Models

September 19, 2026 · View on GitHub

All System One models share POST /v1/systemone. The model field selects which one.

Current (fetched 2026-09-19)

Versioned IDjev-1.13.0
Aliasesjev-latestjev-1.13.0; jev-preview → same (no preview build right now)
Price$42 / Btok input = $0.042 / Mtok. Output tokens free.
Rate limits250k tokens/s and 1,200 rpm (dynamic; 429 over either). Higher on enterprise.
Context64k tokens/request; 32k for state + longest question
InputText only. String, object, or array of text.

Pin jev-1.13.0 if you have tuned thresholds against that version; aliases move on release. Response model reports the ID that answered.

Jev is not fine-tuned per account. Shape behavior with state, instructions, and criteria. Not trained on customer requests. English is the primary training language.

evaluate default: jev-latest (TypeSafe) or ~typesafe/jev-latest (OpenRouter).

Official: models

Jaggedness (jev-1.13, reviewed 2026-09-17)

FailureDo this
Literal readingWrite the exact condition; put boundaries in criteria
Math, counting, numeric encodings (hex, RGB)Compute in code; ask one noul per item and sum
Date/time order, duration, windowsExtract parts as choice (include "not stated"); compare in code
Indirection / double negativesDirect instructions; name the relevant state paths
Large irrelevant stateFilter first; optional noul for relevance
Adversarial / injected instructionsPrecise criteria; test edges
Contradictory instructions vs criteriaAlign them
Expected identities (P + P(not) = 1, noul vs yes/no choice)Ask each decision one way; enforce identities in code
Generation / free-form extractionCandidate in code or a generative model; Jev selects

Score interpolation between levels is weakly calibrated — threshold, don't reconstruct a magnitude.

Avoid: asking what code can compute; hiding several judgments in one question; System Two multi-hop reasoning; stuffing unused context.

Official: jaggedness

Training (why not an LLM)

Jev is trained with RLCD (reinforcement learning for calibrated decisions): typed answers and probabilities, not preferred prose (RLHF) and not long chain-of-thought (RLVR). Calibration is a group property (P=0.8 should be right ~80% of the time), not a guarantee on one answer.

TypeSafe's claim on System One tasks vs LLMs: orders of magnitude faster/cheaper; no string generation. Treat marketing multiples as theirs, not ours.

Official: AI primer · manifesto