Biggest LLM Release Week Ever!

Welcome back to the AI Edge (by Miles Deutscher) weekly newsletter!

This week is an exciting one. We’re covering the biggest LLM release week of the year.

GPT-6 Sol & Luna, Jev, Opus 5.5, and Grok 4.7.

For our main anchor piece, we’re focusing on the most exciting new model of the bunch, Jev.

Most AI models you use today focus on text generation. They write, code, research, reason, and return a text output answer.

Jev is a probability model created by one of the co-inventors of ChatGPT. You give it a decision and a set of outcomes, and it returns a score or probability.

200x cheaper than GPT, 400x faster, with a 0% hallucination rate.

IN TODAY'S READ

Jev Explained in 30 Seconds
Jev versus a Typical LLM
How Jev Works Under the Hood
How to Test Jev & 3 Workflows to Try Now
Weekly AI News Recap
Tools and Research Worth Your Time
New From AI Edge

After our main guide, stick around for the Midweek Edge for this week's AI news and Looking Ahead, where we compile our top AI research, trends, and workflows.

WEEKLY BUILD & WORKFLOW

Jev Explained in 30 Seconds

Think of Jev as an AI decision framework.

You give it possible decisions upfront, then Jev evaluates the outcomes and returns a probability score.

Real Example

Take a customer support ticket, for example. Instead of asking a model to write a long analysis, you could ask three narrow questions about the same message:

→ Which team should handle this? Billing, technical, or sales.

→ How frustrated is the customer? Calm, frustrated, or very frustrated.

→ Does this message need urgent attention? Yes/no.

Jev evaluates all three questions against the same ticket and returns the results in a single structured response.

Jev is basically a judgment model.

Jev versus a Typical LLM

The easiest way to understand Jev is to compare the job it does to that of a normal LLM like Claude.

At its core, Jev does not try to “predict” tokens the same way a model like Claude would.

Instead, it grades the fixed set of outcomes you provide.

More technically speaking, when you ask a typical LLM for JSON or other formats, it generates tokens, and your software still has to check whether the answer follows the format, whether the content makes sense, and what action should happen next.

	Typical LLM	Jev
Main job	Generate text, code, or reasoning	Make a defined decision
Output	Free form or structured generated text	Choice, score, or probability
Best fit	Writing, coding, research, long reasoning	Routing, classification, scoring, verification
Uncertainty	Usually added through prompting or another check	Included in Choice and Score outputs
Pricing	Depends on the model and both input and output usage	$0.042 per 1M input tokens, output is free
Speed	Depends heavily on the model and task	Published Jev tests land around 70 to 500 ms

How Jev Works Under the Hood

Jev has three question types: Choice, Score, and Noul.

Here’s how we recommend using each:

1. Choice

Use Choice when the answer should be one item from a fixed list.

Back to our support ticket example, you might ask: Which team should handle this ticket?

→ Billing

→ Technical

→ Sales

Jev returns the winning option, a probability for every option, and a confidence score.

2. Score

Score is best used when the answer is on an ordered scale.

For the same ticket, you could enter: “Score this customer's emotions”

→ Calm

→ Frustrated but civil

→ Very frustrated

3. Noul

Noul is Jev's probability of yes.

For example: “Is this message urgent?”

The result is a number from 0 to 1.

A value close to 1 means yes looks very likely.

A value near 0.5 means the model is unsure.

A value close to 0 means the evidence points toward no.

You can mix Choice, Score, and Noul in the same request.

Then use all the output scores from Jev to route the data to where it needs to go.

How to Test Jev & 3 Workflows to Try Now

The easiest place to test Jev is the Playground, where you can start running inputs right now.

You can also visit Typesafe directly to test Jev now.

4 WORKFLOWS TO TRY NOW

1. Model Router (save tokens)

This is a good prompt to try if you want to save tokens on your other LLMs like Claude.

You can put Jev before your main model call to have it judge the request's type and task difficulty.

Easy work can go to a cheaper model, hard coding can go to a coding specialist, while complex or high-risk work can be escalated to a frontier reasoning model.

Send this input structure to Jev:

State:

[Paste the incoming request]

Choice — Which model should handle this?

- Cheap model: simple or routine work

- Coding specialist: coding, debugging, or repo work

- Frontier model: complex or high-risk reasoning

Score — How complex is this request?

Low / Medium / High

2. Agent Guardrail

Use Jev to check an agent action before it runs.

For example, score whether a tool call matches the user request, whether a message contains sensitive information, whether a proposed action is risky, etc.

Input structure:

State:

User request: [paste request]

Proposed action: [paste action or tool call]

Noul — Does this action clearly match the user's request?

Noul — Does this action need user approval before it runs?

Choice — What should happen next?

Proceed / Ask for approval / Block and review

3. LLM Output Verification

By acting as a verification step following another model's response, Jev can help ensure output quality.

Use this:

State:

Original task: [paste task]

Relevant source/context: [paste source]

Model output: [paste answer]

Noul — Does the output follow the original request?

Noul — Are its factual claims supported by the supplied source?

Choice — What should happen next?

Accept / Recheck / Human review

10 Other Use Cases for Jev

We also found this cool Jev directory with 400+ use cases and prompts

MIDWEEK EDGE: Manually curated list of the most important weekly AI news updates

Weekly AI News Recap

→ Grok 4.7 is Live. SpaceXAI just deployed Grok 4.7. Available now in Grok Build, Cursor, and the xAI API.

TL;DR: One of the best value-per-dollar models on the market right now.

→ OpenAI Releases GPT-6 Sol and Luna. OpenAI just expanded the GPT-6 family with Sol and Luna.

Cheaper and faster versions joining Astra in the GPT-6 family.

→ Anthropic Releases Claude Opus 5.5. Opus 5.5 delivers roughly Fable 5.1-level performance while costing about 40% less to run than Opus 5.

Opus 5.5 requires less compute to serve and works longer within limits than Opus 5. It communicates more naturally in sessions and follows your writing rule much better. It feels closer to AGI than any other model

In the next section, we attached an Opus 5.5 vibe-coding prompt you can use now.

→ Muse agent adds Shopify checkout. Shop Pay is now directly in Muse. Find products and complete purchases across Shopify stores without leaving the agent.

→ Muse for Marketing. You can now connect Muse to 75+ marketing tools through Ryze, including Google Ads, Meta Ads, and more.

→ OpenAI Launches Astra for Law. A version of GPT-6 Astra, built for professional legal work with a dedicated US legal search index.

LOOKING AHEAD: Manually curated list of useful AI tools, skills, workflows & more

Tools and Research Worth Your Time

→ Fun Fact: 3 Companies vs 45 Years of Tech IPOs

At current and expected valuations, SpaceX, OpenAI, and Anthropic are collectively worth more than the combined first-day value of US tech IPOs over the last 45 years.

Pretty wild!