TypeSafe AI launches JEV, a decision model hitting 70-500ms at $0.042/M input tokens
sven_ai · x · 2026-09-22
TypeSafe AI released System One models and JEV on 2026-09-15: not a chat model, but a typed decision engine — unstructured state in, typed probabilistic decisions out (probabilities, confidence, structured answers) — targeting the many places in your stack where an LLM "makes a judgment" but no text generation is needed.
Official claims: 70-500ms latency, $0.042/M input tokens with free output, and up to 193.6x faster / 444.6x cheaper on some workflow evals (vendor-benchmarked, no third-party reproduction yet).
Suggested first use cases after getting a key:
- Routing: support ticket triage, agent model routing (cheap models for easy tasks), tool-call routing, document pipeline dispatch — via Choice outputs
- Moderation/guardrails: content safety checks and agent guardrails (risk judgment before bash, payments, destructive ops, emails)
Related event: Ex-OpenAI researcher launches Jev, a non-chat 'System 1' decision model(20 posts)→
More from coding & agent
- Dev's early take: Grok 4.7 excels at complex coding but falters on 3D — Daniel_Farinax · 2026-09-22
- Ex-Apple engineer open-sources jevals: fast local agent evals with one-shot decision models — byebaybay · 2026-09-22
- One prompt, 90% of tax docs gathered: Muse agent handles tax season prep — armand_ruiz · 2026-09-22
- $20 LoRA fine-tune of Qwen2.5-1.5B beats general API on sensitivity tiering, with 8x asymmetric loss — TrifleHopeful5418 · 2026-09-22
- Grok Build ships v1.0.38-40: subagents default to parent model, long-run agent workflows get steadier — XFreeze · 2026-09-22
- How do you separate agent identity from tool authorization in MCP setups? — PuzzleheadedNet2204 · 2026-09-22