Assemble a 3D scene from assets that already exist
A 3D assembly lead centered on choosing and arranging existing assets rather than generating every object from scratch.
Published 2026-09-21 · Free case note
Jev: small decisions, new kinds of software.
A free collection of creator projects using Jev for selection, routing and interaction. See what exists, what it needs, and what has only been demonstrated.
In focusOpenAI DevDay · what launched →Personal agents compared →Sonnet 5.5 · released Sep 28 →GPT-6 Sol + Luna · official release →Opus 5.5 · launch & reactions →Jev · what people are building →Free access & alternatives →
New this week · checked October 3, 2026
The decision-model race: Jev, OpenAI and Cloudflare
Available
Jev · TypeSafe
The model that started the category: typed questions in, calibrated probabilities over fixed answers out.
Vendor comparison: Median latency 524 ms in Cloudflare’s comparison.
Source (opens in a new tab)Limited preview since Sep 29
Decisions API · OpenAI
Runs on GPT-6 Luna. You define the questions and the allowed answers; it returns a choice for classification, routing or an agent’s next step.
Early test · one task: In one task tested by Every: right in 76 of 78 steps vs 73 for Jev, about 230 ms vs 500 ms.
Source (opens in a new tab)Available since Oct 1 · open weights
Clef and Clef-flash · Cloudflare
Cloudflare’s first in-house models (27B and 9B, built on Qwen). Jev-API compatible, 64K context, up to four images per request, Apache 2.0 weights on Hugging Face.
Vendor-reported: Median latency 209 ms (Clef) and 39 ms (Clef-flash) vs 524 ms for Jev; beats Jev on most of Jev’s own index, by Cloudflare’s tests.
Source (opens in a new tab)On Workers AI, one independent review lists Clef at $0.24 and Clef-flash at $0.09 per million input tokens, with output not billed; Cloudflare’s launch post does not state a price. The review ↗
Our read · 01
Three vendors in one week
A niche model type became a product category: OpenAI and Cloudflare both shipped one within days, and Cloudflare copied Jev’s API so switching takes one line.
Our read · 02
Open weights change the math
With Clef on Hugging Face, you can run decisions on your own hardware. For high-volume routing, that can beat any per-token price.
Our read · 03
For builders: don’t marry one vendor
Put decisions behind one interface and test each model on your own labelled examples. Speed and price differ more than accuracy does.
Not another chat window.
TypeSafe describes Jev as a model for typed decisions with confidence estimates. Think choosing an action, ranking candidates or classifying a row. Writing a reply, seeing a screen or executing an action may require other parts of the stack.
TypeSafe’s model description (opens in a new tab)Public selections
Browse 12 Jev selections From idea to something you can inspect
Turn clipboard context into suggested form entries
A form-assistance experiment uses clipboard context to suggest values for fields, reducing repeated transfer from copied text into structured inputs.
Published 2026-09-22 · Free case note
Draft replies beside the conversation, but leave sending to the person
An Android conversation copilot reads visible chat context, ranks reply drafts, and fills the chosen text without pressing send.
Published 2026-09-22 · Free case note
Search with visible sources instead of generating an answer
A search interface lets Jev choose queries, sources, and time ranges, then ranks links without composing a synthetic answer.
Published 2026-09-17 · Free case note
Choose browser actions from an indexed control table
A browser-agent loop asks Jev to choose an operation and live DOM target while a separate model supplies text only when needed.
Published 2026-09-17 · Free case note
Compose an interface from components the application already trusts
An experimental composition path lets Jev select and arrange catalog components instead of freely writing UI code or arbitrary properties.
Published 2026-09-18 · Free case note
Fork a game state, try several futures, and keep one timeline
A game experiment pairs Jev move selection with branchable virtual machines that fork after failure and preserve a surviving timeline.
Published 2026-09-18 · Free case note
Let the form choose its next useful question
A branching form uses Jev to choose the next question while json-render supplies the interface and XState owns application state.
Published 2026-09-18 · Free case note
Name a spreadsheet column, then classify rows by that intent
A spreadsheet experiment treats a heading such as “Urgency” as an instruction for classifying each row into useful labels.
Published 2026-09-18 · Free case note
Classify construction-plan sheets before downstream estimation
A construction workflow uses typed decisions first to triage plan sheets, then to classify bill-of-materials items for downstream estimation.
Published 2026-09-19 · Free case note
Use voice and outfit context to choose the next virtual look
A virtual-try-on experiment combines spoken intent and the current outfit to select another item from a digital closet.
Published 2026-09-19 · Free case note
Coordinate a 3D character’s face, gaze, and body per message
A character-performance experiment asks ten questions per message, then maps probabilistic emotion and body-region choices onto an existing rig.
Published 2026-09-18 · Free case note
What we add to the links.
We separate the task, the role Jev plays, the rest of the toolchain and the unresolved conditions. All of this collection is free. There is no membership or model API access included.
More candidates at Jevable (opens in a new tab)