-A4B, a mixture-of-experts model with roughly 4B active parameters, adds multimodal input (text and images), and publishes a full performance story [2][10].
The contract is unchanged: you send a state (text or JSON) plus typed questions with fixed option sets, and get back probabilities — not prose. Three question primitives are documented: yes/no, choice from a labelled set, and scored levels returning a probability-weighted average. The model runs on a custom inference engine written for decision workloads rather than a text-generation stack, served at /providers/microsoft/v1/systemone on Foundry [2][10].
Price is $0.042 per million input tokens (Global Standard), $0.05 in EU and APAC data zones. Output tokens are free — there are effectively none. The model is hosted in Foundry's "Direct from Azure" portfolio and is also accessible via OpenRouter and Vercel; no download or fine-tuning path exists [2][10].
The two speed-and-cost claims #
A figure of "10x faster and 200x cheaper than a general model" circulated on 17 November, sourced to a TestingCatalog post on X. Neither number appears in Microsoft's announcement [2][5].
Microsoft publishes two specific, scoped comparisons:
- Ignite post / Foundry launch : Decision 1.1 is "50x faster and 79x cheaper than GPT-Sol 6.1" — a single competitor, a single routing task, Microsoft's own harness[1][8][9] .
- Foundry launch post : Decision 1.1 is "more than twice as fast as our first model, at the same cost", with a Copilot example: routing between Copilot and Cowork at the same accuracy as a large language model, 2.2x faster and 4.4x cheaper[2][10] .
The second claim is checkable against a third party. Artificial Analysis measures Decision 1.1 at 123.25 ms per task on its Decision Index v2.1, against 274.25 ms for Microsoft-Decision-1 — a 2.23x speedup, matching "more than twice as fast" [2]. On cost, AA computes $139.64 per million tasks for 1.1 vs $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 [2].
The 50x/79x claim against GPT-Sol 6.1 is a vendor measurement on Microsoft's own traffic and harness. It describes the gap between a specialised decision endpoint (zero output tokens, no reasoning loop) and a general model doing the same job in JSON mode (tens to hundreds of output tokens plus reasoning). The gap is real; the exact multiple is Microsoft's choice of baseline [2].
Where it ranks independently #
On the Artificial Analysis Decision Index v2.1 (read 18 November 2026), Decision 1.1 scores 67.78, placing eighth of eighteen models [2]. Microsoft's launch post says it "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)" — every part of that is true [2][10]. Jev 1.14 sits 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 [2].
What the sentence does not say: seven models score higher, and the leader (SubQ Snap 1.2 at 72.58) 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 normal for a vendor and a poor proxy for a general ranking [2].
Sub-scores reveal 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 rises from 40.59 to 48.45, and its Brier score (calibration, lower is better) improves from 0.1965 to 0.1604 — the most meaningful number in the release, because it says the probabilities got closer to being true, which is the only property that makes thresholding a decision model a sane engineering choice [2].
The base-model switch is the real story #
Version 1 was a post-train of Qwen3.5-9B (dense 9B). Version 1.1 is a post-train of Gemma 4 26B-A4B (MoE, ~4B active). Microsoft's explanation is about serving economics: 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 [2][10].
Three consequences follow. First, a MoE with 4B active is cheap to serve per call 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 25-language claim is about the new base, and any calibration measured on version 1 does not transfer. Third, licensing on the base is Google's; the post-trained model is Microsoft's and hosted only — so the open-weight lineage buys nothing operationally. You cannot download this model [2].
Pricing footnotes on the coding models #
The MAI Code 2 headline — 98% of GPT-6.1 Sol's performance at 40% of the cost — uses a 50% promotional price that ends 17 May 2027. At list price ($1/$5 per million input/output vs promo $0.50/$2.50), MAI Code 2 costs about 81% of GPT-6.1 Sol per task, not 40% [4][5]. Against Claude Opus 5.5 the model stays dramatically cheaper either way (~10% of cost per task for 96.8% of the score). MAI Code 2 Flash ends up 30% more expensive per task than GPT-6 Luna at list price, though it still scores higher [4][5].
MAI Cyber 2 Flash's claim ("under 2% of the cost per task" vs GPT-6 Astra) is priced at list price ($0.10/$0.50 per million tokens vs Astra's $10/$50), making it the sturdier of the two cost claims — though Microsoft excluded 92 of 905 CyberGym-E2E tasks (10.2%) that overlapped its training data, and the model ships only inside Microsoft security products (MDASH from mid-December) [4][5].
What shipped, what's real, what's marketing #
- Seven models announced , all running on Maia 200 silicon in production[1][3] .
- Decision 1.1 is real : GA in Foundry, OpenRouter and Vercel; used by 4,000+ customers per Microsoft[10] ; multimodal; 2.23x faster than v1 on independent benchmarks; better calibrated (Brier 0.1604).
- The 200x figure is not Microsoft's ; the company's claim is 79x vs one model (GPT-Sol 6.1) on one harness.
- Independent ranking : 8th of 18, top half but not the leader.
- Promotional pricing drives the sharpest coding-model ratios; list prices are notably less aggressive.
- MAI Licensed Weights announced: custom MAI models in customers' own Azure environments, coming soon via Frontier Company engagements[1][4][5] .
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.
Sources #
1. [Seven new world class models at scalable price-performance](https://microsoft.ai/news/microsoft-ignite-2026) (microsoft.ai)
2. [MAI Decision 1 Is Really Decision 1.1: Whose 200x?](https://www.orcarouter.ai/blog/mai-decision-1-ignite-2026) (orcarouter.ai)
3. [Ignite 2026: Microsoft AI Announces 7 New Models](https://www.thurrott.com/a-i/microsoft-ai/343999/ignite-2026-microsoft-ai-announces-7-new-models) (thurrott.com)
4. [MAI Code 2: Price, Benchmarks and Copilot Plans](https://felloai.com/mai-code-2) (felloai.com)
5. [Microsoft's MAI Code 2 Costs 40% of Sol at Its Promo Price](https://fourweekmba.com/ai-microsofts-mai-code-2-costs-40-of-sol-at-its-promo-price) (fourweekmba.com)
6. [MAI Decision 1 — это на самом деле Decision 1.1, а заголовок о 200x](https://www.orcarouter.ai/ru/blog/mai-decision-1-ignite-2026) (orcarouter.ai)
7. [La décision 1 de MAI est en réalité la décision 1.1 : à qui appartient le 200x ?](https://www.orcarouter.ai/fr/blog/mai-decision-1-ignite-2026) (orcarouter.ai)
8. [Microsoft Ignite 2026: MAI keynote transcript | Microsoft AI](https://microsoft.ai/news/microsoft-ignite-2026-mai-keynote) (microsoft.ai)
9. [Seven new world class models at scalable price-performance | Microsoft AI](https://aka.ms/MAI-Ignite-2026) (aka.ms)
10. [Microsoft-Decision-1.1 is now generally available for fast, low-cost decision making](https://commandline.microsoft.com/microsoft-decision-model-ga) (commandline.microsoft.com)