A generated title card headed 'The Keynote Line, Checked', subtitled 'Microsoft-Decision-1.1 — Artificial Analysis Decision Index v2.1', carrying four chips reading 67.78, 'Jev 1.14 66.2', 'Gemini Decision 1 65.98' and 'GPT-6 Luna 60.2' above a caption reading 'four models named on stage', with two greyed side labels reading '25 entries ranked' and 'ten above 67.78'. The OrcaRouter logo is composited in the bottom-right corner.
Guides & Insights

Microsoft-Decision-1.1 Reached the Ignite Keynote. The Benchmark Line It Quoted Is True and Narrow.

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Microsoft-Decision-1.1 became generally available on 17 November 2026, six weeks after Microsoft-Decision-1 and on a different base model entirely: Gemma 4 26B A4B, the mid-sized mixture-of-experts system Microsoft names in its own launch post, replacing the Qwen3.5-9B that version 1 was post-trained from. That swap is the substance of the release. The line that travelled further came from the Ignite keynote, where Decision 1.1 was introduced as model routing "integrated right into your applications" and then supported with one sentence — "on the AA Decision Index it beats Jev, Google and OpenAI." Those three names resolve to Jev 1.14 at 66.2, Gemini Decision 1 at 66.0 and GPT-6 Luna (Decisions) at 60.2, against 67.78 for Microsoft-Decision-1.1. All three comparisons are correct.

They are also the most flattering four-model subset on a board that currently ranks twenty-five entries. Ten decision models score above Microsoft-Decision-1.1 tonight, from d1.2-ultra at 74.8 through Drex Max v1.1 and SubQ Snap 1.2 at 72.6 down to SubQ Snap 1.1 at 67.9. The keynote sentence was not false. It was a curated comparison presented in the grammar of a complete one, and that distinction is worth more than the launch itself, because it is what happens the first time a benchmark index gets a slide at a hyperscaler's developer conference.

What the stage said, and what the board says

The keynote transcript Microsoft published the same day is short on the subject. After a run through the coding models, it turns to decision-making: "agents are increasingly needed to make decisions, especially where speed and cost is critical. So Decision 1.1 provides model routing integrated right into your applications. And on the AA Decision Index it beats Jev, Google and OpenAI."

Unpacked, that is a four-way comparison the index does support. On the Decision Index v2.1 composite, published by Artificial Analysis and read here on 19 November 2026, Microsoft-Decision-1.1 scores 67.78, Jev 1.14 scores 66.2, Gemini Decision 1 scores 65.98 and GPT-6 Luna (Decisions) scores 60.2. The gaps are 1.6, 1.8 and 7.6 points. Nothing about that is arguable.

What the sentence omits is the half of the board above it. The same page ranks Microsoft-Decision-1.1 eleventh, and the ten entries ahead of it are not obscure:

• d1.2-ultra (Liquid AI) — 74.8, the board leader, at $133 per million tasks

• Drex Max v1.1 (Nace) — 72.6 at $103, and SubQ Snap 1.2 — 72.6 at $137

• GLiDE v1.2 (Fastino) — 72.5 at $780, Mammoth 27B (Baseten) — 71.4 at $264, d1-ultra — 70.6

• Decider v1.3 (Perplexity) — 70.4 at $68, Decider v1.2 — 69.8, d1.2 — 68.7, SubQ Snap 1.1 — 67.9

A vendor choosing the three named competitors that flatter its launch is not a scandal; every launch post does it, and Microsoft is unusually explicit that these are the models it chose to name. The interesting part is the venue. A proprietary third-party index — one whose v2 revision shipped on 8 November 2026, nine days before Ignite, built on 45% private datasets — was quoted on stage as the authority for a claim about Microsoft's own model. That is a different kind of fact from a vendor benchmark screenshot, and it cuts both ways: it raises the index, and it puts the index in a position where being quoted selectively is now part of its reputation.

