What it is
Jev is a model from TypeSafe AI that answers typed questions instead of writing text. You send a state (an email, a CRM record, a contract clause, as text or JSON) and one or more named questions, and it returns structured answers: a choice from up to 255 options you define, a score on a rubric of up to 10 levels, or a noul, a 0-to-1 probability that a statement is true. Choice and score answers also return the full probability spread and a confidence value. There is no prose to parse and no malformed JSON to retry.
TypeSafe came out of stealth in September 2026 and, on 9 October 2026, announced an $870M round at a $7.5B valuation led by Andreessen Horowitz with Sequoia and DCVC. The product is still labelled early access on typesafe.ai. Access is through one REST endpoint (POST https://api.typesafe.ai/v1/systemone) plus Python and JavaScript SDKs. The docs list no native Zapier, Make or n8n connector, so in those tools you call it from an HTTP request step.
Where it fits in ops automation
Jev is for the decision inside an automation, not the automation itself. The scoped use cases where it earns its place:
- Ticket triage. One call asks “which queue?” (choice), “how urgent?” (score) and “is this a refund request?” (noul) against the same ticket. TypeSafe evaluates the questions in parallel, so adding a question barely changes latency.
- Lead routing pre-checks. Classify an inbound form or reply into segment, intent and “existing customer?” before your router assigns an owner.
- Approval gates. Expense pre-approval, discount-desk triage, or “does this clause deviate from our playbook?” as a yes/no with a probability attached.
- Tool-call approval in agents. Ask Jev whether an agent’s proposed action matches the user’s request before it runs.
The pattern TypeSafe documents is confidence-gated routing: act automatically above a high threshold, ask for confirmation in the middle band, and send anything below roughly 0.5 to a human.
Pricing reality
Jev 1.13 and Jev 1.14 both cost $0.042 per million input tokens, and output tokens are free (TypeSafe’s models page, checked 21 October 2026). TypeSafe offers no seat licence. Enterprise plans add higher rate limits and zero data retention through sales.
At ops volumes the bill is a rounding error. 100,000 tickets a month at about 1,500 input tokens each is 150M tokens, or about $6.30. The same volume on OpenAI’s Decisions API ($0.10 per million input, no output charge) is about $15. On a general model such as GPT-6 Luna through the Responses API, it is about $17.50 once you pay for roughly 50 output tokens per answer. Price will not decide this purchase. Accuracy on your own labels, and whether the confidence numbers mean something on your data, will.
Best for
A RevOps, support-ops or legal-ops engineer who already runs triage or routing through n8n, a custom service or an agent framework, and is spending money or latency on an LLM to produce one label per record. The win is speed, near-zero cost, and a probability you can set thresholds on.
Not for a team without someone who can call an API and keep a labelled test set, or a workflow where the decision needs multi-step reasoning, arithmetic or date logic.
Versus the alternatives
OpenAI Decisions API is the direct competitor and the default if you already run on OpenAI. It entered public beta on 6 October 2026 on gpt-6-luna with the same three question shapes (predicate, choice, score), at $0.10 per million input tokens. Pick it when you need ZDR, HIPAA eligibility or EU data residency today without a sales call, or one vendor contract. Pick Jev when input cost or latency at very high volume is the constraint, and when it beats Luna on your own eval. At least one commenter in the HN funding thread reported that it did in production evals.
A general LLM with structured outputs (Claude, ChatGPT models) is what most teams run now. Keep it when the decision needs reasoning across steps, or when you want the explanation text alongside the label for an audit trail.
Native triage in your helpdesk or router (Zendesk, Intercom, LeanData) is the right call when the vendor’s built-in AI or rules already cover your categories. Adding a model call adds a dependency for no gain.
Watch-outs
- “Confidence” is not calibrated accuracy. TypeSafe’s own docs define it as a statistic of how peaked the probability spread is, not as a measured hit rate. HN commenters disputed the calibration claims, and some found general LLMs more accurate on business-domain labels. Guard: run 200 to 500 historical records with known outcomes through Jev, plot accuracy per confidence band, and set each threshold from that plot, per question.
- Documented blind spots. TypeSafe’s Jev 1.13 notes list first-option bias in choice questions, weak handling of numbers, dates and double negatives, and degraded accuracy when the state carries irrelevant detail. Guard: compute dates, amounts and SLAs in code before the call, trim the state to the fields that matter, and shuffle option order or use the documented self-consistency pattern on high-stakes choices.
- The state is not treated as hostile. The same notes say injected instructions in the input can change answers, despite the “zero hallucinations” marketing line. Guard: never let a Jev answer alone trigger a refund, deletion or contract action on customer-supplied text; require a second check or a human above a cost threshold.
- Early-access vendor, moving aliases.
jev-latestandjev-previewrepoint when releases ship, and the docs state no deprecation policy. Guard: pinjev-1.13.0(or whichever version you validated), re-run your eval set before changing versions, and keep the model call behind one function so you can switch to OpenAI’s endpoint in an afternoon.