TypeSafe AI opened early access to Jev, its first System One model. The product bet is blunt: stop asking chat models to invent a string you then parse, and ship a model whose contract is typed decisions with calibrated probabilities. Jev does not generate prose, code, or free-form JSON. You send shared state plus one or more typed questions; you get Choice / Score / Noul-style answers your code can branch on.
Founder Diogo Almeida (ex-OpenAI instruction-following / RLHF work behind InstructGPT and ChatGPT) frames the gap TypeSafe is chasing: models that are great at chat still leave a hole for automation inside software. Co-founders include Erik Gafni and Sasha Sheng. The company came out of stealth with about $40M seed funding led by DCVC, per TypeSafe’s Business Wire release — treat that as company/press-release confirmed, not a launch-blog footnote.
Access paths that matter for builders now: TypeSafe’s own API (early access / waitlist pacing), OpenRouter (typesafe/jev-1.13 and Decisions API), Vercel AI Gateway (typesafe-ai/jev via AI SDK experimental_evaluate), and a LangChain integration (langchain-typesafe / TypeSafeClassifier).
Image credit: TypeSafe
Key points
- What shipped: Jev — TypeSafe’s first System One model: non-generative decision API (Choice / Score / Noul), calibrated probabilities + confidence, parallel multi-question evaluation over shared state.
- Not a chat replacement: Complements LLMs for routing, triage, guardrails, scoring, and verification; LLMs still own generation and open-ended reasoning.
- Training / stack (vendor): New architecture, parallel sampler, RLCD (Reinforcement Learning for Calibrated Decisions). No public weights; proprietary.
- Latency / cost (vendor, self-reported): End-to-end roughly 70–500 ms; company contrasts ~40×–200× faster vs frontier LLMs on System One–shaped queries. List price often cited at $0.042 / MTok input, output free (OpenRouter / TypeSafe / Vercel Gateway pages align on that list). Homepage peaks like ~193.6× faster / ~444.6× cheaper are TypeSafe’s own workflow-eval highs — they say those sit on the higher end.
- Distribution: TypeSafe API, OpenRouter, Vercel AI Gateway, LangChain middleware examples (model routing, AutoMode-style tool guardrails).
- Company: ~$40M seed led by DCVC (Business Wire / FinSMEs / SiliconANGLE); founders Diogo Almeida, Erik Gafni, Sasha Sheng.
- Caveats: Speed/cost multiples and “can’t hallucinate” are vendor claims. “Hallucination-free” here means no off-schema string generation, not perfect judgment. Workflow evals use expensive LLM consensus as reference, not human ground truth. 9to6AI has not independently re-benched.
What shipped
TypeSafe calls System One models a class built to make fast, structured decisions software can consume directly — brand language borrowed from Kahneman’s System 1 (fast / pattern-matching), while arguing constrained outputs can be more reliable for automation than free-form chat.
Jev is positioned as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out. Three question primitives (names vary slightly by platform):
| Primitive | Role | Typical return |
|---|---|---|
| Choice | Pick from a fixed option set | Selected option + per-option probabilities + confidence |
| Score | Rate on ordered levels you describe | Continuous score, level distribution, confidence |
| Noul (Boolean on some gateways) | Yes / no proposition | Probability the statement holds |
Multiple questions share one state read and are evaluated in parallel in a single request — TypeSafe’s efficiency argument versus autoregressive token generation.
Named company pieces: RLCD training, parallel sampler, new architecture (not fully disclosed). Demos highlighted in the launch essay include a Doom bot on structured game state and Wikiracing over high-cardinality link choices (cardinality up to 255; higher uses a two-stage score-then-choose path). Treat demos as latency / typed-choice illustrations, not production SLAs.
OpenRouter’s developer explainer documents live Decision API shapes against typesafe/jev-1.13, including how Score’s continuous value is a probability-weighted average of levels, and how confidence should be thresholded in your code rather than trusted as a single oracle.