What actually changed between 1 and 1.1

Strip the slide away and the engineering is specific. Microsoft-Decision-1.1 is not a tweak of the October model. The base changed from Alibaba's Qwen3.5-9B, which the v1 catalogue page still credits, to Gemma 4 26B A4B. Microsoft describes the choice in cost terms: enough active parameters for judgment, small enough to run across A100, H100/200 and GB200/300 fleets. A custom inference engine, built for scoring rather than token-by-token generation, is what the company credits for the latency.

The capability that did not exist in v1 is image input. Microsoft-Decision-1.1 accepts text, JSON, or images — photos, screenshots — and still refuses audio, video, and any form of generated explanation. Question types are unchanged in name: noul for a probability on a yes/no statement, choice for one label out of a fixed set with a probability for every option, score for a position on an ordered scale. The request goes to a path that ends /providers/microsoft/v1/systemone, which is either the least self-explanatory endpoint name shipped this quarter or a hint about where Microsoft thinks this category sits in the stack.

Context is the number to watch if you are migrating: Microsoft documents "inputs of up to 32K tokens", and Artificial Analysis records a 33k window. Version 1 was the same. A decision model that reads a 40,000-token contract is not in this release.

Three claims from the launch post, checked against the board

Microsoft's own post makes three quantitative claims, and the independent data settles each of them differently.

• "Scores 68" — the index value is 67.78. This is rounding, not inflation, and it is the one claim that held exactly.

• "More than twice as fast" — version 1 measures 274.25 ms per task, version 1.1 measures 123.25 ms. That is 2.23x, and it makes the new model the fifth-fastest entry on the board rather than the twentieth.

• "At the same cost" — true at the token level, and misleading in practice. Both versions list $0.042 per million input tokens. But measured cost per million tasks is $139.64 for 1.1 against $142.56 for v1, a 2% improvement, not a halving. Same price, roughly the same bill, twice the throughput. If your constraint is latency rather than spend, the upgrade is large; if it is spend, the upgrade is nearly nothing.

A generated single-column scoreboard headed 'Microsoft-Decision-1.1 — the scoreboard' with six rows: Decision Index 67.78, 11th of 25; When2Call at 10% weight 95.08; AA-Paperclip at 30% weight 58.67; HLE-MultiChoice 6.29% raw accuracy; non-hallucination 49.4%, was 21.7%; speed 123 ms per task, was 274 ms. A footer reads 'All figures per Artificial Analysis Decision Index v2.1, read 19 November 2026.' The OrcaRouter logo is composited in the bottom-right corner.

Where the composite stops agreeing with the marketing

The Decision Index v2.1 composite is not a mystery. It is six weighted components, and Microsoft-Decision-1.1's 67.78 is reproducible from them: 30% AA-Paperclip, 15% EvalSafe-O*NET-AA, 10% When2Call, 15% AA-Omniscience-Decision v1.1, 15% HLE-MultiChoice v1.1, 15% AA-Symmetry. Multiply and add and you land on 67.78 exactly. That transparency is what makes the sub-scores readable, and the sub-scores are where the launch narrative gets interesting.

• When2Call — 95.08. This is the routing test, and it is the strongest number in the set by a wide margin. Microsoft's pitch is "model routing integrated into your applications", and on the one component that measures exactly that, the model is near the top of the board.

• AA-Paperclip — 58.67 Decision Score, 36.8% of documents answered perfectly. This is the 30% component, 1,500 private documents each carrying five decisions and ten facts, scored all-or-nothing. Version 1 passed 20.9%. A fifteen-point jump in perfect-document rate is the single largest engineering gain in the release, and it is also why the composite moved as much as it did.

• HLE-MultiChoice v1.1 — 46.11 Decision Score on 6.29% raw accuracy, with 38.4% non-hallucination. Six percent accuracy on a hard knowledge test is not a strength; the Decision Score is respectable only because the model is honest about being lost. Microsoft's own post says these models are "explicitly not yet tuned for the harder problems", and this row is where that admission shows up in public.

• AA-Omniscience-Decision v1.1 — 60.46 on 33.2% accuracy and 49.4% non-hallucination. Version 1 scored 21.7% on non-hallucination, so the model learned to decline roughly twice as reliably while gaining two points of accuracy. For a decision model, that trade is the whole game.

