OpenAI has opened its Decisions API to developers in public testing and added the ability to receive images to make fast decisions from visual content. The public version runs on the GPT-6 Luna model and enables text or images to be converted into three types of structured outputs: probabilities about whether a particular statement is true, a choice from a specified list, and a numerical score that determines the input’s position within a range.
OpenAI prices the API at $0.10 per million input tokens, with no fees for output tokens or cached-memory read and write operations. This model places the service in a different position from general reasoning models, as the API focuses on issuing a specific decision quickly rather than producing a lengthy answer.
From Games to Robots
In a demonstration presented by OpenAI, the API analyzes frames from a video game featuring a car moving amid traffic, then chooses whether the car should change lanes or continue. The company says that making these decisions takes a fraction of the time a reasoning model might need to perform the same task.
The company also demonstrates using the API with camera footage from a programmable bipedal robot from Hugging Face called Microduck. By combining Decisions API with the real-time voice model GPT-Live, the system can identify the position of a piece of fruit in an image and direct the robot toward it in response to a command such as “follow the apple.” It can also handle a more ambiguous command such as “follow the fruit.” Other examples include selecting expressions for an animated character during a voice conversation.
Competition Between Hosted and Open Models
OpenAI’s move does not come in a vacuum. Perplexity has launched the pplx-decider-v1-27b model on Hugging Face under an Apache 2.0 license. It is a model fine-tuned from Qwen3.8-27B. According to the model card, it achieved 85.71% across 11 benchmarks, compared with 84.51% for TypeSafe’s Jev model, while Jev outperformed it on six of the 11 benchmarks, including WinoGrande, BBH, and TruthfulQA binary.
Amazon introduced the downloadable Strands Decider 2B model, built on Qwen3.5-2B, and said that it ranks second among general models with around two billion parameters on the public JevBench dataset. Cloudflare, meanwhile, launched the Clef and Clef-flash models with open weights and an Apache 2.0 license, while also making Clef available through its Workers AI service. Clef and OpenAI’s API support images, whereas Jev and Strands Decider remain dedicated to text.
What Matters to Developers?
The most important change is not limited to the addition of images, but also concerns how the model is deployed. OpenAI’s API is available as a hosted service, while Perplexity, Amazon, and Cloudflare allow developers to download the weights and run the models on their own infrastructure. As a result, considerations such as where execution takes place, control over data, and operating costs will join accuracy and speed in determining which decision model to choose.
The reported results remain tied to the benchmarks tested by each party, and the material does not provide a comprehensive independent comparison of response time, operating costs, or visual performance. Based on the available information, Decisions API expands the use of decision models to visual and interactive applications, but it does not yet settle whether the convenience of a hosted service will outweigh the flexibility of local operation.