Jev (TypeSafe AI)

Overview

Jev is TypeSafe AI’s first System One model — a new class of AI model purpose-built for decisions inside software, not text generation. You send it a state (any text) plus a set of typed questions, and it returns structured answers — a boolean probability, a choice from a list, or a rubric score — each with a calibrated confidence value. There’s no parsing, no prompt scaffolding, and no hallucinated prose: the model’s output is constrained to exactly the schema you declare. It launched in public early access on September 15, 2026.

Read our full Jev AI review.

Key Features

• Three primitives: Choice (pick an option), Score (rate on a rubric), Noul (0–1 truth probability) — mix them freely in one call
• Parallel evaluation: every question runs against the same state in a single pass; adding questions barely changes latency
• Calibrated confidence: every answer reports how sure the model is, so your software can act at high confidence and escalate at low confidence
• No hallucinations structurally: outputs can’t be malformed, undeclared, or wrong-typed — the schema is enforced at the model level
• Fast and cheap: 70–500ms latency, about 193x faster and 440x cheaper than frontier LLMs on System One-style tasks

Use Cases

• Support triage: classify tickets by department, urgency, and refund-triggering sentiment in one call
• Agent orchestration: decide which tool or subagent to invoke next, or whether to continue, retry, or stop
• Guardrails and moderation: verify model outputs, detect policy violations, and block risky actions
• Scoring and ranking: rate fit, risk, or quality with a custom rubric
• Routing: choose the right pipeline, provider, or queue per request
• Email and message triage: flag urgency, intent, spam-likelihood, and sentiment

Pricing

• Input tokens: $42 per billion ($0.042/M) — charged per input token
• Output tokens: free (the model returns tiny structured answers)
• Rate limits: 250,000 tokens/second, 1,200 requests/minute
• Context: 64k tokens per request (32k for state plus the longest question)
• Access starts on the early-access waitlist at typesafe.ai; API keys are issued in console.typesafe.ai