What it is
Jev is the first public model from TypeSafe AI, a San Francisco lab that came out of stealth on 15 September 2026 with a $40 million seed round led by DCVC. TypeSafe calls it a "System One" model, after Daniel Kahneman's fast, intuitive mode of thought. It does not write prose, code or explanations. You send a state (a string, a JSON object or an array of text) and a set of named questions, and it answers every question in one call.
Each question is one of three types: a choice from options you define, a score on a scale you define, or a "noul", the probability that a statement is true. Some types return a confidence value, so your code can act on its own above a threshold and hand the rest to a person or a larger model. TypeSafe says it trained Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD).
Jev is API-only, text-only and closed. Two versions are live: jev-1.13.0 (the jev-latest alias, the SDK default) and jev-1.14.0 (jev-preview, from 14 October 2026). Requests may hold 64K tokens, 32K per question, or 262K with a long-context header that TypeSafe warns is slower and less accurate.
Our review
Jev now sells convenience more than quality. On the community Jev Decision Index, which now scores on private tests, Jev 1.14 ranks 41st of 229 entries at 60.24 and the stable 1.13 is 48th at 59.22. Fastino's GLiDE 1.2 leads at 71.52, and Perplexity's open-weights Decider v1.2 is fourth at 67.98 while charging $0.02 per million input tokens to Jev's $0.042.
The product idea still holds. Jev works as a cheap function call for judgements code already makes with an LLM: which team owns a ticket, whether a message is urgent, whether a tool call should go ahead. The answer can only be one of the options you defined, so nothing needs parsing. In nibzard's independent run, all 40,226 decisions came back valid for a recorded $1.71, with 79.2% accuracy on Banking77, 88.6% on CLINC150 and median latency of 267 to 316 ms under load.
Confidence is the feature that is meant to let you automate, and it needs testing on your own data. TypeSafe's docs say calibration holds across groups of predictions, not for any single answer. Nibzard fitted thresholds for at most 5% error: on Banking77 a cut-off of 0.96 automated 51.2% of test messages at 3.68% error, and on CLINC150 the chosen threshold missed the target at 6.65%. On a 77-way intent queue, expect to automate about half, not most.
The operating details are mostly sensible. TypeSafe says Jev is not trained on customer requests or responses, though zero data retention is Enterprise-only. Aliases move when versions ship, so tuned thresholds should be pinned to jev-1.13.0 or jev-1.14.0. English gets the best accuracy, input is text only, and TypeSafe caps a choice at 255 options; nibzard's pilot had 12 of 256 requests rejected for going over it.
What Jev keeps over the open field is that someone else runs it. That advantage is shrinking: vLLM's main branch has served the same /v1/systemone endpoint since 14 October, though not yet in a release, and models such as d3 load into it without conversion. Moving off Jev is now close to a change of base URL.
Where it earns its place
A support-engineering lead at a bank with a 77-category intent taxonomy
Route incoming chat messages to the right queue without an LLM bill
One choice question per message; at $0.042 per million input tokens, a million 300-token messages cost about $13. Nibzard's tests suggest a 0.96 confidence cut-off automates about half at under 4% error, with the rest going to agents.
A platform engineer building an agent framework
Gate tool calls with a fast yes/no check before they run
A noul question returns a probability for "this call is safe to run" in roughly 200 to 320 ms. Pin jev-1.13.0 so thresholds do not shift when the alias moves, and send low-confidence calls to a larger model.
A product manager at a SaaS company with no ML team
Add urgency scoring and spam filtering to a feedback inbox
Score and noul questions answered in one call per message, with no GPUs to run. If volume grows or images enter the inbox, open models such as Perplexity's Decider cost half as much and read images.
Before you commit
- Setup
- Create an API key, then POST a state and named questions to /v1/systemone or use the SDKs, which default to jev-latest.
- Integrations
- TypeSafe HTTP API and SDKs; Vercel AI Gateway; the same /v1/systemone request shape is served by vLLM for open decision models.
- Data and privacy
- TypeSafe says Jev is not trained on customer requests or responses and every account uses the same weights; zero data retention is offered to Enterprise customers. The models page gives no retention periods; the DPA sits on the Legal page.
- Cost at scale
- Input only is billed, at $0.042 per million tokens: 100M decisions of 500 tokens each cost $2,100. Nibzard's 40,226-decision benchmark cost $1.71.
- Limits
- 64K tokens per request and 32K per question, or 262K with the TypeSafe-Danger-Zone-Long-Context header; rate limits set per organisation and returned as 429s; choice questions take at most 255 options.
What's good
- Output is always one of your defined options, so nothing to parse
- $0.042 per 1M input tokens, output free
- Not trained on customer requests or responses
- Versioned model IDs keep tuned thresholds stable
What's not
- 41st of 229 on the community Decision Index, behind open models
- Text only, and best in English
- Closed weights; no self-hosting or fine-tuning
- Zero data retention only on Enterprise
Best for
Teams that want a hosted, low-cost API for text routing, triage and yes/no gates without running their own GPUs.
Alternatives
Jev: common questions
- Is Jev worth it?
- Jev turned typed decisions into a product, and open models have since passed it on both score and price. We rate it 3.8 out of 5.
- How much does Jev cost?
- $0.042 per 1M input tokens ($42 per billion); output tokens are free. Same price at long context. Enterprise limits through sales.
- What are the downsides of Jev?
- 41st of 229 on the community Decision Index, behind open models; text only, and best in English; closed weights; no self-hosting or fine-tuning; zero data retention only on Enterprise.
- Who is Jev best for?
- Teams that want a hosted, low-cost API for text routing, triage and yes/no gates without running their own GPUs.
- How much does Jev cost for a team or in production?
- Input only is billed, at $0.042 per million tokens: 100M decisions of 500 tokens each cost $2,100. Nibzard's 40,226-decision benchmark cost $1.71.
- Is Jev safe for company data?
- TypeSafe says Jev is not trained on customer requests or responses and every account uses the same weights; zero data retention is offered to Enterprise customers. The models page gives no retention periods; the DPA sits on the Legal page.
- What are the best Jev alternatives?
- The closest alternatives we have reviewed are Perplexity Decider (pplx-decider), d3 (vLLM Semantic Router), Clef, Laya and GLiNER2.5-Decide.
Answers checked Oct 23, 2026.