A generated title card headed 'Karpo Decision 31B vs Microsoft-Decision-1' with the line 'One has a leaderboard score and no price. The other has an Azure price list and no published score.' Below, two cards: 'Karpo Decision 31B' listing '73.86 Full score', 'no price published' and 'Apache-2.0 weights'; and 'Microsoft-Decision-1' listing '62.07 on Artificial Analysis', 'Azure API only' and 'no benchmark table'.
Guides & Insights

Karpo Decision 31B vs Microsoft-Decision-1: One Has a Price, the Other Has a Score

Author

Gideon Frost

Date Published

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

Put Karpo Decision 31B next to Microsoft-Decision-1 and the first thing you notice is that neither company has given you the thing you need to choose between them. Karpo Decision 31B, the 31.3-billion-parameter Gemma 4 fine-tune that MachinePulse pushed to Hugging Face on 5 November 2026 without a launch post, has a leaderboard number and no price. Microsoft-Decision-1, on Microsoft Foundry since 8 October 2026, has a price, an Azure contract, a responsible-AI packet and no published benchmark table at all. One is measurable and unsellable. The other is sellable and, by its vendor's own choice, unmeasured.

That inversion is the whole comparison, and it is worth resisting the urge to resolve it with arithmetic. The 73.86 that Karpo carries and the 62.07 that Artificial Analysis assigns to Microsoft-Decision-1 are not two readings of one thing. They come from different boards, different task mixes, different scoring formulas and different ideas of what a decision model is for. The useful question is not which number is bigger. It is which of the two gaps — the missing price or the missing score — your deployment can actually absorb.

What Microsoft is selling, exactly

Microsoft-Decision-1 is a post-trained Qwen3.5-9B that answers with a distribution over options instead of prose. The contract is narrow: yes/no, multiple choice, a rating scale, a classification set, or a rubric, over up to 32,768 tokens of text. No images, no audio, no explanation field, no generated tokens. Microsoft is explicit that it is not for open-ended questions, translation or summarisation, and it recommends against using it as the sole basis for decisions about credit, employment, housing, insurance, education or healthcare.

What it does bring is the part an open checkpoint cannot: a Foundry deployment under the Direct from Azure portfolio, serverless or unified-endpoint, standard SKU, Azure authentication, one invoice, batch inference deliberately switched off, and no weight download or fine-tuning path. For a bank or a hospital group, that last cluster of facts is often worth more than twenty points on a leaderboard, because it is the difference between a model a platform team can approve and a model a platform team has to build a serving stack around.

The price is real but awkward to quote. The model page does not print a per-token rate; it links out to Azure's pricing surface. What Artificial Analysis tracks for it is $143 per million tasks at the input rate, against $345 per million tasks for Google's Gemini Decision 1. Karpo has no comparable figure because Karpo has no rate card of any kind — no host, no console, no trial, no published per-GPU cost.

What the board measured, and what it did not

A generated two-panel diagram titled 'Two numbers, two instruments'. The left panel is headed 'Karpo Decision 31B — Decision Index 0.4.1' and lists 'Full score 73.86 — private tests, board-computed', 'Same-skills 70.94', 'New-domains 61.5' and 'Public index 75.05 — vendor run, all 140,178 requests, 11 refused'. The right panel is headed 'Microsoft-Decision-1 — Artificial Analysis Decision Index v2.1' and lists 'Composite 62.07', 'When2Call 88.0', 'AA-Paperclip 51.8 — 20.9% document pass' and 'HLE-MultiChoice 39.4 — 8.6% raw'. The footer reads 'The two panels are different instruments and are not subtractable.'

Karpo's number is a community one. The Decision Index, the open board that tracks reproductions of TypeSafe's Jev, puts it at 73.86 Full in edition 0.4.1 — first equal of 307 rows in the board's published dump, tied with GLiDE 27B at 73.37 because the board groups anything within 0.9 points of the leader. That Full score is third-party in the sense that matters: the board's maintainers ran private tests the vendor never saw, split roughly 60% same-skills and 40% new-domains. Karpo's split reads 70.94 on the familiar skills and 61.5 on the unfamiliar ones, which is the ordinary shape for a model that has been tuned hard against a known suite.

The 75.05 on the model card is a different animal, and the card is honest about it: that is MachinePulse's own run of the official scoring kit, 140,178 scored requests with eleven refused for exceeding the window, adopted by the board after a kit rescore came back equal to the claim. Public and reproducible, still first-party. The board's own row confirms the adoption and notes that the private-test agreement with the submitted rows was 100% across 827 rows.

