Jev
Jev is not another chat model. It never writes text or code. You give it a situation and a few typed questions; it answers with a choice, a score, or a probability, fast and cheap. Think of it as the agent's reflex, not its brain.
1. What it is
- System One modelIntuition, not reasoning. Made by TypeSafe AI (CEO Diogo Almeida, an InstructGPT co-author).
- Typed answers onlyYou define the options in advance, so it cannot invent text. It can still be wrong.
- Probabilities, not proseEvery answer carries a probability and confidence your code can threshold.
- New, with two doorsDirect from TypeSafe: early access, waitlist. Through OpenRouter: no waitlist, just an OpenRouter key. Also on Vercel AI Gateway and OpenCode Zen.
Not a replacement for the model behind Claude Code or Codex. TypeSafe's own docs say so.
The pattern is: keep the big model for thinking, add Jev for quick decisions around it.
2. LLM vs Jev, same input
LLM (System Two)
Reads, reasons, writes. Flexible but slower and pricier, and the output is free text you must parse.
"Sounds like an integration bug. I'd send it to the technical team, it seems fairly urgent."
Jev (System One)
Picks from options you gave it and says how sure it is. Typed by construction.
department: technical (p 0.85) confidence: 0.78 is_urgent: 1.0
3. Turn the dial: when to trust it
The key habit: act on Jev's answer only above a confidence threshold, otherwise hand it to the LLM or to you. This is a sample response from the docs (a support ticket), not a live call.
4. Where it plugs into the agent loop
Each badge is a place a fast decision helps. Click one. (Loop basics: Agent.)
5. In your setup
You use Claude Code and Codex as the thinking models, run through omp and Orca. Status below is from this machine today.
| Tool | How Jev connects | What you get | Status here |
|---|---|---|---|
| omp | native Set TYPESAFE_API_KEY (or /login typesafe) and the judge role. OpenRouter credits may also work (see below). |
Typed judge() in eval (bulk triage, jevify keyword), the find tool, judged TTSR rules, auto-thinking and stop detection. |
No key set; omp models --kind judge lists none. Not configured. |
| Codex | gateway jev-codex from the community jev-gateway package. |
Jev picks the tool, Codex fills arguments (forced or even direct, no LLM call). Keeps your ChatGPT login. | Not installed. |
| Claude Code | gateway jev-claude. |
Hint mode only (thinking and prompt cache block forcing). Better picks with huge tool lists; not cheaper or faster. | Not installed. |
| Orca | untested Orca runs agents in terminals; jev-claude / jev-codex are ordinary commands. |
Same as above, per worktree. One gateway per client port serves all worktrees. | Unverified: orca errored here (“Unable to determine Orca.app path”), so I could not check for a custom agent-command setting. |
| Your own apps | API POST api.typesafe.ai/v1/systemone, SDK typesafe-sdk, or the TypeSafe agent skill. |
Idea, not built: auto-categorise expenses in the family cost app, asking you only when confidence is low. | Idea only. |
omp: the zero-risk start
# ~/.omp/agent/config.yml
modelRoles:
judge: typesafe/jev-latest
export TYPESAFE_API_KEY=... # or /login typesafe inside omp
omp models --kind judge # should now list jev-latestCodex / Claude Code through the gateway
npm install -g jev-gateway # Node 22.15+
jev-codex # use exactly like `codex`
jev-claude # use exactly like `claude`
jev-codex --dashboard # watch decisions, tokens, latency
jev-codex --routing off|on # A/B on your own tasksPlain
codex and claude stay untouched. Only sessions started with the jev-
commands go through the gateway, and if Jev fails the request just goes to the LLM.Using your OpenRouter credits
Yes. OpenRouter serves Jev to any account with an API key: no waitlist and no TypeSafe account,
billed to your credits. Output tokens are free; you pay per input token (price is on the
model page, and each response reports usage.cost).
| Where | What to set |
|---|---|
| Model id | typesafe/jev-1.13, or the alias ~typesafe/jev-latest |
| Your own code | POST https://openrouter.ai/api/alpha/decisions (Decisions API), or /api/v1/systemone with the TypeSafe SDK pointed at OpenRouter |
jev-gateway | JEV_PROVIDER=openrouter and OPENROUTER_API_KEY (or answer option 2 in its first-run prompt) |
| omp | Its built-in judge chain already lists openrouter/~typesafe/jev-latest after the direct TypeSafe model. Whether it works with your login is untested here: no OPENROUTER_API_KEY is set in my shell, and omp models --kind judge lists nothing. |
Caveat: the gateway's README says its OpenRouter path is covered by tests but not yet run with a
real key. Its first-run key check will say at once if something disagrees. OpenRouter also publishes a
“Jev Router” (Jev picks the model and reasoning effort per request) and a cookbook for auto-answering coding-agent
permission prompts with a Jev safety check. Neither is tried here.
6. Does routing pay off?
Third-party benchmark from the gateway's README: the same coding agent does a task with routing on vs off. Negative = better. Medians, only 5 runs per cell.
Reading it: debugging gets cheaper for every model; feature work is mixed, and Opus 5 and Sonnet 5 got worse. GPT-5.6 Luna also solved 3 of 5 tasks with routing against 5 of 5 without. So measure on your own work first.
7. Suggested rollout
- omp judge role. Native, no behaviour change to your main model, and it unlocks bulk triage and judged rules. Verify with
omp models --kind judge. - Codex through
jev-codex, with a baseline. Do a few similar tasks with--routing off, thenon. Keep it only if the dashboard shows a win on your tasks. - Claude Code last, or skip. Hint mode brings accuracy, not savings. Worth it only if a 280-tool list is confusing it.
8. Caveats
- Speed and cost claims (“193x faster, 445x cheaper”) are the vendor's headline. Real examples reported: Vercel's safety classifier 5 to 18 times faster; one team found Gemini slightly more accurate but 10 to 20 times pricier.
- Jev reads text only, has a 32k-token window, and is most accurate in English. Architecture is undisclosed.
- The gateway adds a Jev call (about 0.5 to 1 s) to every turn that carries tools. A wrong forced tool can derail a turn; raise the confidence threshold if it does.
jev-gatewayis independent of TypeSafe. Its launcher gateway has no key, so any local process can use it (localhost only).- This is days-old news. Expect details to change.
Sources
- TechCrunch, A new kind of AI model from a ChatGPT inventor is thrilling developers (18 Sep 2026)
- TypeSafe docs, Jev with coding agents and Quick start (sample response in section 3)
- OpenRouter, Jev documentation (access, model ids, endpoints, pricing model)
- vinilana/jev-gateway README (commands, modes, benchmark table, trade-offs)
- omp docs (this machine):
settings.md,environment-variables.md,tools/find.md,magic-keywords.md,ttsr-injection-lifecycle.md - Tom's Hardware (speed and cost claims, not independently checked), DataCamp (waitlist, API route)