Gemini Decision 1 vs Microsoft-Decision-1: Two Hyperscaler Scorers, Four Points Apart

The two largest cloud vendors now both sell a model that answers with a probability instead of a sentence, and neither will tell you how well it works. Microsoft-Decision-1 went generally available on Microsoft Foundry on 8 October 2026: a Qwen3.5-9B base post-trained by Microsoft, text in, 32,768 tokens of context, calibrated distributions out, weights not distributed. Gemini Decision 1 arrived on 11 November 2026 on the Gemini Developer API: text, image, video, audio and PDF in, 262,144 tokens of context, up to 1,000 typed questions per request, $0.10 per million input tokens, weights not distributed either. Both ship a methodology section and no results table, which leaves one independent instrument as the only place they can be compared — Artificial Analysis, where they sit four points apart, ninth and tenth of eighteen.

Four points is not the interesting gap. The interesting gap is that on the evaluation most decision models are actually bought for, the two are only five points apart while both sitting near the bottom of the field, and the model that wins the overall ranking does it somewhere else entirely.

What each one puts in front of you

The request shapes are close enough that a client written for one adapts to the other in an afternoon, and different enough that the choice is usually made before any benchmark is read.

• Input — Gemini Decision 1 takes text, image, video, audio and PDF, up to 100 MB inline per request and 15 MB per external URL. Microsoft-Decision-1 takes text, serialised to 32,768 tokens. A scanned form or a call recording is native input for one and a preprocessing job for the other.

• Context — 262,144 tokens for Gemini Decision 1, shared across the input and every question. 32,768 for Microsoft-Decision-1, with the input plus the option set consuming it and no generation budget to subtract because there is no generation.

• Question shapes — Gemini Decision 1 formalises three: yes/no bounded to 0.01–0.99, choice over 2 to 255 named options, and score over 2 to 255 ordered levels returning a probability-weighted expected level. Microsoft-Decision-1 covers yes/no, multiple choice, rating, classification and rubric scoring, and explicitly supports abstention options such as "cannot tell" inside the option set.

• Batching within a request — up to 1,000 independent questions for Gemini Decision 1, each isolated from the others, all against one billed read of the shared input. Microsoft-Decision-1's published contract is one state and its question set per invocation, so breadth comes from grouping questions rather than from a documented per-request ceiling.

• Price — $0.10 per million input tokens for Gemini Decision 1, with no output charge because output is zero tokens. The board records $345 per million tasks for it, and $143 per million tasks for Microsoft-Decision-1 at the input rate it tracks.

• Access — Gemini Developer API at v1beta only, with a free tier at 200 requests per minute and 10,000 per day, unlimited daily requests from Tier 2, and no Vertex AI listing in the documentation Google shipped. Microsoft-Decision-1 as a Foundry deployment under the Direct from Azure portfolio, with serverless and unified-endpoint options, standard SKU, Azure authentication and one billing path, batch inference disabled, and no weight download or fine-tuning path.

The access line is where procurement decisions get made rather than technical ones. If your organisation buys through an Azure commitment, Microsoft-Decision-1 is an invoice you already know how to pay. If it buys through Google Cloud, Gemini Decision 1 currently is not — the only door documented is a Gemini API key, which is a different budget line and, for some enterprises, a different approval process.

The scoreboard, component by component

Artificial Analysis's Decision Index scores each evaluation as accuracy minus the Brier penalty for the confidence attached to the answer, weighted across six components. Here is how the two models split them:

• AA-Paperclip, 30% weight — Gemini Decision 1 56.8 against Microsoft-Decision-1 51.8. Both are weak. The all-or-nothing document pass rate is 33% for Gemini Decision 1 and 20.9% for Microsoft-Decision-1, against 63% for the field leader. Neither hyperscaler model extracts fifteen facts from a dense document reliably.

• EvalSafe-O*NET-AA, 15% — 87.1 against 86.8. Effectively identical, and both near the top of a tightly clustered field.

• When2Call, 10% — 80.5 against 88.0. Microsoft-Decision-1 wins this one clearly and sits near the leaders; Gemini Decision 1 is fourteenth of eighteen.

• AA-Omniscience-Decision v1.1, 15% — 66.3 against 54.9. Gemini Decision 1 takes first place in the field; its 49.6% non-hallucination rate is the reason. Microsoft-Decision-1 reports 21.7% non-hallucination here, which is the number that should worry anyone planning to threshold its abstentions.

