TypeSafe introduced Jev, a model built specifically to make decisions rather than to generate free-form text. Jev accepts data representing a state (for example, a support ticket, an agent trace, or a JSON blob) plus a list of typed questions (pick one of these options; score this on a scale; is this statement true?). Instead of emitting a token stream, it returns typed answers with probability distributions in a single parallel pass.
Performance and pricing
According to TypeSafe, Jev responds in 70 ms to 500 ms and is priced at $0.042 per million input tokens. TypeSafe’s own evaluations claim Jev can be 40×–200× faster and 40×–400× cheaper than LLM-based solutions for similar tasks. In one small independent comparison, a general-purpose model used about 910 output tokens on average to reach a yes-or-no answer, while Jev used 85 tokens — and those weren’t billed.
Training approach and guarantees
TypeSafe trained Jev using a method they call Reinforcement Learning for Calibrated Decisions (RLCD), intended to align stated probabilities with actual correctness: a higher reported probability should correspond to a higher chance the answer is right. TypeSafe also asserts Jev "can’t hallucinate," but that is limited by the schema: Jev cannot return answers outside the provided schema, though it can still be wrong within that schema. The model's probability scores are intended to signal confidence.
Early results: mid-tier accuracy at large cost savings
In TypeSafe’s published evals across four decision workflows (one being review of finished agent traces to decide if a human needs to look), Jev averaged 68% accuracy at $0.0004 per decision and 0.4 seconds per case. By comparison, GPT-5.6 Terra also reached 68% but at $0.03 and 10 seconds; Opus 5 scored 73% at $0.18 and 38 seconds. Jev sits a few points below Opus 5 in accuracy but claims a roughly 440× cost advantage in that comparison.
Early independent tests show similar directions. One team ran 777 judgments in under 0.7 seconds for about a quarter of a cent. UK events site NearHere tested listing moderation and reported 96% accuracy from Jev versus 86% from Gemini Flash-Lite, with Jev 58× cheaper per decision. A developer ran Jev zero-shot on 18,514 spam emails and observed a statistical tie with a trained classifier.
In the spam evaluation called out in the coverage, Jev achieved 98.3% accuracy with no labeled examples; a TF–IDF logistic regression trained on about 14,800 labeled emails hit 98.4%. The two disagreed on 466 emails and split them roughly evenly, with no statistically meaningful difference between the systems.
What you gain and what you lose
The most obvious trade-off is explanation: TypeSafe’s documentation states that System One models do not produce explanations of their reasoning, and NearHere noted the same in its test — a category and probabilities returned, nothing more.
For many use cases that matters. LLM judge explanations provide not only the label but the why, which is a strong signal for automated retraining, debugging, or coding agents. Jev provides only a probability, which is less instructive when trying to determine how to improve a system.
That said, Jev’s low cost enables hybrid approaches: run Jev on every trace for broad, cost-effective measurement and monitoring, and sample failures for re-evaluation with an LLM judge to obtain directional signals. Doing so may require rearchitecting systems.
Calibration matters for automation and escalation
A model that’s correct 95% of the time but can’t flag which 5% need human review can’t fully automate workflows. LLM judges have been evaluated by accuracy against gold sets, which shows expected error rates but not where the errors will occur. Production systems need a decision threshold, an escalation path to humans, and drift monitoring — all of which rely on trustworthy probability scores. LLM judges often produce overconfident probabilities.
A 2025 JudgeBench study of 14 models found judges clustering predictions at 90%–100% confidence while achieving lower accuracy, motivating a shift from accuracy-centric to confidence-driven evaluation. Jev’s spam test provides an example of usable calibration: of the emails Jev scored under 0.1, 0.1% were spam; of those scored 0.9 or above, 99.9% were spam. In the 0.5–0.6 band, only 38% were spam. Sending the 4.6% of emails scored between 0.3 and 0.7 for human review left the rest with 99.5% accuracy.
Another practical metric is tokens per decision: a component that spends hundreds or thousands of output tokens to emit a label is likely the wrong tool for many decision tasks. UkisAI’s Swift-Qwen3.8-27B, for example, cut 58% of its reasoning tokens on GPQA-Diamond and lost only 0.1 points of accuracy — suggesting many tokens don’t materially improve outcomes.
Implications for application design
At $0.0004 and 0.4 seconds per decision, sampling becomes optional: you can check every output, tool call, and agent step in real time. To benefit, designers should decompose work into many small typed questions and only invoke expensive LLM generators when text must be produced. That yields a stack where the decision layer can be versioned, measured, and swapped independently of the generation model — the first time decision-making has been cheap, fast, and calibrated enough to make that practical without labeled datasets.
A practical next step for teams is to inventory how many LLM calls in their systems end with a small set of labels and evaluate which of those calls could be replaced with a decision-first model that returns calibrated probabilities.
Conclusion
TypeSafe’s Jev introduces a new class of decision-first models that deliver calibrated probabilities in zero-shot settings at dramatically lower latency and cost than generative LLM-based judges. The primary limitation is the absence of free-form explanations, which affects automated improvement workflows. Early benchmarks are promising but limited; more independent evaluation and longer-term operational experience will be needed to understand how broadly this approach should change AI architecture and monitoring practices.
