AI Decision Models: Jev vs Clef vs Strands Decider

On September 15, a San Francisco startup called TypeSafe AI came out of stealth with $40 million from DCVC and a model that can’t write a sentence. That was the point. Jev reads a chunk of state, like a support ticket or a step in an agent’s plan, and answers questions you’ve already written the answers to: yes or no, pick one of these four, rate this from one to five. It returns a probability for each option, and nothing else.

Sixteen days later it has company. OpenAI announced a Decisions API at DevDay on September 29. On October 1, Cloudflare released two open models called Clef and Amazon released Strands Decider 2B, also open. TechCrunch’s headline for the Amazon one called it a “Jev clone,” which is blunt but fair.

I read all four launch posts side by side. The questions below are the ones I had, and the answers are only as good as the companies’ own numbers, which is a problem I’ll get to.

What does a decision model do that ChatGPT can’t?

It’s faster, mostly, and it can’t go off script. A normal chat model generates text token by token, and if you want it to pick “billing” or “technical” you have to hope it does, then parse the answer. A decision model scores only the options you hand it, in one pass. Amazon’s engineers explain the trade in their own post: these models are “significantly worse at solving complex problems than reasoning models” and are no use for coding, chatbots or summaries.

So nobody’s replacing GPT-6 with one. The pitch is that an agent makes hundreds of small calls a day (retry this step or not, which tool next, does this need a human) and paying a big model to think about each one is slow and wasteful.

Who shipped what, and when?

Model	From	Released	Weights	Price
Jev	TypeSafe AI	September 15	Closed, API only	$0.042 per million input tokens, no output fee
Decisions API (on GPT-6 Luna)	OpenAI	Announced September 29, limited preview	Closed	Not published
Clef (27B) and Clef-flash (9B)	Cloudflare	October 1	Open, Apache 2.0	Free to download; also hosted on Workers AI
Strands Decider 2B	Amazon (Strands Labs)	October 1	Open, Apache 2.0, with training data and scripts	Free to download

OpenAI’s version got the least airtime. Sam Altman mentioned it almost as an aside at DevDay, according to TechCrunch, saying “by focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections.” There’s no public price and no docs page yet. Luna itself lists at 10 cents per million input tokens, which I covered in OpenAI’s GPT-6 price cut, but nobody has confirmed the Decisions API bills the same way.

Is the free stuff actually better than Jev?

Depends whose chart you read, and right now almost every chart comes from someone selling something.

Cloudflare says Clef beats Jev on 3 of 4 workflow tests from TypeSafe’s own eval suite: invoice processing, customer service and security incidents. Jev won the fourth, agent trace observability, 71.6 to Clef’s 68.5. Cloudflare also measured median latency of 209 milliseconds for Clef, 39 for Clef-flash and 524 for Jev across its runs. That last number is odd, because TypeSafe says Jev answers in under 100 milliseconds. Network distance, hardware and prompt size can all move these numbers a lot, and Cloudflare’s test was run by Cloudflare.

Jev also still wins some of Cloudflare’s own rows. On When2Call, a test of whether a model knows when to call a tool, Jev scored 80.97 and Clef 72.37.

Amazon’s claims are smaller and, honestly, easier to trust. It says Strands Decider ranks 3rd of 33 models in its size class on the public JevBench set, gets every easy task right, and makes a decision in a median of about 115 milliseconds on an Nvidia RTX 3090 (about 153 on an M3 MacBook for small tasks). VentureBeat looked at Amazon’s chart and found it puts the model around 72% accuracy. A competing open model, Mapika’s decider-2b, sits higher at about 76%. Jev isn’t on that chart at all, so there’s no head-to-head.

Every benchmark figure here is self-reported by the company that built the model, or by its direct rival. I couldn't find an independent comparison of all four yet.

How did three big companies copy it so fast?

Because the recipe is cheap. Clef and Strands Decider both start from Alibaba’s Qwen models. Amazon took Qwen3.5-2B, cut off the part that predicts the next word, and bolted on a “pointer head” of just over a million parameters that scores each option. The version it released is v19.

Marc Brooker, the Amazon distinguished engineer behind it, started it as a side project after seeing Jev, according to TechCrunch, and it briefly topped the JevBench ranking for its size. TechCrunch, summing up his view, wrote that for smaller markets the cost of building something interesting is in the hundreds or thousands of dollars. That’s a striking thing to hear from an Amazon engineer about a category a startup raised $40 million to open.

What does TypeSafe say?

Its CEO, Diogo Almeida, worked on RLHF, InstructGPT and ChatGPT at OpenAI before starting the company in 2024. He isn’t rattled, at least in public. “I get that people think it’s a gold rush, but they might be underestimating the difficulty of making the models actually smart,” he told TechCrunch. He called the new batch “more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful.”

He might be right about quality. Calibration, meaning whether “85% sure” is really right 85% of the time, is the hard part of these models, and it’s the part a weekend project is most likely to fudge.

My read

I don’t think TypeSafe’s problem is that its model is worse. It may well be better. The problem is that Jev is a hosted API, and two of the biggest infrastructure companies in tech just made a decent version of the same thing free to run on your own GPU, with Amazon even publishing its training data. When the copy is free and the original costs 4.2 cents per million tokens, buyers start asking whether “better” is worth a vendor dependency.

Then again, 4.2 cents is already close to nothing, and running your own GPU isn’t free either. VentureBeat made the same point: Amazon hasn’t shown that self-hosting beats Jev’s bill once you count hardware and ops. For a small team, a cheap API you don’t maintain can still win.

What I’d watch is OpenAI’s price. Once the Decisions API leaves preview with a number attached, we’ll know whether this category is a business or a feature.

A startup's AI model got three rival versions in 16 days. Two of them are free.