• HLE-MultiChoice v1.1, 15% — 48.7 against 39.4. First place against eleventh. Raw accuracy is 13.2% against 8.6%, and non-hallucination 35.7% against 15.1%.

• AA-Symmetry v1.1, 15% — 70.5 against 70.4. A dead heat, both mid-table, holding the same answer under reordering and negation on 38.3% and 51.7% of question families respectively.

The composite lands at 65.98 and 62.07, ninth and tenth. Read the components and the shape of each product becomes clear. Microsoft-Decision-1 is a solid tool-call and workplace-judgement scorer whose knowledge and abstention behaviour is weak. Gemini Decision 1 is the reverse: the best knowledge and abstention scores on the board, paired with mid-table extraction and poor tool-choice discrimination.

What they share is the thing to be suspicious of

Neither vendor publishes a benchmark table. Microsoft-Decision-1's card describes its methodology — public and community decision benchmarks, held-out internal sets, accuracy and calibration error, option order varied, paired statistical tests — and claims the model "performs on par with leading decision models and ahead of other open decision models evaluated with the same methodology". There is no table under that sentence. Gemini Decision 1's release notes announce the model and link the API documentation, which lists supported and unsupported capabilities and no accuracy figure at all.

This matters because a decision model's entire interface is a number your code branches on. A chat model that is a few points weaker produces worse prose; a scorer that is miscalibrated silently routes 8% of your tickets to the wrong queue, and nothing in the response tells you. Two frontier vendors shipped products whose central claim — that the probability means what it says — is asserted in prose and measured only by a third party.

The third party, to its credit, measures both on the same instrument, which is why this comparison is possible at all. But one board is one board, its largest component is 45% private data, and it has been scoring decision models for a matter of weeks. Treat 66 and 62 as the first outside reading, not as settled capability.

Which one to put behind your threshold

Choose Microsoft-Decision-1 when the decision is a tool call or a workplace-document judgement, when the input is text that fits comfortably in 32K tokens, and when the deployment gate is an Azure contract with unified billing and a Responsible AI package attached. Its 88.0 on When2Call and 86.8 on EvalSafe-O*NET are real strengths, and its explicit abstention options let you design a "cannot tell" escape hatch into the option set rather than bolting one on afterwards.

Choose Gemini Decision 1 when the input is long, multimodal, or shared across many questions, when knowledge questions and hallucination control are the point, and when you are prepared to pay for context. Its first-place finish on both knowledge components and its 49.6% non-hallucination rate on AA-Omniscience are the reasons to pick it over the cheaper Microsoft model, and its 33% document pass rate is the reason not to pick it for extraction work.

Choose neither for extraction at volume. Both hyperscaler scorers sit in the bottom half of the 30%-weighted document component, and the models that lead there — a 27B multimodal from Nace.AI and a 421M encoder from Convai Innovations you can fine-tune yourself — cost less per task and, in the open-weight case, can be trained on your own documents.

The rest of the pipeline does not have to be hostage to this

OrcaRouter carries neither model. Both are probabilities-only endpoints — one on the Gemini Developer API, one on Foundry — and neither is a chat-completions target, so we do not list them. What the catalogue does carry is TypeSafe's typesafe/jev-1.14, a decision model at 66.23 on this same index, which is a reasonable third option if you want the scoring call on the same key as everything else.

The call that surrounds the scorer is where a router earns its place, and it is the half these two products deliberately do not address. Something has to write the rubric the scorer grades against, summarise the 60-page document down to 32K tokens, or draft the candidate answer being ranked — and that something is a generative model whose price and quality move every few weeks. That call runs across one endpoint with more than 200 models on it at provider list price passed through at 0% markup, with automatic failover when a provider degrades and a routing DSL if you want several models composing one answer. Keep the writing half portable and the scoring decision stays reversible, which is worth more than four index points when both vendors revise on a quarterly cadence.

What to watch

Three signals would change the ranking, and none has arrived. A Microsoft benchmark table with calibration curves would let its Azure customers set thresholds on evidence; the card currently offers a methodology and a claim. A Vertex AI listing for Gemini Decision 1 would move it from an API-key purchase to a cloud-contract purchase and change the procurement comparison entirely. And a second independent board scoring both models would tell us whether the four-point gap is signal or the noise of a leaderboard three weeks old.

Until then, the defensible read is this: Gemini Decision 1 is the better model on the two components where knowing when to decline is the skill, Microsoft-Decision-1 is the better one on tool selection and marginally cheaper, and both are mid-table on the document extraction that most enterprise decision workloads actually consist of. Run 200 of your own labelled items through each before either number decides anything.