OpenAI Decisions API: Verdicts, Not Prose, at $0.10/1M
The OpenAI Decisions API is the most quietly radical thing to come out of DevDay 2026. Instead of generating text, it answers one fixed question about your text or images — with a probability, a pick from your list, or a score. It runs on GPT-6 Luna, costs $0.10 per million input tokens, and charges nothing for output. For anyone building moderation, routing, or agent logic, this changes the unit economics overnight.
What the Decisions API Actually Does
Every other OpenAI endpoint returns prose you have to parse. The Decisions API flips that: you define the question and the finite set of acceptable answers up front, and the model returns a typed verdict your code can branch on directly.
Three verdict shapes, per OpenAI’s announcement:
- Probability — “How likely is this support ticket to churn?” → 0.82
- Choice — “Which category does this image belong to?” → one of your labels
- Score — “Rate this lead’s fit” → 7.4 out of 10
Because the answer space is bounded before inference — not parsed after generation — the round trip is dramatically faster. Early coverage puts it at roughly 10x faster than the Responses API for the same classification workload, since the model never generates tokens you throw away.
The Price That Breaks the Old Math
Here’s the number that matters: $0.10 per million input tokens, no output charge, no cache charge in public beta.
Compare that to what classification used to cost. Running the same verdict through a general-purpose model meant paying for both input and output tokens, plus the latency of generation, plus the engineering cost of parsing JSON that sometimes wasn’t valid JSON. The Decisions API collapses all three costs at once.
The Firecrawl team’s technical breakdown adds important nuance: this isn’t a new model. It’s a constrained interface over the existing GPT-6 Luna model — Sam Altman described it on stage as giving Luna “a predefined set of options to choose from.” Structured outputs constrain format after generation; the Decisions API bounds the answer space before inference. That’s where the speed comes from.
One caveat from the same analysis: as of late September, OpenAI hadn’t published the request schema in its API reference or SDKs. Treat the beta as moving fast and check the official OpenAI developer documentation before building production systems on it.
What This Means for You
If you’re a freelance developer, this is a new product category to sell. Every SaaS company with user-generated content needs moderation. Every support operation needs ticket routing. Every sales team needs lead scoring. Previously you’d quote these projects with “the AI part costs $X/month in API bills” as the scary line item. At $0.10 per million input tokens with no output charge, that line item rounds to zero for most workloads — which makes your proposal dramatically easier to close.
If you run an AI agency or build agents, the Decisions API is the missing control-flow primitive. Agents constantly need to make small routing decisions: is this email urgent? Which tool should handle this request? Did the last step succeed? Each of those used to burn a full model call. Now they’re sub-cent verdicts at 10x the speed.
If you’re a creator or operator (not a coder), the takeaway is simpler: the tools you buy are about to get cheaper and faster. Every app with “AI-powered moderation,” “smart routing,” or “auto-triage” just had its cost basis cut by an order of magnitude.
The Bottom Line
Watch the pattern: OpenAI is unbundling the model call into purpose-built endpoints. Chat for conversation, Codex for code, and now Decisions for verdicts. Each unbundling makes one specific job radically cheaper.
For builders, the move is straightforward: audit every place in your stack where you call a general model just to get a yes/no, a category, or a score. Each one of those is now a ten-cent-per-million-tokens API call.