OpenAI's Decisions API: More expensive, slightly faster with images

The announcement. OpenAI's Decisions API, first shown during its Dev Day, is now publicly accessible to anyone with an API key. The service resembles Jev: a set of specialized models meant to answer simple questions with small, predefined outputs. It supports three question types: predicate (yes/no), choice (multi-class classification), and score (a numerical rating). The only model available is GPT-6 Luna, and OpenAI positions this as a tool for real-time decisions where latency around 100-200 milliseconds is acceptable.

Cost and capability comparison. The API costs $0.10 per million input tokens, which is 2.5 times Jev's $0.04 per million. Output tokens are free on both platforms. The creator shows that, for text-only tasks, running an open model locally via Ollama was faster (1.1 seconds) than the Decisions API (2.5 seconds) and Jev (2.9 seconds), as well as being free.

The trade-off. Teams already in the OpenAI ecosystem may benefit from minimal integration work. But the premium over Jev is meaningful, and the speed advantage over open models disappears entirely for text tasks. Image processing is the only scenario where OpenAI wins on speed, and that win is hardware-dependent: a local setup with a consumer GPU may reverse the numbers. The video makes no mention of rate limits, concurrency support, or batch processing.

What's missing. The video compares the Decisions API with Jev and Ollama, but does not evaluate a standard GPT-4o prompt structured for classification. That baseline would reveal whether the dedicated route offers meaningful latency or accuracy gains over a general-purpose model. Without that comparison, it is unclear if the Decisions API solves a real problem or merely carves out a narrower, pricier version of an existing capability.

The organizational impact. For engineering leaders, the decision hinges on how much they already pay for text classification. If a team runs hundreds of millions of classification calls per month, the increased per-token cost here versus Jev becomes a significant line item. The image capability is genuine and may justify the premium for pipelines that mix text and image classification. But for most workloads, the fastest and cheapest option remains a local open model or Jev.

Bottom line for teams. The Decisions API is a credible option for teams that already depend on OpenAI and need image support with low latency. Every other team should check whether Jev or a local model matches their requirements first. The video is a useful walkthrough of the API, but the final verdict supports adopting it only in specific, bounded scenarios where speed and multimodal support justify the cost.