A post on OpenAI’s developer community forum presents the Decisions API as a public beta that any developer can try. The post is dated October 6 and appears under the handle sps. Nothing in it says who sps is or whether the account belongs to OpenAI staff. It reads like a release note, but the thread does not confirm that, so every claim below is the post’s own and has not been independently checked.
The API is built on GPT-6 Luna and accepts text and image inputs. Instead of writing an answer, it returns a judgement about what you send. The post describes three kinds of result. A predicate takes a statement and returns the probability that it is true, which is a yes-or-no call with a confidence attached. A choice takes a list of options you supply and returns its pick along with how sure it is. A score rates an input somewhere along a numeric range you define.
The headline speed claim needs careful reading. The post says decisions arrive up to ten times quicker than sending the same question to Luna through the Responses API. That is a ceiling, not a typical result. It is measured against one specific route to one specific model, and the post publishes no latency figures, no test setup, and no independent benchmark. The only timing in the thread comes from a forum user who reported about 0.8 seconds for image inputs over a slow connection of roughly 3 Mbps. That is one anecdote, and it says nothing about the ceiling.
The pricing is the unusual part. Input costs $0.10 per million tokens, and that is the whole bill. OpenAI charges nothing for output tokens, and nothing for cache reads or cache writes.
The post does not explain why, but the product’s shape suggests an answer. A normal model call generates text, and generation is the expensive step, so output tokens carry the highest price. A decision model returns a probability for each option you hand it. There is essentially no generated text to bill, so the meter runs on input alone. That is an inference, not something the post states.
The missing cache charges are less generous than they sound. A forum reply says the Decisions API currently has no caching at all, and the replier guessed it might arrive later. Another developer replied that this makes switching classification workloads from cached Luna calls unattractive. Zero cache fees therefore reflects a missing feature rather than a discount, and a workload that repeats a long shared prompt may cost more here than it does today.
Public beta means unfinished, and the post says nothing about rate limits, stability guarantees, or planned changes. It also does not describe the model behind the endpoint beyond naming Luna. One forum user asked for a blog post on the architecture.
Early community testing points to a real caveat. A user named platypus ran a weighted coin that landed heads 70 percent of the time. Framed as a predicate, 1,000 trials returned heads about 70 percent of the time. Framed as a choice, the model picked heads 98 percent of the time. The same user later reported that the order of the options shifted the probabilities, and concluded for now to avoid choice questions. This is one person’s test, not an OpenAI finding.
The practical split follows from that. If your task is a yes-or-no gate, such as screening images or routing tickets, run your own sample through the predicate format and compare it with your current setup on cost and accuracy before trusting the 10x figure. If you need calibrated odds across several options, treat the choice format as unproven until OpenAI says otherwise.
Posted to OpenAI’s developer community forum on 6 October 2026 under the handle sps; the post does not identify the poster as OpenAI staff.
