On 20 September 2026, TypeSafe AI released Jev, the first of what it calls System One models. It is not another large language model. It does not write paragraphs. It reads a piece of state and returns a typed decision with a calibrated probability — in 70 to 500 milliseconds, at $0.042 per million input tokens, with output tokens free. That combination reopens a question every AI team has quietly settled the expensive way: which of my LLM calls were never really about language at all?
This is a builder’s list. Seven things that are a good fit for a fast, cheap, structured decision model — each with what Jev returns, what it replaces in a typical stack, and the honest fallback if you cannot use it yet. If you would rather see the reverse — how to strip these decisions out of an existing GPT or Claude app — read the companion piece, Replacing OpenAI/Anthropic apps with Jev.
01The 30-second version
What a System One model actually is
TypeSafe frames the difference as System One versus System Two, echoing fast, instinctive thinking against slow, deliberate reasoning. An LLM is System Two: it generates language token by token and is unbeatable when you genuinely need prose. Jev is System One: it is non-autoregressive, produces all of its output in a single parallel query, and only ever returns a value from a space you defined up front. TypeSafe trained it with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD), which is why the probabilities it attaches are meant to be trustworthy rather than decorative.
Practically, that gives you three question types — a true/false judgement, a choice from a fixed set, and an ordered score — each returned with a confidence number. Because the answer can only come from your predefined options, it cannot wander off-format the way a prompted LLM can. That is the whole pitch: the boring 80% of AI calls that are really just classify, route, score or check, done two orders of magnitude faster and cheaper.
The performance figures here are TypeSafe’s own published numbers for an early-access model: 70–500ms latency against 3–329 seconds for frontier LLMs, and one workflow evaluation quoted at 193.6× faster and 444.6× cheaper. Treat them as vendor claims to verify against your own workload, not measured Appycodes results. What is not in dispute is the shape of the tool.
02The list
Seven things to build with Jev
Every item below shares one test: the output is a decision, not a document. If you can enumerate the possible answers before the call runs, it is a System One job.
A router in front of your models
Real-timeA Choice — which queue, model or tool should handle this input, plus confidence.
A cheap-LLM classifier or a brittle regex/keyword router.
Most AI products quietly pay an LLM to answer “where does this go?” before doing any real work — pick a model tier, choose a tool, sort a request. That pre-flight call adds latency and cost to every request. A System One router returns the branch in tens of milliseconds and lets you reserve the expensive model for the requests that actually need language.
Because the route comes back with a calibrated probability, you can set a threshold: high confidence auto-routes, low confidence falls through to a human or a heavier model.
Content moderation & safety triage
GuardrailA Noul (is this policy-violating?) or a Score across severity levels.
A moderation LLM prompt, or a keyword blocklist that over-blocks.
Moderation is a classic “decision, not prose” task, and it runs on the hot path where every millisecond and every fraction of a cent is multiplied by traffic. A System One check can screen each message, upload caption or comment for the policies you define, return a calibrated risk, and hand only the genuine grey-area cases to a slower reviewer.
The calibrated confidence matters more here than anywhere: it is the dial between “block aggressively” and “escalate for review”.
Lead & record scoring
Batch or liveA Score against ordered levels — fit, intent, priority, risk.
An LLM scoring prompt run per record, or manual triage in a spreadsheet.
Sales and ops teams burn real money scoring inbound leads, support tickets or applications with an LLM prompt per row. When the output is “1 to 5” or “low / medium / high”, that is a System One score. Run it live on a form submission to prioritise the queue, or in a nightly batch across the whole table for a fraction of the token bill.
Free output tokens change the economics of anything you score in bulk — see our per-token economics of AI features for why the output side of the bill is usually the one that scales badly.
LLM output verification & guardrails
VerifierA Noul — did this generation follow the rules, stay on policy, avoid a jailbreak?
A second ‘judge’ LLM call grading the first one’s output.
The standard way to make an LLM safe or reliable is to grade its output with… another LLM, doubling latency and cost on every generation. A System One verifier is purpose-built for this: screen the model’s answer for jailbreak attempts, off-policy content or missing required fields before it reaches the user, in a single fast pass.
This is exactly the “smart if-statement” pattern TypeSafe pitches — a check that used to need a full model, now cheap enough to put on every response.
RAG retrieval gating & rerank
RetrievalA Noul (does this chunk answer the question?) or a relevance Score.
A rerank/cross-encoder pass, or an LLM asked to judge relevance.
In a retrieval pipeline, the quality gate — “is this retrieved passage actually relevant?” — is a decision, not a paragraph. Putting a System One judgement between retrieval and generation lets you drop weak chunks, short-circuit when nothing is relevant, and stop paying to stuff marginal context into an expensive prompt.
