← WIZ
// EXPERIMENTS
REFLEX

TypeSafe shipped a model that does not write. You hand Jev some messy state and up to six typed questions, and it hands back typed answers with calibrated probabilities in about 400 milliseconds for a fraction of a cent. Kahneman would call it System 1. This page races it, floods it, and then puts it where it actually belongs: in front of the expensive thinking.

by Pawel Jozefiak

This is a recording of a real run

Not a simulation. Not numbers I made up and typed in because they looked about right. One run happened, against the real APIs, and this page replays what came back.

Every latency and every cost you are about to see was measured on 21 September 2026. Jev through OpenRouter's decisions endpoint, the challengers through chat completions on their default settings.

It is recorded rather than live because a public page that calls a paid API on every click is a bill with a stranger's finger on it.

the whole run cost
$0.162
real money, already spent
every jev call in it
$0.000946
all three acts, together
the challengers
$0.161
the same work, text models

The challengers cost 171x what Jev cost for the same work. That split is the argument the rest of this page spends three acts making.

1. Pick something that was judged

Route a ticket, decide if it can wait, read the temperature.

Order #48812 was supposed to arrive Tuesday. It is Friday. Tracking has said 'label created' for four days straight. I have emailed twice and nobody has replied to either one. I paid extra for express shipping, so I want that fee back at minimum, and if I do not hear from a human today I am calling my bank and doing a chargeback.

These five scenarios are the ones that were actually run. Pasting your own would need a live key on a public page, which is the thing this page is deliberately not doing.

teamchoice
Which team should handle this message?
billing / shipping / technical / account
urgentnoul
Does this need a human reply today?
angerscore
How angry is the person who wrote this?
calm → annoyed → frustrated → angry → furious

2. Pick the challenger

3. The race

JEVtypesafe/jev-1.13-20260917
0 ms
cost pending
GEMINI FLASHgoogle/gemini-3.8-flash
0 ms
cost pending

These are the measured latencies from the recorded run, replayed at real speed. Jev settles at 476ms, Gemini Flash at 2360ms. Each counter is real wall clock time in your browser, stopping where that side actually stopped. No request is being made.

What this is not

  • Jev cannot write a sentence. No summaries, no replies, no explanations. It answers questions you defined in advance and nothing else. If you need prose, you still need a text model.
  • The context window is 32k and a choice question tops out at 255 options. Long documents and high cardinality routing are both outside what it does.
  • The calls were not clean. TypeSafe was returning 529 overloaded on roughly four calls in five during launch week, and several of the Jev answers in this recording took more than one attempt to get.
  • The benchmarks are TypeSafe's own. Nobody independent has reproduced them yet. Treat the speed and price claims as measured here, on these calls, and nothing broader.
  • A calibrated probability is not the truth. A confident wrong answer is still a wrong answer. The confidence number tells you how often it should be right at that level, not whether this particular one is.
  • This is not a frontier model replacement. The argument on this page is the opposite: put it in front of one.

Sources

Every number on this page is read from recorded-run.json, captured on 21 September 2026 through a small proxy that held the API key. The page itself makes no network calls at all.