What changed
For years, classify / route / score / judge loops usually meant: prompt an LLM → parse JSON or tool calls → validate → retry on schema failure → accept residual hallucination and variable latency.
Jev’s launch is a category claim: those jobs should not go through string generation. That changes the failure mode:
- Schema / type errors: TypeSafe says these are impossible by construction (outputs constrained to the declared schema). Easy to falsify with one counter-example; builders should still verify on their keys.
- Wrong answers: Still possible. Calibration and confidence are the company’s answer for auto-act vs escalate vs abstain.
- Eval method: TypeSafe’s workflow evals fix a compute graph and score models against the average of large reference models (launch materials cite Astra / Fable-class references). That measures agreement with expensive LLM consensus under a shared harness — not independent ground truth. TypeSafe discloses capabilities-team authorship bias and that reference choice can underestimate DeepSeek-relative performance.
Ecosystem pull (attribute carefully):
- OpenRouter published a full “what is Jev” guide with real Decision API responses and list pricing.
- LangChain shipped
TypeSafeClassifierplus experimental middleware for model routing and AutoMode-style tool guardrails, with builder anecdotes (Browserbase, live trading, email triage) — these are LangChain’s named examples, not 9to6AI audits. - Vercel added Jev to AI Gateway and documented AI SDK
experimental_evaluate; the changelog repeats TypeSafe’s speed/cost multiples as TypeSafe-reported. Treat any “fastest-adopted on Gateway” marketing as platform self-report, not an independent ranking.
What it means for builders
1. Put Jev (or similar) where you currently burn an LLM for a label.
Support routing, urgency triage, incident severity, “should a human review this agent trace,” jailbreak / policy scoring of LLM outputs — these are the shapes TypeSafe, LangChain, and Vercel all market. Keep an LLM for replies, summaries, and open-ended tool planning.
2. Design for confidence thresholds, not just the argmax.
Wire auto-act / escalate / abstain bands. Calibrate on your labeled tickets before cutting over. OpenRouter’s guide is blunt: a single 0.8 can still be wrong; calibration is about averages.
3. Model cost with published list prices, then re-check.
At $0.042 / MTok input and $0 output on TypeSafe / OpenRouter / Gateway list pages, high-QPS decision layers become budgetable differently than frontier chat. TypeSafe notes it cannot prove pricing is not subsidized and expects prices to fall. Third-party hosted wrappers may meter differently — always read the meter you are on.
4. Prefer parallel multi-question calls over serial LLM classify hops.
One state, many questions, one round trip is the product’s economic and latency thesis. Cardinality and state packing still matter (Wikiracing two-stage path is a hint).
5. Early access / version pinning.
No public weights. Pin versions (e.g. jev-1.13 / jev-1.13.0) once you set thresholds — OpenRouter and docs warn snapshot strings can change. Waitlist vs Gateway/OpenRouter availability may differ by account.
6. Keep generation and decision layers separate in the agent loop.
LangChain’s framing matches the honest architecture: LLM decides what language to produce; Jev decides whether / where / how risky — complementary, not competitive for chat share.
What to watch
- Independent latency, cost, and accuracy benches vs GPT/Claude/Gemini structured-output baselines on fixed public workflows.
- Whether confidence tracks error rates on customer data, not only vendor plots.
- Pricing durability of $0 output + low input rates.
- Waitlist → broader GA, region/latency outside West Coast (TypeSafe notes many published speed evals run West Coast → West Coast).
- How much of architecture + RLCD becomes paper vs closed API.
- Competitive response: cheaper “decision-only” SKUs from major labs aimed at classify-route-judge spend.
Sources
- Introducing System One Models & Jev — TypeSafe AI (primary)
- TypeSafe docs · Workflow evals
- What Is Jev? — OpenRouter
- Building a Harness with Jev — LangChain
- TypeSafe AI’s Jev now available on AI Gateway — Vercel · Classify, route, and score with Jev — Vercel KB
- TypeSafe AI Emerges From Stealth With $40M — Business Wire (funding; DCVC led)
