OpenAI's Decisions API Costs $0.10 Per Million Tokens: Route 10,000 Tickets for $0.50

Most AI pricing bills you twice: once for what you send and once for what the model writes back. OpenAI's Decisions API breaks that pattern. It charges only for input tokens, $0.10 per million uncached, $0.01 per million cached, and zero for output. It went generally available on 22 October 2026, after a public beta that started on 6 October, and it is built for one narrow job: making fast, bounded decisions inside software. If you run a business that sorts, routes, or scores a stream of requests, the math is worth doing properly, because the API cost turns out to be the smallest part of the bill.

What is the Decisions API?

It is not a chatbot endpoint. You send it an input (text, an image, or structured JSON) plus up to 400 questions in a single request, and it returns short, structured answers you can act on in code. It runs on GPT-6 Luna only, through a dedicated endpoint, and there is a Decisions Playground on OpenAI's platform where you can test questions against your own text or images before writing code.

The point is speed and structure. OpenAI says decisions come back up to 10 times faster than putting the same model to work through the Responses API, because the endpoint is built for bounded answers rather than open-ended writing. One company testing it in voice, Speak, reported time to next action falling from 1.4 seconds to 0.13 seconds in its own tests. Treat vendor and customer benchmarks as direction, not guarantees: your latency depends on your input size, region, and question count.

What are the three question types?

Every question you ask must fit one of three shapes, and picking the right shape is most of the design work.

A predicate question returns a probability between 0 and 1 that a condition holds, such as whether a product photo shows visible damage. A choice question returns one of your own preset values plus probabilities and a confidence figure, such as routing a ticket to billing, shipping, or returns. A score question grades input against an ordered rubric of 2 to 10 levels and returns a probability-weighted average, such as grading package damage from none to major.

Independent questions can share one input in a single request, which is where the cost math gets good. Questions that depend on earlier answers need separate requests, so a 400-question call evaluates shared evidence rather than running a sequence of dependent steps. Design your workflow around that: one call, many independent judgments on the same ticket, photo, or lead.

What does it actually cost?

Here is the full price list, with the fine print included, because the fine print is where budgets break. Uncached input costs $0.10 per million tokens. Cached input costs $0.01 per million. Output tokens cost nothing. Cache writes are billed separately, regional processing premiums can apply, and long-context multipliers can apply. For prompt caching, the reusable prefix must be at least 1,024 tokens, and question text does not count toward that threshold.

Now the ticket math, assuming 500-token tickets and uncached input:

Read that table twice. A small business with 2,000 support tickets a month pays about ten cents in Decisions API fees. A busy shop with 100,000 tickets pays about five dollars. The API is not the cost of a routing service. The cost is building it, testing it, and keeping it accurate.

What can a small business actually do with it?

Four jobs fit immediately. Support routing: classify every incoming ticket into billing, shipping, returns, or technical, and send each to the right queue with a confidence score, while low-confidence tickets go to a human lane. Lead scoring: grade every enquiry against your rubric and prioritize callbacks by score instead of arrival order. Damage and quality grading: score product photos against a severity scale for returns or insurance intake. Moderation triage: flag the small share of user content that needs human review instead of paying a human to read all of it.

The pattern is the same in each case: the API makes the first judgment at machine speed and machine cost, and humans handle the exceptions. That is also the honest limit. The API does not resolve tickets, close sales, or fix products. It sorts the pile so humans spend time on the items that need them.

How do you build a first version?

A working first version fits in a day or two of focused work, and the steps matter more than the code. First, define your questions and your allowed answers before touching the API, because vague questions produce vague scores. Second, test in the Playground with 50 to 100 real examples from your own history and check the answers against what your team actually did. Third, set thresholds: high-confidence answers route automatically, low-confidence answers go to a human, and you log every disagreement. Fourth, run it in shadow mode for a week, where the API judges but humans still decide, so you can measure accuracy before anything customer-facing changes. Fifth, turn on automatic routing only for the categories where shadow mode was right nearly every time.

Skip any of those steps and you get the classic failure: an auto-router that confidently misroutes a visible slice of customers while nobody watches the logs.

Could you sell this as a service?

Yes, and the economics explain why. The API cost for a client with 10,000 tickets a month is about $0.50, so the fee you charge is for setup, monitoring, and tuning, not for tokens. A realistic offer is a fixed monthly retainer for one routing workflow: setup, threshold tuning, a weekly accuracy report, and a human review lane for low-confidence items. What you cannot honestly promise is a number: no guaranteed response time improvement, no guaranteed savings figure, no guaranteed accuracy. Accuracy depends on the client's data, and the first month is calibration, not magic.

The real effort is also easy to underestimate. You need to collect labeled examples, handle the categories the API gets wrong, explain the system to the client's team, and revisit thresholds when their products or policies change. Clients pay for that ongoing care. The API being nearly free is what makes the service viable for small clients, not what makes it easy.

What are the honest limits?

Four limits deserve space. First, it needs development work: this is an API, not a plug-in, and someone has to build and maintain the integration. Second, accuracy needs calibration: scores are probabilities, and a threshold that works for one business misfires for another. Third, the fine print costs: cache writes are billed, regional processing premiums apply, and long-context inputs cost more, so a workload built on huge shared documents will not match the headline $0.10 rate. Fourth, it is one model, GPT-6 Luna, so you cannot swap in a different model if Luna's judgments do not fit your domain.

The bottom line

The Decisions API is the cheapest serious decision-making endpoint OpenAI has shipped: $0.10 per million input tokens, nothing for output, 400 questions per request, and up to 10 times faster than the general Responses API by OpenAI's own comparison. For a support routing job, 10,000 tickets cost about $0.50 in API fees. That makes automated routing affordable for businesses far smaller than the ones who used to need it. The work that remains is human: define good questions, calibrate thresholds, watch the logs, and keep a human lane for the cases the machine should not decide alone.

Sources: OpenAI Decisions API developer guide and pricing, as reported by EconoTimes, WebProNews, Unite.AI, and Search Engine Watch on the 22 October 2026 general availability release. Accessed 25 October 2026.

Related reading: TopUseAI blog, Open Source LLM API Pricing, and n8n vs Zapier Make.