Almeida names two faults in OpenAI's Decisions API
An architecture objection, posted in a reply thread, with no benchmark numbers attached from anyone.
Diogo Almeida has put a second, more specific layer on his objection to OpenAI's Decisions API, which he earlier summed up as Luna with constrained decoding. Pushed by Justin Schroeder to say what he actually meant, he named two things.
The first is that Decisions appears to run on a model shared with other objectives, Luna, rather than a dedicated one, which in his words can cause jaggedness. The second is that it applies constrained decoding on top of string outputs instead of being, as he puts it, more machine-native.
Why he thinks constrained decoding hurts
Almeida has made the underlying argument before, in a reply to @LukeberryPi. Constrained decoding works by stripping probability from tokens that would make the output invalid. His point is that the need to strip anything is itself the signal. If the model ever assigns probability to an invalid token, the model is confused, and masking that confusion pushes the generation further out of distribution rather than fixing it. He says he will write a longer post on it one day.
That is the whole of the technical case as posted. It is an argument about how the thing is built, not a measurement of how it behaves.
How this started
The thread kicked off when Ian Nuttall posted the launch and asked whether the Decisions API, built on Luna, is a Jev killer. Maz posted it the same way, with the line that OpenAI said hey guys we have jev too. Almeida's reply came in under that.
For people building on it
Nobody in this thread has posted a number. There is no accuracy figure, no latency figure and no cost figure for Decisions, and no public run of Decisions against Jev on the same task. What you have is a well formed architectural objection from someone who has been consistent about it, which is worth something when you are deciding what to test first, and worth nothing as evidence that one is better than the other.
If you are evaluating Decisions for a routing or classification path you already run through Jev, the two claims to check yourself are the ones Almeida flagged. Does a shared base model produce uneven results across your task mix, and does forcing valid structured output degrade answers that the unconstrained model would have got right. Both are measurable on your own data in an afternoon, and so far that measurement is the thing missing from the argument.
constrained decoding often makes models dumber for reasons I've talked about previously
if EVER your model assigns ANY probability to invalid tokens, by definition, your model is confused
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.
