📰 Key Takeaways

At Dev Day, OpenAI quietly rolled out a “Decisions API,” revealed by CEO Sam Altman almost as an aside. The concept is similar to Jev, the model TypeSafe AI launched earlier this month — a “super-classifier” built on top of an LLM that lets developers hand it a set of options and get back fast, cheap, probability-based outputs.

Altman described OpenAI’s Decisions API as letting the company’s Luna model pick from a predefined set of options — say, sorting an image into a specific category, or deciding between different agent behavior modes. “By letting the model focus on that one choice, we can make it extremely fast while still keeping image understanding, broad language support, and safety guardrails intact,” he said.

Diogo Almeida, TypeSafe’s CEO and a former OpenAI engineer who co-invented reinforcement learning, joked on X that this is the start of a “clone war,” hinting that OpenAI’s move might signal that “building in a System One-compatible way” is where things are headed (TypeSafe calls fast, intuitive thinking System One, as opposed to deliberate System 2).

The logic here: full LLMs are too slow and too expensive for a lot of software use cases. Developers who pair an LLM with something like Jev find it’s faster and cheaper. Almeida says his company’s moat is using synthetic data to produce statistically meaningful outputs. One likely application for this kind of decision model is monitoring and safeguarding AI agents — OpenAI has recently turned to a separate model, run at “significant compute cost,” to watch for erratic behavior after agents went off the rails on the open web.

💬 JudyAI Lab Take

Sam Altman’s reveal of the “Decisions API” at Dev Day — letting the Luna model pick fast from a predefined set of options instead of running a full conversational inference pass every time — is basically the “super-classifier” concept TypeSafe’s Jev model already beat them to. It’s a sign the big labs are starting to confront the reality that full LLMs are too slow and too expensive for a lot of jobs.

For AI builders, this points to a split forming in how the industry thinks about design: not every use case needs an LLM’s full reasoning power. For “multiple choice” scenarios — image classification, deciding an agent’s behavior mode — a model focused on a single decision can be faster and cheaper while still keeping things like image understanding and multilingual support intact. TypeSafe calls this kind of fast, intuitive judgment System One, versus the more deliberate System 2 — and that division of labor is worth paying attention to, especially given OpenAI’s mention of agent behavior monitoring, where a low-cost decision model could end up being part of agent safety infrastructure too.

Next time you’re designing an AI application, it’s worth asking: does this step actually need full LLM reasoning, or is it really just a multiple-choice question?

📅 Source Info

🔗 Further Reading