A generated title card headed 'One Test, Most of the Lead', subtitled 'Microsoft-Decision-1.1 vs Gemini Decision 1, Decision Index v2.1', with a two-by-two grid of four label cards reading '67.78 vs 65.98', '95.08 vs 80.50', '123 ms vs 211 ms' and '139.64 vs 344.90 dollars', a caption reading 'per million tasks, read 19 November 2026', and two grey bar blocks labelled Microsoft and Gemini. The OrcaRouter logo is composited in the bottom-right corner.
Guides & Insights

Microsoft-Decision-1.1 vs Gemini Decision 1: A Two-Point Lead Built Almost Entirely on One Test

Author

jinhao.song

Date Published

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

Put Microsoft-Decision-1.1 and Gemini Decision 1 side by side on the Artificial Analysis Decision Index v2.1 and the gap looks like a verdict: 67.78 against 65.98, with Microsoft's model also nearly twice as fast at 123 ms per task versus 211 ms, and less than half as expensive at $139.64 per million tasks versus $344.90. Take the composite apart and the verdict gets smaller and stranger. Microsoft-Decision-1.1's entire advantage comes from one component worth 10% of the index and one process-level decision about how a document is billed. On the two knowledge components, Google's model wins — one of them by more than six points, with nearly twice the raw accuracy.

That split is the whole comparison, and it decides more than a leaderboard position does. Gemini Decision 1 reads 262,144 tokens and accepts audio, video and PDF; Microsoft-Decision-1.1 reads 32,768 and accepts text, JSON and images. Google's model admits when it does not know at 49.6% on the Omniscience test; Microsoft's at 49.4%. On the routing test, Microsoft's is twelve and a half points ahead. If your decision is "which of these five tools does this request need", the answer is not close. If your decision is "is this clause in this 90-page contract true", the answer flips, and it flips because of a context window rather than a model.

Both are scorers, and they are scorers in the same shape

The two APIs converged on nearly the same primitive, which makes the comparison unusually clean. Each takes a shared input, a set of typed questions, and returns one typed answer per question with a probability attached. Neither writes a word.

• Question types — both expose the same three: a yes/no probability, a choice from a fixed set with a probability per option, and a position on an ordered scale. Microsoft calls the first noul, Google also calls it noul, which is not a coincidence so much as a category settling on vocabulary.

• Output tokens — zero on both, on every call. Billing is input-only, so the token price you read is the whole price.

• Parallel scoring — both evaluate every question against the shared input independently, so questions cannot see each other and adding a tenth does not change the first nine.

• Request size — Google documents 1 to 1,000 questions per call, choice questions with 2 to 255 options, score questions with 2 to 255 levels. Microsoft documents "one or more typed questions" without a ceiling, and its launch post describes requests carrying "hundreds of questions".

• Image input — both accept it now. Microsoft added it in 1.1; Google has it from launch, along with audio and video, which Microsoft explicitly does not take.

• Endpoints — this is where they diverge physically. Gemini Decision 1 is reachable only through Google's Decisions API on the Gemini Developer API v1beta. Microsoft-Decision-1.1 deploys as a Foundry resource and answers at /providers/microsoft/v1/systemone on that resource.

The composite, decomposed

A 1.8-point index lead that sits in one 10%-weighted component is not a rounding artefact, but it is a much narrower claim than "better model". Here is the full decomposition of both entries, as Artificial Analysis publishes it, read 19 November 2026:

• Overall index — Microsoft-Decision-1.1 67.78 vs Gemini Decision 1 65.98

• Accuracy index (the index rebuilt with accuracy alone, confidence removed) — 48.45 vs 45.53, so Microsoft leads on correctness too, just by less

• When2Call, 10% weight — 95.08 vs 80.50. A 14.6-point gap and the source of the lead

• AA-Paperclip, 30% weight — 58.67 vs 56.83 Decision Score, and 36.8% vs 33.0% of documents fully correct

• EvalSafe-O*NET-AA, 15% — 88.85 vs 87.05, with 83.4% vs 82.0% raw accuracy

• AA-Omniscience-Decision v1.1, 15% — 60.46 vs 66.28, with 33.2% vs 44.8% raw accuracy and 49.4% vs 49.6% non-hallucination

• HLE-MultiChoice v1.1, 15% — 46.11 vs 48.71, with 6.29% vs 13.16% raw accuracy and 38.4% vs 35.8% non-hallucination

• AA-Symmetry, 15% — 75.72 vs 70.47, on rewording pass rates of 47.5% against 38.3%

Two things fall out of that list. The first is that the largest single gap is on the smallest-weight component — When2Call contributes 9.5 points to a 100-point scale, so the 14.6-point difference there is worth about 1.5 index points, which is most of the 1.8-point total. The second is that Gemini Decision 1's knowledge advantage is genuinely large: 44.8% against 33.2% on Omniscience raw accuracy is not noise, and it is why Google's model has a higher accuracy index than its composite suggests.

In other words, Microsoft did not out-measure Google across the board. It out-measured Google decisively on the one test that asks whether the model knows when to call a tool, and roughly matched it everywhere else, losing the knowledge tests and winning the consistency one.

A generated two-column scoreboard headed 'Microsoft-Decision-1.1 vs Gemini Decision 1 — the scoreboard'. The Microsoft column reads Decision Index 67.78, When2Call 95.08, Omniscience accuracy 33.2%, HLE accuracy 6.29%, context 32,768 tokens and cost per million tasks $139.64. The Gemini column reads Decision Index 65.98, When2Call 80.50, Omniscience accuracy 44.8%, HLE accuracy 13.16%, context 262,144 tokens and cost per million tasks $344.90. 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.

