Jev new

What it is

Jev is a transformer that outputs decisions, not tokens. TypeSafe AI released it on September 15, 2026 and calls the category System One, after Kahneman’s fast, intuitive mode of thinking. The company was founded by Diogo Almeida, who worked on RLHF and InstructGPT at OpenAI, and has raised $40M.

An LLM asked to classify something writes a string that you then parse and hope is valid. Jev skips the string. You declare the possible answers up front and get back a probability distribution over them. It is non-autoregressive: every answer comes out of a single forward pass, so there is no chain of thought and no generated text at all.

It is not a replacement for an LLM. It is meant for the small decisions inside an agent loop that currently cost a full LLM call.

How it works

One endpoint, POST https://api.typesafe.ai/v1/systemone, with three fields:

• state: a string, JSON object or array. Text only, no images or audio.
• questions: a map of named, typed questions about the state.
• model: jev-latest or a pinned id such as jev-1.13.0.

Three question types:

Type	Asks	Returns
Choice	pick one of up to 255 described options	choice, probabilities, confidence
Score	place the state on an ordered rubric of levels	score (probability-weighted, can fall between levels), probabilities, confidence
Noul	is this statement true	noul, P(yes) from 0 to 1

from typesafe_sdk import Choice, Noul, TypeSafeClient

client = TypeSafeClient(model="jev-1.13.0")
res = client.system_one(
    state={"tool": "bash", "command": "rm -rf ./build", "cwd": "/repo"},
    questions={
        "risk": Choice(
            instructions="How risky is running this command?",
            criteria={
                "safe": "Read-only or confined to build artifacts",
                "review": "Modifies tracked files or installs packages",
                "block": "Deletes outside the repo, touches secrets, or exfiltrates data",
            },
        ),
        "destructive": Noul(instructions="The command deletes files."),
    },
)
res.answers["risk"].choice         # "safe"
res.answers["risk"].confidence     # 0.91
res.answers["destructive"].noul    # 0.97

Until the waitlist opens, the practical way in is the Vercel AI Gateway, which serves it as typesafe-ai/jev through the AI SDK’s experimental_evaluate. The shape differs there: the yes/no type is boolean, its answer is probability, and confidence sits in providerMetadata.typesafe. The build uses that route.

The state is processed once and all questions are evaluated in parallel, so adding questions barely changes latency.

confidence is a single 0 to 1 statistic computed from the shape of the probability distribution: concentrated means sure, spread out means unsure. TypeSafe’s training method, Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for those probabilities matching real hit rates. The training data is entirely synthetic.

Numbers, all from TypeSafe:

• Latency 70 to 500 ms end to end.
• $0.042 per million input tokens. Output is free.
• 64k tokens per request; 32k for the state plus the longest question.
• 1,200 requests per minute, 250,000 tokens per second.
• SDKs: typesafe-sdk (Python), @typesafe-ai/sdk (JS), langchain-typesafe.

When to use it

• Tool-call gating. Score each proposed tool call for risk before it runs, and send low-confidence ones to a human. LangChain ships this as AutoModeMiddleware.
• Model routing. Decide which LLM a request deserves before paying for one.
• Agent monitoring. Ask yes/no questions about each step of a trace: is it looping, did it drift off task.
• Bulk classification. Tickets, emails, moderation, anything with a fixed label set and high volume.

Gate different actions at different thresholds. A read-only action can run at a confidence that a payment should not.

Do not use it for anything open-ended: summaries, code, extraction of free text, multi-step reasoning.

Watch out for

• The benchmarks are the vendor’s. The 40 to 200x speed and 400x cost figures against GPT-5.6 Terra, GPT-6 Astra and Fable 5.1 have not been independently reproduced. Early third-party reports are smaller: a Vercel engineer measured 5 to 18x faster, Bryo AI 10 to 20x cheaper. Calls for the build on this site took 450 to 800 ms through the Vercel gateway.
• “Cannot hallucinate” means it cannot return an option you did not define. It can still pick the wrong one. Measure that with your own eval set.
• Calibration is a claim until you check it. Bucket answers by confidence on labeled data and compare with accuracy before trusting a threshold.
• It reads literally. Negations, implied conditions and indirection are taken at face value. Put boundary cases in the criteria, or split into several direct questions and combine them in code.
• No arithmetic. Counting, date comparison and numeric reasoning are unreliable. Do them in code. Do not interpolate a Score back into an exact number.
• Irrelevant state lowers accuracy. Filter to the fields the question needs.
• State is not treated as adversarial. If a user controls it, a prompt injection can steer the decision. Do not make Jev the only guard in front of a dangerous tool.
• jev-latest moves. Thresholds tuned on one version may not hold on the next. Pin the version and log the model field from each response.
• English first. Other languages work with lower accuracy.
• Early access. Waitlist only, and rate limits adjust with demand.