Chat that thinks while it talks.
While one model writes your answer, a second one reads it as it comes back. That second model is Jev, a decision model: it never writes a word. It reads each sentence as it arrives and classifies it: the substance you asked for, extra detail you didn't, or padding such as filler, hedging and repetition. You can see at a glance what is worth reading. It also decides which model should answer and which follow-up questions to offer. Each decision takes a fraction of a second, and Reflex shows you every one.
What Jev decides on every turn
Jev reads your message and decides whether the cheap writer is enough. If it is unsure, the turn escalates to the big model. This is where most of the saving comes from: the small model costs about a tenth as much.
Each sentence is labelled as it arrives, three per call: substance (what you asked for), extra (true and related, but more than you asked), or padding (filler, hedging, repetition). Once the answer is complete, Jev also picks the one key sentence and it is highlighted. The answer restyles itself as you read, so you can see what matters.
With every batch Jev is also asked whether your question is now fully answered. At 90% sure, the stream is cut. In practice this rarely fires: see the note below.
The writer drafts five follow-up questions; Jev scores them and keeps the best three. Selecting is a decision, so Jev does it. Writing is not, so it doesn't.
Jev does not measure carbon: the estimate is arithmetic on token counts. But its decisions shape the footprint and explain it. Routing keeps most turns on the small model, and the difference from always using the big one is shown as carbon not emitted. Its sentence labels show how much of each turn's footprint went on padding rather than on the answer. And Jev's own calls are counted in the total, so its overhead is never hidden.
Is it fast enough to cut an answer short?
Partly. A verdict takes 0.3 to 0.6 seconds to come back, and the writer streams roughly a hundred tokens a second, so thirty to sixty more tokens appear before a cut can land. Reflex keeps any whole sentences from that window and drops a half-written one, so a stopped answer never ends mid-sentence.
The bigger limit is not speed. Across the test questions we ran, Jev only became 90% sure an answer was complete once the writer had, in fact, finished. The writer is told to be concise, so there is rarely a tail of recap to cut. Early stop is a safety net for answers that ramble, not the main saving. When it does nothing, the app says so rather than claiming a stop.
A decision model from TypeSafe. You give it some text and a typed question (pick one of these, yes or no, rate on this scale) and it returns the answer with a probability. It cannot write a sentence.
About a quarter of a second per call once a connection is open, up to 0.7 s on a cold one. Its servers are in Oregon; most of that time is the round trip from Europe, not the model.
$0.042 per million input tokens, output free. A typical turn spends less on Jev than a hundredth of a cent, but that is often more than the cheap writer's answer cost. Judging every sentence is not free, and the app shows it.
Everything is on the table
Beside the chat, a panel lists each decision as it fires, with its latency and the probabilities behind it. Every answer carries its token count, its cost and an estimate of its carbon. The conversation is also drawn: a tree that grows a branch per turn and a leaf per useful sentence, above layers of earth whose thickness is the estimated CO₂ each turn cost. The carbon figures are rough ranges built from published estimates, and the app explains how.
Writer: z-ai/glm-5.3-flash, escalating to z-ai/glm-5.3, via OpenRouter. Decisions: Jev (jev-latest) on TypeSafe's API. Costs are as billed; comparisons price the same tokens on GLM 5.3 with no decision layer.