Jev did not stay alone for long. Within three weeks of its September 15 launch, Cloudflare, Liquid AI, Perplexity, and OpenAI had each shipped a model that answers typed questions with probabilities instead of generating text. The category is now real; the ranking is not settled, and the gap between the two is the story.
The rivals
The loudest challenge came from Cloudflare's Clef, a pair of open-weight models released on October 1: Clef, a 27B model that reads text, JSON, images, or video, and Clef-flash, a 9B model. Across 43 of its own benchmark runs, Cloudflare reports a 209.3 ms median for Clef and 38.8 ms for Clef-flash against 524.1 ms for Jev. A week later it added Clef-omni, which accepts audio and video, and cut Clef-flash's input price below Jev's.
Liquid AI's d1 took a different route to the same claim. Liquid says d1 was the first model to outperform Jev on Hugging Face's Decision Index, matching or beating GPT-6.1 Sol on four of six tested applications at 19x to 200x lower cost, and it supports image input. OpenAI's Responses-based Decisions API arrived in public beta, pitched as up to 10x faster than its own GPT-6 Luna. Perplexity shipped an updated open-weight decider.
The open reproductions
Behind the closed rivals sits an open layer that answers the same shape. Kev trains Qwen models from 0.8B to 27B to return typed answers; Bespoke Nimble, Strands Decider 2B, and Mapika's decider fill in other sizes. Because they speak the same System One wire protocol, an existing Jev caller often changes only a base URL and a key.
The benchmark politics
This is where the claims get slippery. FOMA's roundup ranks thirteen decision-model APIs and interleaves vendor figures with independent ones, which is useful as a map and misleading as a verdict. Cloudflare leads with a self-run index; Liquid leads with its own execution of the Decision Index; both are vendor runs of a public benchmark. The counterweight is the field: one Hacker News tester reported hosted Clef two to three times slower than Jev and worse on a hate-speech task, the opposite of the vendor table.
The community's Decision Index has tried to settle it by scoring open reproductions on a frozen suite with a private half, and Let's Data Science's timeline tracks how fast the copies arrived — dozens within two weeks, by TechCrunch's count. TechCrunch's own coverage of OpenAI's entry rounds out the picture.
What the convergence means
Almost all of these models now expose the same request: a state, a set of typed questions, and an answer with a probability per option. Switching is usually a base URL and a key. That convergence is good news for buyers and bad news for anyone hoping a benchmark would do the choosing, because it pushes the decision onto the things a vendor table hides: input modalities, price per million tokens, output charges, license, self-hosting, and calibration on your specific questions.
How to choose
- Fix the axis first. Image or video input narrows the field immediately; text-only keeps Jev and most rivals in play.
- Read the per-axis numbers, not the composite. A harmonic mean of intelligence, calibration, speed, and cost can hide a bad latency tail or a calibration miss.
- Measure on your task. Every vendor here runs its own benchmark; the only number that transfers is the one you produce on your own labelled set.
- Pin a version. Aliases like
jev-latestand rival equivalents move under you. - Check the license and the price. An open weight you can self-host and a hosted API with free output are different products, even when the request shape is identical.
A category that was one model in September is a dozen by October, and the next round will not wait. The primitive — state in, a typed decision out — has won. Which model answers it is now a procurement decision, and the evidence for it lives on your task, not on a launch post.
