Microsoft-Decision-1 vs Jev: price per decision, latency, accuracy and what switching costs

Accuracy is close, independent tests lean Jev, and option order favours Microsoft

On accuracy, the independent tests split. JevBench puts Jev 1.13 ahead on intelligence, 63.6 to 57.2, and on calibration, 90.6 to 84.4. Artificial Analysis puts Jev ahead, 57.5 to 53.8. Anand found Jev 1.7 points ahead on macro-F1 with a 95% interval of -1.8 to +5.2, "not statistically separable". jev-ai.pro, a site that sells Jev access, found Microsoft-Decision-1 slightly ahead on two tests.

Microsoft-Decision-1 is steadier when the options are reordered. In Anand's harness, it gave the same route across all six option orders 97.2% to 97.6% of the time, against 91.6% to 91.8% for Jev and 64.2% for OpenAI Decisions. Microsoft claims zero flips when options are shuffled, yet its own docs warn that scores "can change based on how you phrase or order questions and options".

Thresholds do not carry over. With the confidence thresholds Anand had tuned for Jev, Microsoft-Decision-1 sent about 50 cases per run to unnecessary manual review, against one or two for Jev. The benchmark post covers how to set them on your own data.

A cheap general model still has a job when the decision needs words

A general model such as Gemini 3.5 Flash-Lite still has a place next to a decision model when the answer has to include arguments or a reason, or when the prompt is loose. Mike Swift of Major League Hacking measured this on 8 October, before Microsoft-Decision-1 launched.

Swift's router picks a deterministic checker for each eval check or hands it to an LLM judge. Out of the box Jev made three dangerous calls on eight real checks and Flash-Lite none. One added sentence took Jev from 35 false-positive runs to 0. Swift still uses Flash-Lite, but behind Jev: a decision model returns the checker's name "And stops", so Flash-Lite writes the arguments only for the checks that need them.

Swift's cost table, per 1,000 checks, billed through OpenRouter: Jev handing off to Flash-Lite only when arguments are needed cost $0.10, Flash-Lite alone $0.52, and Jev alone about three cents. Flash-Lite lists at $0.30 per million input tokens, about seven times the decision models' rate, plus $2.50 per million output tokens. The logprobs post covers getting typed answers from a general model.

Switching between Jev and Microsoft-Decision-1 is a URL change and a threshold retune

Moving a Jev integration to Microsoft-Decision-1 changes the URL, the authentication and the model field, and keeps the request body. Both take a state, a questions map keyed by your names, and the same noul, choice and score types with the same criteria fields. Moving either one to the OpenAI Decisions API rewrites the body.

An example request in the shape both docs describe:

{
  "model": "jev-1.13.0",
  "state": "The API returns 500 on every call.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this ticket?",
      "criteria": {
        "billing": "Payments, invoices and refunds",
        "engineering": "Errors, outages and bugs",
        "support": "How-to questions and account help"
      }
    }
  }
}

What changes for Microsoft-Decision-1:

1. URL. https://api.typesafe.ai/v1/systemone becomes {your Foundry resource}/providers/microsoft/v1/systemone.
2. Auth. A TypeSafe bearer key becomes an Entra ID token or an api-key header.
3. Model. jev-1.13.0 becomes your Foundry deployment name.
4. Limits. Inputs over 32,768 tokens fail, and images are not accepted.
5. Thresholds. Retune them on labelled data; Microsoft's docs say to "validate it against labeled examples before you automate a gate".

Microsoft does not document a sample response or a confidence field, and publishes no limits on questions per request or options per choice. Gateways hide the URL difference: OpenRouter serves all three as typesafe/jev-1.13, microsoft/microsoft-decision-1 and openai/gpt-6-luna-decisions, which is how Anand ran the comparison. The OpenAI Decisions vs Jev post has the body rewrite for OpenAI.

On Stackness, one public profile lists Jev as of 17 October 2026 (mine), and none yet list Microsoft-Decision-1 or the OpenAI Decisions API (data sources). On OpenRouter's decision rankings for the week to 16 October, Jev 1.13 served 1.05 billion requests and Microsoft-Decision-1 19.5 million. The AI tools developers list page shows where the slot sits in real stacks.

Pick Jev if, pick Microsoft-Decision-1 if

Pick Jev if:

• each decision sits in a user-facing path where 100 ms matters
• your inputs run past 32,768 tokens
• you want the most independent benchmark data behind your choice
• you need images, even behind an experimental flag

Pick Microsoft-Decision-1 if:

• you already run on Azure and want Entra ID auth and Azure billing
• inference has to stay in an Azure data zone
• you make millions of decisions a month and the cost per decision decides
• answers should move as little as possible when options are reordered

Pick the OpenAI Decisions API if you need image input without a flag or OpenAI's data terms, as the earlier comparison covers. Pick a general model like Flash-Lite if the decision has to come with arguments or a reason.