OpenAI's Decisions API is generally available on GPT-6 Luna
AspireAI · OpenAI
Most of what an AI agent does isn't thinking - it's deciding: is this urgent, which tool runs next, does this look like fraud. OpenAI's Decisions API, generally available on October 22, 2026, is built for exactly those bounded questions. You define the question and the allowed answers, hand it text, images or JSON, and get back a probability (predicate), one of your values (choice, with probabilities and confidence), or a weighted score across a rubric. It runs on GPT-6 Luna, and OpenAI claims it is up to 10x faster than the same model through the Responses API. Customers quoted in the post: Datadog reports 40x faster p95 with comparable accuracy when screening agent-generated shell commands before execution; Sierra reports under 280 ms p95 and roughly 3x faster with 90% lower cost per call; Box reports about 5x faster routing at the same accuracy as the Responses API. These are vendor and customer numbers, not independent benchmarks. New since public beta: native GPT-Live integration for voice agents, prompt caching for reused context, US and EU inference residency, native JSON inputs, and up to 400 questions per request. Pricing reads $0.10 per million input tokens and $0.01 per million cached, with no output-token charge - but the model doing the actual work still bills normally, so the decision step is the cheap part, not the whole job.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.