Run structured image and text decision queries with OpenAI using an llm plugin

You can call OpenAI's new Decisions model from the llm tool this week and include images; the API charges for input tokens and a community plugin makes testing immediate.

What actually changed

OpenAI published a Decisions-style API similar in concept to Jev and exposed a model called gpt-6-luna that accepts image input as well as text. Simon Willison has released an llm plugin, llm-openai-decisions 0.1a0, built after having GPT-6 Astra read the OpenAI documentation; the plugin follows the same decision question types as Jev (yes/no, choices, scores). The two models charge only for input tokens rather than output. You can install the plugin with llm install llm-openai-decisions and try an image query with the example command Simon publishes:

llm -m openai-decisions/gpt-6-luna -a https://static.simonwillison.net/static/2025/two-pelicans.jpg -s ' Does this image contain any mammals? '

The sample response in the post shows a structured JSON reply type, including a field labelled "predicate" in the example output.

Who it affects

• Developers and teams that already use Jev-style decision APIs or structured decision formats. The API supports the same question types you would expect from Jev.
• Teams that need simple, machine-readable answers from images as well as text, because gpt-6-luna accepts image attachments.
• Anyone using the llm command-line tool or the llm plugin ecosystem: Simon's plugin plugs this new API into that workflow.

What it costs or what it replaces

• Pricing: Simon reports OpenAI's Decisions model billing at 10 cents per million input tokens and notes Jev charges 4.2 cents per million input tokens. The post emphasises both providers charge for input and not for output.
• What it replaces: the API is described as conceptually similar to Jev rather than an outright replacement; the llm-openai-decisions plugin is modelled on the earlier llm-typesafe approach and provides a drop-in way to call the OpenAI Decisions endpoint from the llm tool.

What we don't know

• Exact latency and throughput characteristics of gpt-6-luna under load.
• Whether image content billing is treated differently from text input tokens in practice.
• Rate limits, regional availability and enterprise contract terms for the new Decisions API.
• Any accuracy or behaviour differences between Jev and OpenAI Decisions beyond the shared question types.
• SDK or official client support beyond the examples Simon included in his plugin and README.

What to do next

1. Install the plugin and run the example: llm install llm-openai-decisions then run the example image query Simon published to verify the workflow against your environment.
2. Read the plugin README and test the three question types (yes/no, choices, scores) with a small set of representative inputs and images from your own projects.
3. Estimate expected input-token volume and apply the published rates (OpenAI: 10 cents per million input tokens; Jev: 4.2 cents per million) to forecast monthly cost for typical usage patterns.