OpenAI's Decisions API Went GA at $0.10 per Million Input Tokens. Here's When It Beats Jev and a Local Model, and When It Doesn't.
Follow-up to: 2026-09-26, "Someone Shipped Ollama for Jev." What is new: OpenAI's Decisions API became generally available on October 22, so you can finally compare rent versus run with real numbers instead of a limited-preview slide.
OpenAI's changelog lists the Decisions API as generally available on October 22, running on GPT-6 Luna. Two coverage pieces report the price as $0.10 per million input tokens with nothing charged for output. TypeSafe's Jev sits at $0.042 per million input on OpenRouter. Divide one by the other and you get roughly 2.4x. That is the premium, and the question is what it buys you.
What Decisions actually is
It is not a chatbot endpoint. You send questions and get typed answers back: a probability for a yes/no predicate, a pick from a fixed set of choices, or a score against a rubric. It is built for the boring high-volume work a product does all day, like routing a request, tagging a support ticket or deciding whether an input needs a human.
The changelog gives the timeline. Beta on October 6, a GPT-Live integration on October 15 so a voice session can call it directly, and GA on October 22. Since beta, shared input can be cached for prompt-cache discounts, images can go inside individual questions, structured data can be passed as JSON, and one request now carries up to 400 questions, up from 200.
The price, carefully
Here is where I would slow down. The changelog does not list a Decisions-specific rate. The only GPT-6 Luna rates it gives come from a September 22 entry: $0.10 input, $0.01 cached input and $0.50 output per million tokens. The "no output charge" claim comes from OpenAI's guide as reported by HighCircl and repeated by Crypto Briefing, and I could not confirm it on the pricing page. Open that page before you build a margin model on it.
Even so, the math at the input rate is gentle. Suppose your product classifies 2 million requests a month at 300 input tokens each. That is 600 million tokens, or about $60 at $0.10 per million and about $25 at Jev's $0.042. The gap is roughly $35 a month. If that number matters to your business, you have bigger problems than your classifier. If you are at 200 million requests a month, it is $3,500, and it matters a lot.
Speed is a vendor claim for now
OpenAI's launch material says up to 10x faster than the standard Responses API for bounded decisions, with demos around 150 ms. That is the vendor's number, with no stated baseline and no independent measurement that I can find. Jev's own numbers on OpenRouter, as of October 1, were a 0.21 second median and 0.34 seconds at the 95th percentile. Different measurement conditions, so I would not call either one faster until you run both on your own traffic.
Where the premium might be worth it
Images. Per the comparison I read, Jev takes text only. If your decision involves a screenshot or an uploaded photo, Decisions may be the only option that does it in one call.
Compliance paperwork. Decisions is described as supporting Zero Data Retention, with eligibility subject to OpenAI's approval, and as HIPAA-eligible under an executed agreement. If a client's procurement form has a box for that, being a single-vendor integration saves you a week.
One vendor to monitor. If your product already depends on OpenAI for generation, adding a second vendor for decisions is another contract, another outage page and another place for data to go.
Where it probably isn't
Raw price per decision at scale. Local is cheaper still. A decision model running on hardware you already own has no per-request bill, and the local port I wrote about last month shows it can be done. If your classification is simple and stable, a tiny local model with a rule-based fallback may beat everything.
Anything where accuracy has to be proven rather than assumed. These APIs return probabilities, and you pick your own cutoffs from labeled examples. Neither vendor does that part for you, and Jev has simply been generally available longer, since September 15.
The test I'd run before switching
Pull 500 real inputs from your product and label them by hand, which takes an afternoon. Run all three options, Decisions, Jev and a local model, against the same set. Record accuracy at your chosen threshold, p95 latency from your own server and the monthly cost at your real volume.
If the accuracy difference is under a point or two, pick the cheapest. If Decisions wins by five points on the cases that matter, pay the 2.4x. Do not make this call from a launch slide.
Where I could be wrong
Prices move. GA often comes with changes, and OpenAI has already adjusted limits once since beta. A free output token and cached input at $0.01 could make Decisions cheaper than Jev for workloads with a long shared prefix, which would flip my conclusion for some products. I have not run that comparison, and I would not trust anyone who claims they have without showing their traffic.
What I'd actually do
For a new product with under a million decisions a month, I would start on Decisions because it is the fewest moving parts, then log every input and output so I can swap later. For an existing product that already runs Jev or a local model and works, I would leave it alone and re-run the comparison in a quarter. The decision layer is the easiest part of your stack to change, which is exactly why you should not spend a week choosing it.
Author
Lukas
@lukcombinator