Closed SourcePreviewTypeSafe AIReleased 2026-09

Jev

Structured decision model for programmatic automation with calibrated probabilities

Context window

32K

Input / 1M tokens

$0.042

Output / 1M tokens

Free

Provider

TypeSafe AI

Data verified 2026-09-16

Jev is TypeSafe's first System One Model designed for fast, structured decisions inside software systems. Rather than generating text token-by-token, it evaluates typed questions in parallel against input state and returns constrained outputs (choices, scores, or yes/no probabilities) with calibrated confidence scores. Optimized for automation workflows requiring high-volume, repeated semantic decisions.

Capability index

Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.

Coding
0
Reasoning
3
Math
0
Multimodal
0
Long context
2
Speed
9
Cost efficiency
10

How to access it

Early access via waitlist at typesafe.ai. Access through Python/TypeScript SDKs or HTTP API with API key.

Strengths

  • 70-500ms latency (20-200x faster than LLMs)
  • Extremely low cost ($0.042 per million input tokens)
  • Returns typed, schema-constrained outputs eliminating hallucinations
  • Calibrated confidence scores for autonomous decision-making
  • Parallel evaluation of multiple questions from single input

Best for developers who...

High-volume classification and routing decisionsSupport ticket triage and escalationInvoice and transaction verificationRisk assessment and fraud detectionAutomated decision gates in production workflows

When to choose it (and when not to)

Reach for Jev when...

  • When you need fast, repeated structured decisions at scale
  • For decisions that should run in software background, not chat
  • When you want reliable probability calibration over raw speed
  • For cost-sensitive high-volume automation

Look elsewhere if...

  • When you need text generation or open-ended explanation
  • For tasks requiring image/audio input
  • When you need a general-purpose chat interface
  • For reasoning tasks beyond bounded classification

How to use it

  • Define bounded question types upfront (Choice, Score, Noul/yes-no)
  • Decompose complex reasoning into separate judgment questions
  • Send only necessary context to keep costs low and quality high
  • Set confidence thresholds based on your specific use-case cost tolerance

Quickstart

Python
from typesafe import TypeSafe
client = TypeSafe(api_key='...')
response = client.decisions(
  state='customer ticket...',
  questions=[
    {'type': 'Choice', 'name': 'action', 'options': ['approve', 'escalate', 'deny']},
    {'type': 'Noul', 'name': 'needs_human_review'}
  ]
)

Returns typed outputs with probability distributions and confidence scores per decision.

API model id: jev-latest

Benchmarks

BenchmarkScoreNotes
System One Decision Tasks (TypeSafe internal)Frontier-level intelligence on classification/routing tasksVendor-reported on security alert triage, support escalation, invoice validation, vendor decision tasks. Independent testing limited to accuracy sampling.

Source: TypeSafe AI Launch Materials

Compare Jev

Compare Jev with any other model

Build a comparison →
All model comparisons →

Learn the concepts