Jev TypeSafe: Decision Model for EU Teams 2026

TL;DR

Jev is TypeSafe AI’s System One decision model: you send application state plus typed Choice, Score or Noul questions and get calibrated probabilities back in 70 to 500 ms, at $0.042 per million input tokens with free output. Use it in front of Claude for high-volume routing. Hosting is US-only today, so EU personal data needs a documented SCC transfer and confidence thresholds.

What is Jev and how does it differ from an LLM?

Jev is a frontier model built for decisions inside software, not for chat. TypeSafe released it on 15 September 2026 as the first public System One model after about two years in stealth, with a $40 million seed round reported by heise. You send a state (string, JSON object or text array) and a map of typed questions; the model returns values your code can branch on. It does not generate prose, explanations or tool-call strings.

That design is the point. An LLM optimised with RLHF writes the next token. Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD) so every answer is a distribution over options you declared in advance. TypeSafe’s launch post positions it as “frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.” Schema violations are structurally impossible; wrong-but-valid options still are. Treat outputs as probabilities with a fallback path, not as ground truth.

Dimension	Frontier LLM (e.g. Claude Sonnet 5)	Jev 1.13
Output	Generated text / JSON you parse	Typed Choice, Score or Noul
Typical latency	Seconds to minutes	70 to 500 ms (vendor-stated)
Input price	$2 / MTok for Sonnet 5 global	$0.042 / MTok
Output price	$10 / MTok for Sonnet 5 global	Free
EU processing option	Yes on Bedrock / Google EU	No; US-hosted API
Hallucinated fields	Possible	Impossible by construction

Sonnet 5 prices come from the Anthropic pricing page; the Bedrock EU geo premium of 10% ($2.20 / $11) is listed on Amazon Bedrock pricing. Jev figures are from the TypeSafe models page. Vendor speed and cost multipliers (up to 193.6x faster and 444.6x cheaper on System One workflows) are TypeSafe’s own evals against GPT-6 Astra and Claude Fable 5.1 averages; reproduce them before you budget.

Which three question types does the Jev API expose?

Every Jev call is built from three primitives. Mix them freely in one request; questions run in parallel against the same state, so adding questions costs tokens but barely adds latency.

1. Choice: pick one option from a fixed set (up to 255). Returns choice, a full probabilities map and confidence. Add an other option when the list might not cover the input.
2. Score: rate the state on 2 to 10 ordered levels you describe in words. Returns a score that may fall between levels, plus legend, probabilities and confidence.
3. Noul: yes/no as a probability from 0 to 1. Returns only noul; there is no separate confidence field because the probability is the signal.

The official spelling is Noul, not “Null”. The primitives docs and introduction are the source of truth. English is currently the strongest language; TypeSafe tells you to test non-English workloads and watch confidence carefully.

curl -X POST https://api.typesafe.ai/v1/systemone \
 -H "Authorization: Bearer $TYPESAFE_API_KEY" \
 -H "Content-Type: application/json" \
 -d '{
 "model": "jev-1.13.0",
 "state": "Help! My payouts have been failing for 3 days.",
 "questions": {
 "is_urgent": {
 "type": "noul",
 "instructions": "Does this convey urgency?"
 },
 "department": {
 "type": "choice",
 "instructions": "Which team should handle this?",
 "criteria": {
 "billing": "Payments, invoicing, refunds",
 "technical": "Bugs, outages, integrations",
 "sales": "Pricing, upgrades, new accounts",
 "other": "Does not fit the other departments"
 }
 },
 "frustration": {
 "type": "score",
 "instructions": "How frustrated is the customer?",
 "criteria": ["Calm", "Frustrated but civil", "Very angry"]
 }
 }
 }'

Pin jev-1.13.0 once you calibrate thresholds. The aliases jev-latest and jev-preview both resolve to that version today, but aliases move when TypeSafe ships. Context budget: 64k tokens for state plus all questions, 32k for state plus the longest single question. Text only; transcribe images and audio first. Early-access rate limits are about 250,000 tokens per second and 1,200 requests per minute.

Five patterns that fit EU product stacks without turning Jev into a chat replacement:

1. Speculative fan-out: ask category, urgency, PII and refund intent in one call; ignore answers your code does not need.
2. Confidence gate: act only when Choice confidence or Noul clears a threshold; otherwise escalate.
3. Composite score: split priority into severity, frustration and evidence quality, then weight in code.
4. Agent supervision: let Jev pick the next tool or decide that the agent is stuck; let Claude write.
5. Batch triage: map-reduce over tickets or leads where output tokens would dominate an LLM bill.

Python setup for a team prototype:

pip install typesafe-sdk
export TYPESAFE_API_KEY="sk-..."

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

with TypeSafeClient() as client:
 response = client.system_one(
 model="jev-1.13.0",
 state={"ticket": "Export produces empty PDFs since the update."},
 questions={
 "queue": Choice(
 instructions="Which queue should own this ticket?",
 criteria={
 "billing": "Payments or invoices",
 "technical": "Product bugs or integrations",
 "other": "Anything else",
 },
 ),
 "urgent": Noul(instructions="The message conveys time sensitivity"),
 "severity": Score(
 instructions="How severe is the reported defect?",
 criteria=[
 "Cosmetic; no user impact",
 "Degraded; workaround exists",
 "Blocking; no workaround",
 ],
 ),
 },
 )

if response.answers["urgent"].noul >= 0.8 and response.answers["queue"].confidence >= 0.7:
 route = response.answers["queue"].choice
else:
 route = "human_triage"

How should EU teams combine Jev with Claude?