We break down where this fits a real pipeline in building a production RAG pipeline — retrieval gating is the stage most teams under-build.
Support-ticket triage & deflection
OpsA Choice of intent/queue plus a Noul on ‘can this be auto-answered?’
A classifier LLM call, then a separate routing prompt.
Support automation lives or dies on the routing decision underneath it — and vendors love to quote deflection rates the reality never matches. A System One triage step classifies intent, decides whether a request is safe to auto-answer, and routes the rest, all before your knowledge-base model is ever invoked.
We measured what real bots actually deflect in vendor says 76%, reality says 41% — the honest gate is what keeps the number defensible.
Map-reduce labelling over big data
ScaleA Choice or Score per row, applied across millions of records.
A batch LLM classification job with a frightening token bill.
TypeSafe explicitly pitches Jev for “map-reducing over big data”. Classifying a product catalogue, tagging a support archive, scoring a million transactions for review — these are embarrassingly parallel decision tasks where an LLM’s per-token cost and latency make the job painful. Single-pass output and free output tokens are exactly the levers that make bulk labelling affordable.
03Get hands on
How to try it today
Jev itself is early access, but you do not have to wait to build against the interface. TypeSafe ships an MIT-licensed Python package, system-one-adapter, described as a drop-in replacement for the system_one API — backed by OpenAI or Anthropic instead of Jev. So you can write your router, verifier or scorer against the real call shape now, measure it on a model you already pay for, and switch the backend to Jev when you have access.
# The public path today: the MIT-licensed adapter, backed by an LLM.
pip install "system-one-adapter[openai]"
from system_one_adapter import SystemOneAdapterClient, Noul, Score, Choice
client = SystemOneAdapterClient(
structured_outputs=True,
llm_answer_mode="probabilities",
normalize_probabilities=True,
)
# One state, one or more typed questions. The answer comes from a space you
# defined in advance, with a calibrated probability attached.
response = client.system_one(
state="This book was a delight to read.",
questions={"positive": Noul(instructions="The book review is positive.")},
provider="openai",
model="gpt-4o-mini",
)The three question types — Noul (a true/false judgement), Choice (one option from a fixed set) and Score (an ordered rating) — are the entire vocabulary. If your task fits one of them, it fits a System One model. The full migration path, including the before/after against an existing OpenAI classification prompt, is in Replacing OpenAI/Anthropic apps with Jev.
04Stay honest
When not to reach for Jev
A decision engine is the wrong tool the moment you actually need language. Do not use Jev to draft an email, write a summary, hold a conversation, generate code, or produce anything a user will read as prose — that is System Two work, and it stays on your LLM. It is also not a retrieval system, an embedding model or a database; it decides, it does not remember.
And treat the vendor benchmarks as a starting hypothesis. Early-access latency, your own input-token sizes and the quality of your defined options all move the real numbers. The right way to adopt it is the boring way: wrap one decision, run it in shadow against your current implementation, and compare accuracy, latency and cost on your own traffic before you cut over.
We build and harden production AI systems — routers, retrieval pipelines and the guardrails around them. If you want help finding which of your model calls are decisions in disguise, and moving them onto a System One model without breaking behaviour, that is our AI systems work. Start a conversation from the contact page.
05Questions
Frequently asked questions
What is Jev by TypeSafe AI?
Jev is TypeSafe AI’s first System One model: a model built to make fast, structured decisions that software can use directly, returning typed values with calibrated probabilities rather than free-form text. TypeSafe quotes 70–500ms end-to-end latency and $0.042 per million input tokens, with output tokens free.
How is Jev different from an LLM like GPT or Claude?
An LLM generates text token by token and is best for open-ended language. Jev is non-autoregressive: it produces its answer in a single parallel pass and only returns a value from a space you defined in advance (a true/false judgement, a choice from a fixed set, or an ordered score). Use an LLM when you need prose; use Jev when you need a bounded decision fast and cheap.
What can you build with Jev?
Anything that is a decision rather than a paragraph: request and model routing, content moderation, lead and record scoring, LLM output verification and guardrails, RAG retrieval gating, support-ticket triage, and map-reduce labelling over large datasets.
Can I use Jev without early access?
Jev launched in early access in September 2026. TypeSafe also publishes an MIT-licensed Python package, system-one-adapter, that exposes the same system_one call backed by OpenAI or Anthropic. You can build against that API today and switch the backend to Jev when you have access.
Primary sources
- TypeSafe AI — Introducing System One models & Jev
- GitHub — typesafe-ai/system-one-adapter (MIT)
- TypeSafe AI — product site
Not affiliated with or endorsed by TypeSafe AI. Product names are used for identification only.