Microsoft-Decision-1 has none of this. Its catalogue page carries a Benchmarks tab containing a methodology paragraph and no figures: public and community decision benchmarks, held-out internal sets, accuracy and calibration error, option order varied, paired statistical tests applied. Microsoft states the model performs "on par with leading decision models and ahead of other open decision models evaluated with the same methodology" and then declines to show the evaluation. Every performance claim about it today is vendor-reported and unreproduced.

The one instrument that has scored it is Artificial Analysis, on Decision Index v2.1, and the components are more informative than the composite. Microsoft-Decision-1 scores 88.0 on When2Call, near the leaders, and 86.8 on EvalSafe-O*NET. It scores 51.8 on AA-Paperclip with a 20.9% all-or-nothing document pass rate, 54.9 on AA-Omniscience-Decision with a 21.7% non-hallucination rate, and 39.4 on HLE-MultiChoice with 8.6% raw accuracy. The pattern is a competent workplace-judgement scorer with weak knowledge behaviour, and the abstention figure is the one to sit with: a model that declines to answer 78% of knowledge questions is protecting its calibration at the cost of being useful.

The request shapes, side by side

A generated two-column comparison titled 'The request shapes, side by side'. The Karpo Decision 31B column reads '31.3B bf16 plus fp32 readout — about 62.5 GB', 'text, JSON and images', '32,768 tokens, split by question, refuses overflow', 'choice up to 255 options, noul, score', '94 ms median per request' and 'Apache-2.0 weights'. The Microsoft-Decision-1 column reads '9B dense, API only, no weights', 'text only', '32,768 tokens, no way to raise it', 'yes/no, multiple choice, rating, classification, rubric', '274 ms per task' and 'no published rate on the model card'. A footer line reads 'Latency figures are not the same measurement: per request for one, per task over a document set for the other.'

• Parameters — Karpo 31.3B in bf16 plus a 255 × 5,376 fp32 readout, about 62.5 GB on disk; Microsoft-Decision-1 a 9B dense base served only as an API.

• Input — Karpo takes text, JSON and images, with a screenshot passed as state.viewport_screenshot; Microsoft-Decision-1 takes text only, serialised into 32,768 tokens.

• Context — 32,768 tokens for both, with an important difference in failure mode: Karpo splits a long request by question and refuses anything that still will not fit, while Microsoft's ceiling is the input plus the option set with no way to raise it.

• Question types — Karpo choice up to 255 options, noul for true/false, score for an ordered level; Microsoft yes/no, multiple choice, rating, classification and rubric, with abstention options such as "cannot tell" supported inside the set.

• Batching — Karpo encodes one state and many questions in a single sequence, so a hundred questions about one invoice cost roughly one pass over the invoice; Microsoft's published contract is one state and its question set per invocation, with breadth coming from grouping.

• Latency — 94 ms median per request on the tracker's reference hardware, one GPU, one request at a time for Karpo; 274 ms per task for Microsoft-Decision-1 on AA's measurement, a per-task figure over a document set rather than a per-request one.

• Licence and access — Apache-2.0 weights with Gemma 4's terms alongside for Karpo; no weights at all for Microsoft-Decision-1, which is API-only under Azure terms.

• Price — no published rate for Karpo; $143 per million tasks on AA's tracking for Microsoft-Decision-1, with the per-token input rate held on Azure's pricing page rather than the model card.

Calibration is the axis where both have something to say

For a model whose entire output is a probability, the number that governs deployment is not accuracy but whether the probability means what it says. Karpo's card reports a calibration error of 4.8 points: across 34,411 scored responses its mean confidence ran 87.96% against an accuracy of 83.19%, and in the top confidence bucket — 68.5% of all responses — it was right 95.7% of the time while claiming 98.7%. That is a model you can threshold, with a known drift at the top end.

Microsoft publishes no expected calibration error, no Brier score and no reliability curve. What AA's component scores imply is a model whose confidence is conservatively spent: the high non-hallucination rates on Omniscience and HLE mean it abstains rather than answers when it is unsure, which is the behaviour a guardrail wants and a classifier does not. Whether its 0.8 means 0.8 outside Microsoft's own eval is unknown, and Microsoft's documentation says so in its own way by telling integrators to validate on representative data and set thresholds from the cost of their errors.

That asymmetry has a practical consequence. With Karpo you inherit a published calibration curve and can set a starting threshold before you run anything. With Microsoft-Decision-1 you inherit a methodology and a promise, and your first two weeks are a measurement project. Neither is disqualifying; they are simply different kinds of work.

