Is TypeSafe's JEV worth a look?
JEV picks from fixed options, and this page explains its claims, evidence, and fit for your teams.
TypeSafe's JEVTL;DR
- 01JEV picks a structured answer from fixed options and returns a probability.
- 02It does not write text, so drafts and explanations still need a language model.
- 03TypeSafe produced the current speed and cost evidence. Independent benchmarks are not available.
- 04JEV is in early access. Test one high-volume decision with labelled outcomes.
Verdict
Too early
Test JEV on one narrow workflow. Do not plan a budget around it yet.
What holds up
- JEV returns a structured value and a probability.
- JEV does not write text. A language model must still produce drafts and explanations.
- TypeSafe offers JEV through an early-access waitlist.
What we doubt
- TypeSafe measured the speed and cost claims on workflows its model team wrote.
- No independent benchmark was available when we checked.
What JEV is
JEV takes unstructured input and a typed question. It returns a typed answer with a calibrated probability. TypeSafe describes it as a function call.[1]
TypeSafe calls this model class System One. The name refers to the fast, intuitive system in Kahneman's work. TypeSafe says it trained JEV with Reinforcement Learning for Calibrated Decisions.[1]
JEV cannot invent a paragraph because it does not write text. It can still choose the wrong option. Its probability shows confidence, not accuracy.
What TypeSafe says
- 1It does not generate text. Summaries, drafts, replies and explanations still need a language model.
- 2A single question can have at most 255 possible answers. Larger choice sets are handled in two stages.
- 3The public demo works on text and structured data, not images.
- 4Availability is early access through a waitlist.
The claims
| Claim | Vendor says | Our reading | Rating |
|---|---|---|---|
| Response time versus frontier models taking 3 to 329 seconds.[1] | 70ms to 500ms | JEV returns one structured value instead of many tokens. Small, fast models may reduce the reported gap. | Partly |
| Input price per million tokens, with output tokens free.[1] | $0.042 | TypeSafe lists this early-access price. The price may change after early access. | Holds |
| JEV is faster than frontier language models.[1] | 193.6x | TypeSafe tested workflows written by its own model team. The company says this result is near the high end. | Unproven |
| JEV is cheaper than frontier language models.[1] | 444.6x | TypeSafe compared JEV with frontier models. Independent tests have not confirmed the reported gap. | Unproven |
Independent evidence
No independent benchmark was available when we checked. The current evidence comes from TypeSafe's evaluation.
Which teams should care
- Customer supportRoute a ticket to a queue, set priority, detect a refund requestStrongJEV can choose a queue, priority, or request type.
- SalesScore an inbound lead, classify a call outcome, flag a stalled dealStrongJEV can choose a score, outcome, or risk flag.
- FinanceCode an expense line, flag an invoice exceptionStrongJEV can choose an expense code or exception flag.
- EngineeringCheck another model's output against a rule before it shipsStrongJEV can check an output against fixed rules.
- LegalClassify a clause type, triage a contract by risk tierPartialJEV can classify clauses, but a language model must draft text.
- MarketingWrite campaign copyNoneJEV picks from fixed options. It does not write copy.
Volume determines the value. Low-volume decisions produce small savings. High-volume or time-sensitive decisions make speed and unit cost more important.
What would change our view
How to test it
- 1Pick one decision you already make with a language model and already have labelled outcomes for.
- 2Run both models on the same inputs. Compare each result with your labels.
- 3Measure the latency your user feels and the cost per thousand decisions, from your own bill.
- 4Check how often high-confidence answers are correct.
- 5Decide what happens to low-confidence answers before you switch anything over.