GILD verdict
Choose Jev when a repeated software decision has a fixed set of answers and needs typed probabilities.
Teams that adopted it report speed and cost first: Isotopes AI runs Jev 1.13 as its production default after it passed 100% across three runs of 88 scenarios, Serverpod roughly halved time to first token, and MotherDuck labeled 100,000 rows in 40 seconds for fifty cents (Sep and Oct 2026). The tests that went against it show where it loses: kapa.ai chose Voyage, which tied it on reranking and cost less; Cisco and Anaconda beat it with trained classifiers; Brief found none of 13 code-review problems with it; dessn moved more than 10,000 calls to one Claude Opus 5.5 call for accuracy. TypeSafe prices input at US$42 per billion tokens with output free. You take on evaluation, thresholds, fallbacks and version pinning; Isotopes held back Jev 1.14 after it scored 93.6% on the same suite.
Jev
AI Model API by TypeSafe AI
+40 PositiveFit and limits
Who it fits
- The current workflow parses generated prose to choose a branch, as Isotopes AI's did before Jev passed 100% across three runs of 88 scenarios.
- Many small decisions share one state payload, such as Serverpod's document choice or Browser Use's next browser action.
- Bulk labeling where cost per row matters, as in MotherDuck's 100,000-row SQL benchmark.
- A team can keep thresholds and fallback behavior in application code.
Where it falls short
- The task needs open-ended writing or judgment in the same call: Brief's code-review test found none of 13 known problems.
- A trained classifier already meets the target, as at Cisco AI Defense and Anaconda.
- A dedicated service already wins on price and speed, as Voyage did on kapa.ai's 1,000-case reranking benchmark.
- Inputs can carry hostile text: TypeSafe's own jaggedness page lists adversarial content as a weakness.
- The call sets severity with no person reviewing it: Casco saw Jev give 550 critical scores to a set holding 55 published critical findings.
- On a Berkeley Function Calling Leaderboard slice, Jev's top label called a function on about 1 in 13 items where none fit, against about 1 in 25 for Gemini Flash-Lite.
Implementation reality
Write the decision contract
Name the state, question type, options or score levels, and the action each answer permits. Keep business rules in code.
Measure beside the current path
Replay labeled examples and compare Jev with the existing classifier or LLM. Track accuracy, latency, input tokens, abstentions, and review rate.
Gate action with confidence
Use confidence as a second axis beside probability. Route uncertain calls to a human or a safer deterministic path. In MLH's test, raising Jev's bar to 0.8 probability cut false positives to 1% and doubled deferrals.
Pin the model and review changes
Pin a Jev model ID, keep regression cases, and re-run them before switching versions or changing thresholds.
What to measure in your pilot
- Replay 500 labeled cases through Jev and your current path; adopt Jev only if its dangerous-error rate matches or beats the current path at your chosen threshold.
- Log the share of calls sent to review at your confidence threshold for two weeks; above 20%, route those cases to a second-stage model and keep the bar.
- Price 1,000 real decisions three ways, Jev alone, Jev with a small writer model, and your current model, and keep the cheapest path that holds your error target.
Sources
- Official documentation
- 9
- Public source
- 1
- TypeSafe: Introducing System One Models & Jev
- TypeSafe docs: Introduction
- TypeSafe docs: Models
- TypeSafe docs: System One API
- TypeSafe docs: Confidence
- TypeSafe docs: Jev jaggedness
- TypeSafe docs: Intent routing
- TypeSafe: Jev 1.14
- Mike Swift, Major League Hacking, on DEV: I got Jev to zero mistakes. I'm still using Flash-Lite. (Oct 8, 2026)
- TypeSafe: Enterprise