What the context window difference actually costs

The 262,144-token window is the number that decides real workloads, and it is an 8x gap: Gemini Decision 1 reads eight times what Microsoft-Decision-1.1 reads. Google's documentation spends that budget in a specific way — the window covers the shared input and all questions together, so a single long contract plus two hundred rubric questions competes for the same 262K.

Microsoft's 32K ceiling has a second-order effect that shows up in the cost column. AA-Paperclip charges a document once for all its questions and resolves 15 answers per document. A model that cannot hold the document in one call has to be applied chunk by chunk, which multiplies calls. Microsoft-Decision-1.1 comes out at $386 per million tasks on Paperclip against Gemini Decision 1 at $946 — but the ranking on total cost per task is already explained by token price, not by document handling.

There is a third mechanism worth naming, because it is the one buyers miss: repeated input counts in full on both. Google's Decisions API explicitly does not support context caching, Batch, Flex or Priority inference, so a rubric sent with every request is paid for with every request. Microsoft-Decision-1.1 has no caching story in its documentation either. Neither vendor is offering a discount for asking the same 200 questions a million times, which is exactly the workload this category is built for.

Pricing, with the SKU trap

Microsoft quotes three input rates — $0.042 per million on Global Standard, $0.046 on US Data Zone, $0.05 on EU and APAC Data Zone — and no output rate. Google quotes one: $0.10 per million input tokens. On the headline number Microsoft is 58% cheaper. On the measured cost across the whole index Microsoft is 60% cheaper — $139.64 against $345.41 per million tasks — because it is also consuming fewer tokens per task on most components.

But the $0.042 rate is the global one, and an EU-resident workload on the data-zone SKU pays $0.05, which narrows the gap to 50%. Neither number includes the cost of the calls around the decision model, which for most production pipelines dwarf the scoring call itself — the model that drafts the candidate answers, the model that writes the rubric, the model that handles whatever the scorer abstained on. That spend is where routing actually saves money, and it is the part OrcaRouter covers: one key across 200-plus models at provider list price with 0% markup, so a vendor price change on the generative half lands the same day, and fusion or the routing DSL can put a cheap drafter behind an expensive judge without a second integration.

A screenshot of the OrcaRouter model page for google/gemma-4-26b-a4b-it, headed 'Gemma 4 26B A4B' with a 262K-token context badge and input modalities listed as text, image and video plus vision, tools, JSON and reasoning tags. The provider line reads Google, 2026-04-05, and the description identifies it as an instruction-tuned mixture-of-experts model from Google DeepMind with 25.28 total parameters and 3.8B active per token. Pricing reads $0.06 input and $0.33 output per million tokens, with zero data retention available and 76.0M tokens of usage.

Which one you would actually deploy

Three shapes of workload, three different answers, and none of them is "the higher index".

• Routing, gating and tool selection — Microsoft-Decision-1.1. A 95.08 on When2Call against 80.50 is the largest single-component gap on the board, and routing is the workload Microsoft built the model for. The 123 ms median also matters here: a router sits in the request path, and the difference between 123 ms and 211 ms is the difference between a step nobody notices and a step the latency budget has to absorb.

• Document extraction and rubric grading over long inputs — Gemini Decision 1. If a single call has to see a 90-page contract plus the questions about it, Microsoft's 32K ceiling simply fails, and no amount of When2Call performance rescues a call that never happened. Google's 44.8% Omniscience accuracy and 13.2% HLE accuracy reinforce the same conclusion from a different direction: it knows more, and it can read more.

• High-volume screening where the bill dominates — neither, and this is the honest answer. Celeris-1.1 Decision scores 62.0 at $14 per million tasks; Decider v1.3 scores 70.4 at $68. Both beat Microsoft-Decision-1.1 on cost per task and one beats it on index. If your decision is cheap and repetitive, the hyperscaler options are the expensive ones.

One caveat applies to all three: the 1.8-point gap and the calibration rows come from an index that revised its own weighting on 8 November 2026, moving Omniscience and HLE down from 25% each to 15% and introducing a 30% document component. A re-weighting can move this ranking further than a model release will. Treat the comparison as current and dated, not permanent.

A screenshot of the Artificial Analysis page for Gemini Decision 1, headed 'Release, Decision Index Score, Performance & Price'. The headline figures read a Decision Index at rank 14 of 25, 211 ms per task and $345 per million tasks. The comparison summary states the model is below average on the Decision Index and particularly expensive, slower than average, supports text and image input and carries a 262k-token context window, with a board median of 206 ms. Component rows for Decision Index, AA-Paperclip, EvalSafe-O*NET-AA, When2Call, AA-Omniscience-Decision v1.1 and HLE-MultiChoice v1.1 sit below.

Where the two models will diverge next

Microsoft-Decision-1.1 is a second attempt on a new base — Gemma 4 26B A4B instead of Qwen3.5-9B — with a custom inference engine and a routing pitch. Gemini Decision 1 is Google's first entry, bound to a single API surface and a single model ID. The trajectories point in opposite directions: Microsoft is iterating on the base model and the serving stack, Google is iterating on the input modalities and the context budget.

What that means for a decision made this month is unglamorous. Check whether your longest input fits in 32K before you read a single benchmark number, because that constraint eliminates one of these two models outright. If it fits, run your own twenty labelled examples through both and compare the confidence distributions rather than the top-1 answers — a scorer you cannot threshold is a scorer you cannot ship, and the two are within 0.2 points of each other on non-hallucination, which is the row that governs whether abstention behaves the same way on your data as it did on a public test set.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily