The Jev AI Model Changes How AI Agents Decide

A new tool called Jev could change how agentic systems get built.

The Jev AI model comes from TypeSafe AI, and it's a decision-making model made to run inside software. Until now I've asked one powerful LLM to do all of it:

• Read the message.
• Understand the context.
• Make the decision.
• Figure out what to do next.

A big share of my agents' work is thousands of tiny decisions:

• Is this an investor or a founder?
• Does this message need a reply today?
• Is this lead worth enriching?
• Is this person in our ICP?
• Is this X post a real builder or engagement bait?

Using Claude for every one of those decisions is like hiring a lawyer to sort your mail.

What the Jev AI model does

You give Jev a list and a question. It makes a decision for every item and returns a probability plus a confidence score. The question comes in one of three shapes:

• Classify: pick one label from a set you define.
• Score: rate on ordered levels you define.
• Yes or no: answer with a probability.

TypeSafe calls this class System One models. They opened early access on September 15, and The Rundown has a short news summary. According to TypeSafe's launch post, pricing is $0.042 per million input tokens, and output tokens are free. The TypeSafe docs cover the full API.

TypeSafe also claims Jev is 40x to 200x faster than frontier models. That's their number, and as TrueFoundry's breakdown points out, it comes from evaluations TypeSafe designed itself that haven't been independently reproduced. I haven't verified it either. Everything below is what I measured myself.

I plugged it into Orbit and tested it on real data

Three small jobs first, all on data Orbit already held. Nothing was scraped for the demo.

• Inbox triage. 12 WhatsApp messages from a VC pitch group: who is this from, and does it need a reply today?
• Lead scoring. 10 founders scored against our ICP, plus a yes or no on whether each is a decision-maker. Six of the ten passed my rule (good fit and decision-maker at 0.7 or above).
• Social filter. 10 X posts. Four came from builders with real problems, five were engagement bait, and one was an off-topic post about FBI agents. Jev sorted all ten the way my Claude agent had sorted them by hand.

Job	Items	Time	Cost
Inbox triage	12	2.1 s	$0.00021
Lead scoring	10	1.9 s	$0.00018
Social filter	10	2.0 s	$0.00018
Total	32	about 6 s	$0.0006

That's 13,677 input tokens for the whole run.

Small demos flatter a model, so I ran a bigger test. I had 188 messages already classified by Claude Sonnet. Jev agreed with Claude on 148 of them, which is 79%, and the whole batch cost about $0.003. That's a fraction of what we pay Claude for the same calls.

About half of the disagreements came from my own test design. Claude had seen the whole thread and knew a sender was the event organizer. Jev only saw a one-line summary of a single message. If you use it, give it the context a person would need.

Confidence gives agents a decision hierarchy

The confidence score is the useful part. Average confidence was 0.86 when Jev agreed with Claude and 0.66 when it disagreed, so an agent can route on it:

• High confidence: act automatically. We use 0.8 as the line.
• Low confidence: send it to a more powerful model.
• Still uncertain: ask a human.

One limit to keep in mind. A few wrong answers still came back above 0.95, so keep the automatic tier for actions you can undo.

A specialized decision layer changes the economics

That means we don't need our most expensive model for every tiny decision an agent makes. We can give agents a specialized decision layer, and I think that changes the economics and the architecture of agentic systems. The pattern is simple: Jev decides, an LLM writes, and confidence decides when a human looks.

Three plays that fit it:

• Score a lead list before you pay to enrich it. If you run outbound like the LinkedIn method behind 20 qualified conversations a day, scoring first means you only enrich the people who pass.
• Tag an inbox and reply only where it matters.
• Turn 300 scraped posts into the 12 worth reading.

It works best when your agents share context, because the threshold and your ICP definition then live in one place. That's the idea behind shared memory for AI agents.

What Jev can't do

Jev can't write a reply, summarize a thread or run an agent. It won't stand in for Claude on anything that needs reasoning or prose. It also gives you a confidence score but no written explanation of why it scored a case the way it did, which DataCamp flags as a problem if you need an audit trail. With one thin line of context its confidence drops, and it tells you so, which is the honest behavior you want from a model that makes decisions.

Jev is now live in Orbit

Any agent or any user can call it to classify, score or make yes or no decisions across up to 500 items. No API key. No setup. We wired it in ourselves through TypeSafe's API, and it's a tool called jev in the Orbit toolbox.

• Three actions: classify, score and yes or no.
• Optional shared context, such as your ICP, and a minimum-confidence setting that flags rows for review.
• Metered like every other provisioned tool, with a one-credit minimum per call.

Because it sits behind Orbit's MCP, you can use it from Claude, Codex, ChatGPT, Hermes and more. If you want to call TypeSafe directly instead, start with their docs. Orbit is the agentic operations layer for Claude, Codex, ChatGPT and more.

I expect we'll see a lot more agents built this way, and it should speed up adoption. One powerful model shouldn't need to make every decision. Try it in Orbit.