TypeSafe AI · Independent

Jev

Jev is TypeSafe's "System One" model, built for fast, structured decisions inside software and agents. It doesn't chat or write code. You send it a state (text or JSON) plus typed questions, and it returns typed answers with probabilities and confidence that your code can act on directly.

✓ Last verified October 15, 2026 from official sources

Pricing

$0.042 per million input tokens ($42 per billion). Output tokens are free. Jev 1.14 and 1.13 are priced the same. Rate limits are set per organization; higher limits on custom and enterprise plans.

Best for

Routing requests to the right handler, whether code, a specialist LLM or a human
Guardrails that screen messages going into and out of an LLM app
Classification, scoring and confidence-gated "smart if" statements in agent workflows

Privacy · Skill

Cloud API. TypeSafe says Jev isn't trained on customer data; enterprise customers can get zero data retention. Skill level: Advanced.

What it's great at

Cheap, calibrated decisions on every agent turn
Confidence-gated routing: act, escalate to a bigger LLM, or ask a human
Guardrails, re-ranking and classification with typed outputs

What it can do

Three question types: Choice (pick one option, with per-option probabilities and confidence), Score (rate against a rubric you define) and Noul (the probability that a yes/no statement is true)
Many questions per request, all evaluated in parallel against one state ("speculative fan-out")
Jev 1.14.0 (released Oct 14, 2026, as jev-preview) brings small improvements across the board, mostly in multilingual support. TypeSafe says jev-latest will move to 1.14 a week after release.
Experimental "danger zone" features behind API headers: image inputs, a new Box type (a bounding box) and long context up to 262k tokens
Context of 64k tokens per request and 32k per question by default. Same weights for every account; TypeSafe says it doesn't train on customer requests.

Example tasks to try

A Choice question routes each support ticket to billing, technical or sales, and anything below your confidence threshold goes to a human.
A Noul question checks whether a quote's source actually supports the claim before your agent cites it.
Score each retrieved passage, and decide in code which ones reach your answering model.

Start in 5 minutes

Get an API key from the TypeSafe console (console.typesafe.ai)
Try it in the browser Playground: paste some text as the state and add a Choice or Noul question
Or install the TypeSafe agent skill so Claude Code or another coding agent can write the integration (npx skills add typesafe-ai/skills --skill typesafe-ai)
Pin the model ID (jev-1.14.0 or jev-1.13.0) if you tune confidence thresholds, because the jev-latest and jev-preview aliases move

Go to TypeSafe AI's official page →

Watch out for

It can't replace the LLM behind Claude Code, Cursor or Grok Bot. Use it alongside one. English is its strongest language. TypeSafe flags the danger-zone features (images, Box, long context) as not production-ready.

Similar agents

Anthropic

Automation you can self-host, with AI agent steps.

Official sources

Prices change often. Confirm on the vendor's page before you buy. We're independent, and no vendor pays for placement.

Jev