The version nobody's title mentions

Microsoft-Decision-1 is no longer the current model in its own family. Microsoft-Decision-1.1 arrived on 17 November 2026 at Ignite on a Gemma 4 26B A4B mixture-of-experts base, at the same price and the same 32K ceiling, and Artificial Analysis scores it 67.78 — 5.7 points above version 1, at 123 ms per task instead of 274 ms, with non-hallucination on Omniscience up from 21.7% to 49.4%. It also got worse at the things a scorer is supposed to be stable at: 6.29% raw accuracy on HLE-MultiChoice against 8.64%, and 47.5% of AA-Symmetry question families answered consistently against 51.7%.

Anyone comparing Karpo to "Microsoft's decision model" in late November is therefore comparing it to a superseded version. That does not make the matchup moot — version 1 is still what a great many Foundry deployments were sized against — but it does mean the number to beat on the Microsoft side is 67.78, and the second-order question is whether a model that trades knowledge accuracy for abstention discipline is the right shape for your pipeline in the first place.

What each side cannot give you

Karpo cannot give you a contract. There is no endpoint, no rate card, no support path, no SLA and no vendor statement beyond a model card and a README. If your organisation cannot run a 62.5 GB checkpoint on its own hardware, the comparison ends there regardless of the score. Loading it as an ordinary chat model will not reproduce its answers either: the branch attention and the 255-code readout are both required, and the Python package and the POST /v1/systemone server exist precisely because a generic inference stack will not do.

Microsoft-Decision-1 cannot give you a number you did not take on trust, cannot be fine-tuned on your own decision data, cannot read a screenshot, and cannot be run offline. It also cannot be benchmarked by you on the same instrument as Karpo, because the Decision Index is a register of open reproductions and a closed API model has no weights to reproduce. The two models will keep missing each other: one board will never list Microsoft's scorer, and Artificial Analysis has no row for Karpo.

There is a third position that is more useful than either column. If what you actually need is a decision layer next to a generative pipeline, the generator is the half you can already standardise. OrcaRouter does not host Karpo Decision 31B or Microsoft-Decision-1, and neither belongs behind a chat-completions route. What it does host is the two models this comparison keeps circling: Gemma 4 31B, the base Karpo was fine-tuned from, and TypeSafe's Jev 1.14, the model the open-reproduction board exists to chase. Both sit behind one OpenAI-compatible key alongside 200+ other models at provider list price passed through with no markup, so a vendor price cut lands on our side the same day, and automatic failover keeps a scoring loop alive when one provider wobbles.

Which one to reach for

A generated decision card titled 'Which one to reach for'. The left card is headed 'Microsoft-Decision-1.1' with the subtitle 'when the estate is Azure' and the rows 'approval trail, one invoice', 'text that fits in 32K', 'budget a calibration study first' and '67.78 on Artificial Analysis'. The right card is headed 'Karpo Decision 31B' with the subtitle 'when you own the hardware' and the rows 'screenshots and JSON inputs', 'Apache-2.0 weights', 'published 4.8-point calibration error' and '73.86 Full score, private tests'. A wide strip underneath reads 'Do not choose on 73.86 versus 62.07. Different boards, different suites, no shared grading.'

Take Microsoft-Decision-1, or more sensibly Microsoft-Decision-1.1, if the decision layer has to live inside an Azure estate, if you need an approval trail more than a leaderboard position, and if your inputs are text that fits in 32K. Budget for a calibration study, because Microsoft has not given you one, and set your abstention thresholds from the cost of your errors rather than from the vendor's description of them.

Take Karpo Decision 31B if you have the GPU, if screenshots or JSON state are part of the job, if Apache-2.0 matters to a product you intend to ship, and if a third-party private-test score plus a published calibration curve is the evidence your team accepts. Accept in exchange that you are the vendor now: no support, no SLA, no price to budget against, and a model two weeks old whose repository was still being edited on 16 November.

Do not take either on the strength of 73.86 versus 62.07. Those numbers were produced by organisations that have never graded the same paper, on suites with different private-test shares and different scoring formulas, and the honest reading is that both models are plausible and neither is proven against the other. The test that would settle it — Karpo appearing on an independent board that also scores Microsoft's model, or Microsoft publishing the results table its methodology section describes — has not happened. Until it does, the decision turns on procurement, hardware and input modality, which is where it should have turned in the first place.