Jev vs Microsoft-Decision-1

Unofficial. Differences only, from sources opened for this page. Not a ranking, and not affiliated with TypeSafe or Microsoft.

Documented differences between TypeSafe Jev and Microsoft-Decision-1: positioning, capabilities, use cases, price, and access. No ranking.

This note lists differences first. It does not say which product is better. If a field is not in a source opened for this page, it is marked not published. Benchmark claims below are each vendor’s own numbers, not tests run by this site.

What Microsoft-Decision-1 is

Microsoft-Decision-1 is a hosted decision-scoring model. Given a situation and a question with a fixed set of answer options, it returns a calibrated probability for each option instead of generated text (Foundry model card). Microsoft announced it on 2026-10-09 (Microsoft announcement).

Differences

Capability boundaries

Microsoft-Decision-1

From the Microsoft Learn guide, the Foundry model card, and the announcement:

• Three question types: noul (probability 0 to 1), choice (one option plus a probability per option), and score (ordered levels, lowest first; the result is the probability-weighted average of level indexes).
• criteria is optional for noul. Microsoft suggests an abstention option such as cannot tell when a forced choice is wrong.
• The model card says it is not designed for text generation, open-ended question answering, conversation, translation, or summarization.
• Known limits listed by Microsoft: possible bias from the base model, sensitivity to wording and option order, calibration strongest on familiar task types, and missed or over-flagged content when used as a safety filter.
• Vendor claims: highest accuracy in Microsoft’s 36-benchmark comparison (nearly 150,000 questions), and P50 latency about 35 times faster than GPT-6 Sol. The announcement’s editor’s note says Jev benchmarks on accuracy and calibration were added later. This site has not reproduced these numbers.

TypeSafe Jev

From the Introduction, System One, API reference, and Models pages:

• Three question types can be mixed in one call: Choice (maximum 255 options), Score (at least two levels, API accepts up to 10), and Noul (0 to 1).
• Choice returns choice, probabilities, and confidence. Score returns score, legend, probabilities, and confidence. Noul returns noul.
• The same weights serve every account. Jev is not fine-tuned with customer data.
• English is the primary training language. The models page says to test non-English content.

Use cases

Both vendors list routing, classification, grading, and gating. That overlap is not a ranking.

Microsoft lists gates and filters, classification and routing, ranking and prioritization, evaluation and grading, and model, tool, or agent selection (Learn). The announcement adds agent controls, data labeling, intent analysis, incident routing, code scanning, safety screening, computer and UI use, robotics, and scientific discovery. It also describes internal tests by Xbox Research (labeling over 10,000 pieces of feedback), the Copilot team (response quality), on-call incident response, and Microsoft Discovery.

TypeSafe’s use-case map lists, among others, model routing, LLM guardrails, universal verification, search and retrieval, moderation, customer support, and scientific discovery.

Price

Both list $0.042 per million input tokens with free output tokens: Jev on its Models page, Microsoft-Decision-1 in its announcement. Azure billing terms beyond that line were not opened for this page.

How to choose

These are questions to ask, not a recommendation.

• Already on Azure? Microsoft-Decision-1 runs as a Foundry deployment with Entra ID auth and data-zone options. Jev needs a separate TypeSafe account.
• Long state? Jev documents a 64k request limit; Microsoft-Decision-1 lists a 32,768-token context.
• Want an SDK? Jev ships Python and JavaScript SDKs. Microsoft’s quickstart uses plain HTTP.
• Many options per choice? Jev documents up to 255 choice options. Microsoft’s limit is not published on the pages opened.
• Code portability: both use state with noul, choice, and score questions at a /v1/systemone path. Endpoint, auth, and the model field differ, so test before switching.
• Either way: validate on your own labeled data. Both vendors say so.

Related on this site

• Compare
• Jev vs OpenAI Decisions API
• Jev vs Kev
• What is Jev?

Sources

Opened before the claims above were written (2026-10-14):

• Introducing Microsoft-Decision-1 (Microsoft)
• Deploy and use Microsoft-Decision-1 (Microsoft Learn)
• Microsoft-Decision-1 model card (Microsoft Foundry)
• TypeSafe Introduction
• System One
• Example use cases
• API reference
• Models