We recommend a two-tier architecture: Jev for high-frequency, fixed-option judgments; Claude for generation, multi-step reasoning and low-confidence escalations. Ticket routing, lead scoring, prompt injection gates, tool selection and RAG chunk keep/drop are Jev-shaped. Drafting a reply, writing code or negotiating ambiguous policy is Claude-shaped.

A practical split for a European support stack:

1. Build a structured state (ticket text, account flags, policy excerpts) inside your EU VPC.
2. Ask every speculative question in one Jev call (category, urgency, PII flag, refund requested).
3. If confidence or Noul clears your threshold, act in code.
4. If not, escalate to Claude on an EU geo profile (Bedrock eu. or Google Cloud EU multi-region) for a reasoned answer.
5. Log both the typed Jev answer and the Claude transcript for audit.

Cost math at list prices: 1 million input tokens on Jev costs $0.042. The same million on Claude Sonnet 5 costs $2 globally or $2.20 on Bedrock EU, plus output. Decision traffic that previously burned Sonnet JSON-mode calls is usually where the savings land. Full Claude EU tables are in our Claude API pricing EU guide. For workloads that must never leave EU infrastructure, a self-hosted classifier is still the alternative; our DeepSeek V4.1 Flash self-hosted guide covers that path.

Gateways matter for ops, not residency. Jev is also listed as typesafe/jev on Cloudflare Workers AI (32k context on that route) and as typesafe-ai/jev on Vercel AI Gateway. Those still terminate at TypeSafe’s US service unless a gateway documents otherwise.

What does GDPR require when Jev runs in the US?

TypeSafe’s privacy policy states the Services are hosted in the United States and that using them from the EEA or UK transfers personal data to the US for storage and processing. It also states TypeSafe will not train or fine-tune models on customer Input and will not disclose Input except to service providers. The DPA (updated 24 April 2026) incorporates EU SCCs Module 2 (and Module 3 when you are a processor), names Ireland as the EEA supervisory authority, lists subprocessors at trust.typesafe.ai, and commits to security-incident notice within 72 hours. Enterprise zero data retention is mentioned on the models page; confirm the contract text before you rely on it.

Our position for EU teams:

• Do not send special-category or high-risk personal data until legal has signed the DPA and completed a transfer impact assessment.
• Prefer hashed IDs and redacted text in state when a decision only needs categories.
• Keep Claude generation on Bedrock or Google EU when the reply contains personal data; see our Claude GDPR provider comparison.
• Document confidence thresholds and human fallback as part of AI Act deployer controls; the calendar is in our EU AI Act LLM deployers guide.

Jev is early access. There is no published EU region, no open weights and no on-prem option in TypeSafe’s materials we checked on 20 September 2026. Speed and price do not create data residency. If legal blocks the US hop entirely, keep classification on an EU self-hosted model or a classical classifier trained on your labels, and keep Claude on Bedrock or Google EU for the generative lane.

When should you not use Jev?

Skip Jev when the job is writing, planning, multi-hop reasoning, creative work or any case where the option set is unknown. It cannot chat, explain itself in prose or read images. Vendor workflow accuracy sits near mid-tier frontier models on TypeSafe’s own harness and trails the strongest reasoning models on harder invoice-style tasks; always keep a Claude (or human) escalation lane. Also skip it when your compliance rule is “processing only in the EU” and you cannot accept SCCs to a US processor.

For EU Claude residency itself, pin eu. inference profiles as in our Claude on AWS Bedrock EU guide and the Amazon Bedrock EU overview. Jev does not replace that setup; it sits beside it for the cheap branch.

FAQ

What is Jev TypeSafe?

Jev is TypeSafe AI’s first System One model. It evaluates application state against typed Choice, Score and Noul questions and returns calibrated probabilities your code can act on. It does not generate free-form text.

How much does Jev cost compared with Claude Sonnet 5?

Jev lists $0.042 per million input tokens with free output. Claude Sonnet 5 lists $2 / $10 globally and $2.20 / $11 on Bedrock EU geo. Exact savings depend on token volume and whether you previously paid for Claude output tokens on classification calls.

Jev vs Claude: which should handle ticket routing?

Use Jev when departments and urgency levels are a fixed set and latency must stay under half a second. Use Claude when the ticket needs a drafted reply, multi-document reasoning or when Jev confidence is below your threshold.

Can Jev run with EU data residency?

Not today. TypeSafe hosts in the United States. EU teams need SCCs via the DPA, subprocessors review and usually redaction. For EU-resident generation, keep Claude on Bedrock or Google EU.

Does Jev hallucinate?

It cannot invent undeclared fields or options. It can still assign high probability to a wrong-but-valid option. Calibrated confidence and a fallback path are mandatory for production.

How do I get API access?

Join the early-access waitlist at typesafe.ai, create a key in the console, then call POST https://api.typesafe.ai/v1/systemone or use typesafe-sdk / @typesafe-ai/sdk. Pin jev-1.13.0 for reproducible thresholds.

Sources

1. TypeSafe AI: Introducing System One Models and Jev (20 September 2026)
2. TypeSafe docs: Models (Jev 1.13) (20 September 2026)
3. TypeSafe docs: Introduction (20 September 2026)
4. TypeSafe docs: Primitives (20 September 2026)
5. TypeSafe Privacy Policy (20 September 2026)
6. TypeSafe Data Processing Addendum (20 September 2026)
7. heise: Jev decisions instead of text (20 September 2026)
8. Cloudflare Workers AI: typesafe/jev (20 September 2026)
9. Anthropic docs: Pricing (20 September 2026)
10. Amazon Bedrock pricing (19 September 2026)