Key Takeaways

  • OpenAI introduced the Decisions API at DevDay 2026, creating a direct competitor to Jev for fast, constrained classification.
  • The API runs on Luna intelligence, restricting outputs to predefined answer sets to cut latency and prevent hallucination.
  • Computer vision gives the Decisions API an edge over text-only routing tools by analyzing visual inputs directly.
  • In live testing, Claire Vo parsed 100 video frames in under 10 seconds to isolate non-awkward podcast thumbnails.
  • The system returned instant single-image classification with 100% confidence in less than a second.

Constrained Decisions with Computer Eyes

Most developers build AI workflows by prompting giant models and praying the output stays in format. When you only need a categorical answer, that approach burns budget and adds seconds of latency you cannot afford. Classification engines like Jev solved part of this problem by restricting outputs to fixed answer sets, but they operated primarily on text.

OpenAI introduced the Decisions API to combine constrained classification with visual intelligence. Vo explains the underlying mechanics clearly: “This is basically Luna with constraints and so it's going to have Luna intelligence on a predefined set of answers and it can return decisions very fast.”

Adding visual capabilities changes what developers can route in real time. “I was able to test a little bit of the decisions API and it has vision. It has vision,” Vo noted. “So we get the speed cost of like a decisions API a Jev style model with eyeballs with computer eyeballs.”

Slashing Latency on High-Volume Media

Running standard multimodal models across large sets of visual assets usually stalls production pipelines. If you feed 100 frames into a general-purpose vision model, you wait minutes and pay full token generation rates just to get a yes-or-no label on each image.

Vo tested the Decisions API against real production media to select thumbnail candidates from video footage. “And so I built this with the decision API. It took less than I don't 10 seconds. It was like very very fast. And it went through 100 frames and found all the ones where we were not looking awkward.”

In a separate quick test, the model identified an object instantly: “It is 100% confidence in less than a second that it is a hot dog. So again AGI is here decisions API.”

When you force a model to select only from allowed values, you remove parsing errors entirely. You get clean categorical tags at the speed of a lightweight classifier, backed by the perceptual accuracy of Luna.

The Real Architecture Shift

For months, developers building complex agents have treated structured classification tools as the glue between specialized tasks. Routing requests, validating user uploads, and tagging raw data require quick determinism rather than creative text generation.

As Vo pointed out, “You all know how bullish I am on Jev. I think this is like the missing piece in a lot of our AI architecture. And so I'm really excited to see OpenAI announce this.”

If your product relies on users uploading images, receipts, or short clips, using a chat completion endpoint for categorization is waste. Switching classification tasks to a constrained decision endpoint drops infrastructure costs while improving reliability.

What to Do With This

Audit your application for any prompt that takes an image and returns a fixed enum or status tag. Replace that general-purpose vision call with a constrained Decisions API call that limits the response schema to your exact enum values, and measure the latency drop across your next 500 requests.