What Jev is
A decision model. Text or JSON in, a typed answer out, with a probability per option and a confidence score. It does not write prose. TypeSafe calls this class of model System One.
| Question type | Asks | Returns | Example |
|---|---|---|---|
| Choice | Pick one option from a set (up to 255) | choice, probabilities, confidence | Which team handles this ticket? |
| Score | Rate against 2 to 10 ordered, described levels | score, probabilities, confidence | How relevant is this passage to the query? |
| Noul | Yes or no | probability from 0 to 1 | Does this message express urgency? |
One request can carry several questions against the same state. TypeSafe calls this fan-out; the state is read once.
Where it fits
Anywhere an LLM makes a short, bounded decision inside software. TypeSafe’s use-case map groups them into five families.
Automation
Route, classify, score or branch inside a workflow your code controls.
Real-time
Decisions inside UI and games, where multi-second LLM calls do not fit.
Map-reduce over data
Label or score large corpora where frontier LLM cost rules it out.
Verification
Guardrails, citation checks, hallucination and jailbreak screening.
Harness engineering
Model routing, context filtering and error detection around agents.
Common enterprise starts: support routing, moderation, entity matching, RAG passage filtering, claims triage, fraud alert routing.
Where it breaks
TypeSafe documents nine known limits for jev-1.13. Each has a mitigation. Most move work into code.
| Limit | What happens | Mitigation |
|---|---|---|
| Literal reading | Answers the written question, not the intent. | State exact conditions and edge cases. |
| Math and counting | Unreliable, worse as numbers grow. | Arithmetic and tallies in code. |
| Dates and times | Reads dates as text, not ordered values. | Parse dates in code; ask Jev only the judgment. |
| Indirection | Double negatives and multi-hop wording lower accuracy. | Ask directly; name the field that matters. |
| Large, noisy state | Irrelevant content distracts. | Filter fields in code before the call. |
| Adversarial input | Does not treat data as hostile by default. | Explicit criteria, adversarial test set. |
| Contradictory instructions | Conflicting guidance confuses answers. | Align instructions and criteria. |
| Cross-type consistency | No guarantee P(yes) + P(no) = 1 across question types. | Do not reuse thresholds between Noul and Choice. |
| Generation | Not trained to write text; slow and poor. | Keep writing on an LLM. |
Input is text only: strings, JSON or text arrays. Convert images, audio and PDFs first.
Confidence routing
The answer says what. Confidence says whether to act. Route on both.
These are TypeSafe’s starting points. Its routing pattern goes lower for low-stakes actions (act above 0.6) and higher for consequential ones (above 0.85, otherwise confirm). Set yours from calibration measured on your own labeled data.
# From TypeSafe's confidence-routing pattern
action = response.answers["intent"]
if action.confidence < 0.6:
route_to_support_agent(account_id)
elif action.choice == "approve_transfer":
if action.confidence > 0.85:
approve_transfer(account_id)
else:
ask_user_to_confirm("Approve this transfer?")Pricing and limits
| Model | jev-1.13.0. Aliases: jev-latest, jev-preview. |
|---|---|
| Input price | $0.042 per million tokens ($42 per billion) |
| Output price | Free |
| Context | 64k tokens per request. State plus the longest question: 32k. |
| Rate limits | 250,000 tokens/s and 1,200 requests/min, adjusted dynamically |
| Latency | 70 to 500 ms end to end. 0.114s median vs 8.566s for LLMs in TypeSafe’s workflow evals. |
| Customization | Same weights for every account. Adapt through state, instructions and criteria. |
| Access | Early access via console.typesafe.ai |
For comparison, current frontier list prices: Claude Fable 5.1 and GPT-6 Astra at $10 input and $50 output per million tokens. Jev’s input is 238× cheaper and its output is free.
Integration
REST
POST https://api.typesafe.ai/v1/systemone · Bearer token
curl -X POST https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"jev-latest",
"state":"...",
"questions":{"urgency":{"type":"noul",
"instructions":"Does this message express urgency?"}}}'Python SDK
pip install typesafe-sdk
from typesafe_sdk import Noul, TypeSafeClient
client = TypeSafeClient()
response = client.system_one(
state="your text here",
questions={"urgency": Noul(instructions="Is this urgent?")},
)
print(response.answers["urgency"].noul)- JavaScript / TypeScript: TypeSafe JS SDK.
- Vercel AI SDK: TypeSafe provider package.
- Coding agents: TypeSafe publishes a drop-in agent skill for Claude Code and similar tools.
- Existing LLM stacks: keep the LLM for planning and writing; call Jev at each decision point.
Vendor risk
Jev entered early access on 15 Sep 2026. Price the upside and the exposure together.
| Exposure | Control |
|---|---|
| Early access. No published SLA or plan tiers. | Current model stays hot as fallback. One flag routes all traffic back. |
| Price sustainability is unproven, by TypeSafe’s own account. | Check headroom: the brief shows how far Jev’s price can rise before savings reach zero. |
| Model updates change behavior. | Pin the version returned in each response. Re-run the labeled set before upgrading. |
| Rate limits under peak load. | Fan-out several questions per request. Request higher limits before cutover. |
| Data handling. | Review TypeSafe’s terms and privacy policy with your security team before sending production data. |
Migration playbook
- Pick one workflow. High volume, bounded answer, fallback already possible.
- Label. 20+ examples to start, per TypeSafe. A few hundred for a decision-grade result.
- Shadow-run for 1 to 2 weeks. Log Jev beside your current model. Switch nothing.
- Read the calibration. Accuracy by confidence band sets your thresholds.
- Automate low-risk paths first. Escalate the rest. Expand as measured accuracy holds.
Want this run for you? Two-week shadow evaluation, fixed scope.
Discuss your workflowCalculator method
- Today: decisions × (input tokens × input price + output tokens × output price), or your actual bill.
- Jev: every eligible decision is tried on Jev first. Jev cost = eligible decisions × Jev input tokens × $0.042 per million.
- Fallback: eligible decisions below your confidence threshold also run on the current model, at today’s cost per decision.
- Unchanged: ineligible decisions stay on the current model.
- Savings: today − (Jev + fallback + unchanged + added operations). Payback uses the one-time migration cost.
- Speed: automated path uses Jev’s 0.114s median. Fallbacks pay Jev’s time plus the current model’s.
- Not assumed: accuracy. A shadow evaluation measures it.
Sources
- TypeSafe AI homepage: latency and cost claims
- Introducing System One models and Jev
- Models: pricing, context, rate limits
- Primitives: Choice, Score, Noul
- Confidence and confidence routing
- Known limitations of jev-1.13
- Use-case map
- Quickstart and SDKs
- Claude pricing, GPT-6 Astra pricing, Gemini pricing