Read together, the profile is coherent: this is a fast, well-calibrated router with a shallow knowledge base. That is a defensible product. It is not the product the keynote sentence implies.

A screenshot of the Artificial Analysis page for Microsoft-Decision-1.1, headed 'Release, Decision Index Score, Performance & Price'. The headline figures read a Decision Index of 68 at rank 11 of 25, 123 ms per task and about $140 per million tasks. The comparison summary states the model is above average on the Decision Index but somewhat expensive and notably fast, supports text input and carries a 33k-token context window, with board medians of 66, 206 ms and $137. Component rows for Decision Index, AA-Paperclip, EvalSafe-O*NET-AA, When2Call, AA-Omniscience-Decision v1.1 and HLE-MultiChoice v1.1 sit below.

Pricing, in the units a bill actually uses

Foundry lists three input rates for Microsoft-Decision-1.1 and no output rate at all, because the model emits no output tokens: $0.042 per million input tokens on Global Standard, $0.046 on US Data Zone, and $0.05 on EU and APAC Data Zone deployments. Choose the data-zone SKU and you pay 19% more than the headline number for locality. That is the first place a "cheapest decision model" claim gets quietly revised.

The second place is the cost-per-task column, which is what happens when you multiply the token price by how much input a task actually consumes. At $139.64 per million tasks, AA puts Microsoft-Decision-1.1 fifteenth of twenty-five on cost — cheaper than Gemini Decision 1 at $344.90, dearer than Celeris-1.1 Decision at $14 and Decider v1.3 at $68. The model is fast and mid-priced. It is not the bargain in its own category, and Microsoft never claimed it was.

Where this leaves a team that wants to call it

The first-party route is Foundry: deploy the model, then send a state plus a map of typed questions to the decision endpoint. Enterprise buyers will care that Microsoft is already using it as a judge inside Foundry Evaluations and as the router in Copilot, where the company reports the same routing decision made 2.2x faster and at 4.4x lower cost than a large language model — a vendor-reported figure from Microsoft's own harness, unreproduced outside it.

What is worth saying plainly is what OrcaRouter does and does not carry. We do not host Microsoft-Decision-1.1, or Gemini Decision 1, or any decision-scoring endpoint, because a probabilities-only API is not a chat-completions target and pretending otherwise would be the kind of availability claim that reads well and breaks in production. What we do carry is the half of a decision pipeline that does generate — and one window into this release is unusually literal: Gemma 4 26B A4B, the exact base Microsoft fine-tuned, is on our catalogue at $0.06 per million input and $0.33 per million output, alongside Jev 1.14 at $0.042, the model Microsoft-Decision-1.1 is measured against. If you want to see how much of the decision behaviour is the base model and how much is the post-training, both ends of that comparison are reachable from one key, at provider list price with 0% markup passed through.

A screenshot of the OrcaRouter model page for typesafe/jev-1.14, headed with the model name and typesafe as the provider. Pricing shows $0.04 input and $0.04 output per million tokens, with a code sample posting a state and two typed questions to https://api.orcarouter.ai/v1/systemone. A performance panel lists a p50 time-to-first-token of 181 ms, a European Union inference region and zero data retention available, with 171.8M tokens of usage shown.

What to watch next

Two things are now testable in public that were not last week. The first is whether the Decision Index keeps appearing on keynote screens: a benchmark becomes a purchasing document the moment vendors start quoting it under their own logo, and the next vendor to do it will quote a different four-model subset. The second is whether Microsoft-Decision-1.1 holds its routing position when the board grows — When2Call carries a 10% weight and supplies most of the model's lead over Gemini Decision 1, which means a routine index revision can move this ranking more than any model update will.

For now the accurate sentence is narrower than the slide and more useful: Microsoft shipped a decision model that is twice as fast as its predecessor at the same token price, markedly better calibrated, newly able to read images, still capped at 32K tokens, honest about what it does not know, and eleventh on the only independent board that ranks it. The keynote said it beats Google and OpenAI. It does. It just does not beat everyone.