- Published on
Llm-openai-decisions 0.1a0
- Authors
- Name
- DeepHorus
OpenAI released its new Jev-style Decisions API on 6th October 2026, following the company's announcement at DevDay the previous week. The API introduces a structured approach to asking models discrete questions rather than generating free-form prose. Developer Simon Willison quickly shipped a companion plugin, llm-openai-decisions 0.1a0, that connects the API to the LLM command-line tool.
The plugin was generated by having GPT-6 Astra read OpenAI's new API documentation and build an implementation inspired by llm-typesafe, Willison's existing plugin for talking to Jev. That workflow mirrors a growing pattern in which developers use capable models to scaffold integrations against fresh documentation. The result is a working plugin available immediately alongside the API launch.
A key differentiator is multimodal input. Unlike Jev, OpenAI's new gpt-6-luna decision model accepts images in addition to text, widening the range of questions developers can pose. Both services charge only for input tokens, leaving output effectively free. OpenAI's model costs 10 cents per million input tokens, while Jev charges 4.2 cents per million, making Jev the cheaper option on paper.
Conceptually, the two APIs are closely aligned. Jev supports three question types covering yes/no decisions, multiple choices, and scores, and OpenAI Decisions supports the same three categories. That symmetry should ease migration for teams already using Jev-style decision tooling. Installation is a single command: llm install llm-openai-decisions.
A sample query demonstrates the image capability. Passing a photograph of two pelicans and asking whether the image contains any mammals returns a structured JSON response: {"type": "predicate", "name": "evaluation", "probability": 0.0}. The model correctly assigns zero probability, showing how decisions are returned as typed, machine-readable objects rather than conversational text.
For developers building classification, moderation, or routing systems, decision-style APIs offer more predictable outputs than chat completions. The pricing structure, which charges only for input, further favours workloads where short, repeatable queries dominate. Willison directs users to the README for details on running the remaining question types.
The launch positions OpenAI directly against Jev in a niche but expanding category of model interfaces. With the plugin already published and image support in place, adoption will likely depend on whether developers value OpenAI's broader ecosystem over Jev's lower per-token cost.
