ExplainedOpenAIDecisions APIclassificationconfidence thresholdsgpt-6-luna

OpenAI's Decisions API, explained: I ran it on 33 egg photos, and the top answer nearly binned a clean egg

OpenAI's Decisions API, in public beta since 6 October 2026, does not write a reply. You send a photo or text, a question and a fixed list of answers, and it returns a probability for each. I ran it on 33 egg photos: it answered in about 0.2 seconds, and one clean egg came back 47% dirty, 44% clean. Here is what that means for your code.

Free PDF: the AI Engineering cheat sheet, 12 pages Get it free

Roni Das

8 min read

Published9 October 2026, 08:00 UTC

Two courses by the author of this page:

AI Engineering

206 lessons · 10 free to read

₹999 in India · $49 elsewhere, once

Get AI Engineering

You own this course

System Design

770 lessons · 18 free to read

₹499 in India · $49 elsewhere, once

Get System Design

You own this course

Every lesson in both courses has an AI tutor beside it. It reads the same lesson you are reading and answers your questions from it. Try the tutor

Egg tray-06 from the run. The Decisions API gave it clean 0.44 and dirty 0.47.

The photo is clean, and the API's top answer is dirty. The two numbers are three points apart, so the API was close to a coin toss on this egg. What your code does with that is the whole lesson.

In short

The Decisions API returns a probability for each answer you allowed. It does not write a reply, and you pay only for input tokens.
On 33 egg photos it answered in a median 0.20 s. The same model writing a JSON answer took 1.26 s.
It was not cheaper per photo in my run: $0.0234 against $0.0197 per 1,000, because it counted the image as more tokens.
Taking the top answer rejected one clean egg. Two cut-offs (pass below 30%, reject above 70%) sent it to a person and got none of the 33 wrong.
I picked those cut-offs on the same 33 photos. Test yours on photos you did not tune on.

Released: 6 Oct 2026, public beta
Endpoint: POST /v1/decisions
Model: gpt-6-luna (only one)
Price: $0.10 per 1M input tokens, no output charge
My run: 33 photos, 102 calls, 9 Oct 2026

This page is the free part.

The course goes deeper on routing and thresholds for production AI

₹999 in India$49 everywhere elseonce, for the whole course

The AI Engineering course covers routing and thresholds for production AI across a run of lessons, not one page. These 3 alone are about 203 minutes of step-by-step reading, every one with code you run in the browser, all with a quiz.

Model Routing and Cascades: Running Three Models Without Chaos
Cheap models for easy traffic, strong models for hard cases. How to route, when a cascade stops paying, and the one thing your gateway must never do.
85 min · diagrams · code you run · quiz
Was the Cheap Model Sure? Choosing Which Questions a Cascade Passes On
A cascade lets a cheap model answer first and passes some questions to a strong model. Everything depends on the clue that picks which ones. On 175 maths problems with known answers, the cheap model said it was 100 out of 100 sure on almost every answer, so asking how sure did exactly as well as random. Asking it three times beat random clearly, but mostly by spotting hard questions, a free rule on the question did about as well, and in tokens it cost as much as the strong model alone.
38 min · diagrams · code you run · quiz
Token Cost Engineering: Where the LLM Bill Actually Goes
Most of an LLM bill is tokens you resend, not tokens users type. Decompose the spend, then cut it with caching, slimming, output caps, and batching.
80 min · diagrams · code you run · quiz

all part of

AI Engineering

206 lessons · about 213 hours · 10 free to read

₹999in India, by UPI

$49everywhere else, by PayPal

Get the AI Engineering course Read 10 lessons free first

You own this course

Continue with Model Routing and Cascades: Running Three Models Without Chaos

pay once,
yours for life

What the Decisions API is, in plain words

“The Decisions API evaluates text, images, or both and returns typed answers about 10x faster than the Responses API.”

OpenAI, Decisions guide

Most AI APIs answer by writing. You ask a chat model whether a photo shows a clean egg, it writes a word or a sentence, and your code has to read that text and hope it is in the shape you asked for.

The Decisions API answers like a multiple choice exam instead of an essay. You give it the evidence (text, images or both), a question, and the answers you allow. It returns a probability for each allowed answer, a number from 0 to 1. Your code never parses a sentence.

There are three kinds of question. A predicate asks whether something is true and returns one probability. A choice picks one of your options and returns a probability for every option plus a separate confidence. A score rates against ordered levels and returns a weighted average of the level numbers.

OpenAI says it is about 10 times faster than the same model through the Responses API. It runs only on gpt-6-luna for now, and it is a public beta.

The hands-on: 33 egg photos, from VS Code

I cut 33 public-domain egg photos from Wikimedia Commons into one egg each: 18 clean, 12 dirty duck eggs and 3 broken shells. I labelled each one myself by eye. Each photo went to /v1/decisions with three questions in one call: a choice (clean, dirty, cracked, unclear), a predicate (is there a visible crack) and a score (how dirty, on three levels).

Every response was saved raw. Every number on this page is read back from that file, not typed in.

What we measured: Median 204 ms per call with three questions, 523 input tokens, 0 output tokens, on 33 photos.

# egg_lab.py (shortened): one photo, one choice question, sent to POST /v1/decisions
body = {
    "model": "gpt-6-luna",
    "input": [{"role": "user", "content": [
        {"type": "input_text", "text": "One egg from a grading line."},
        {"type": "input_image", "image_url": data_url(photo)},   # base64 data URL
    ]}],
    "questions": [{
        "type": "choice",
        "name": "condition",
        "instructions": "What is the condition of the egg shell in this photo?",
        "choices": [
            {"value": "clean",   "description": "Shell is intact with no dirt, stains or droppings."},
            {"value": "dirty",   "description": "Shell is intact but has dirt, stains or droppings."},
            {"value": "cracked", "description": "Shell has a crack, hole, or is broken open."},
            {"value": "unclear", "description": "The photo does not show the shell well enough."},
        ],
    }],
}
# the run also sent a predicate ("is there a visible crack?") and a score (dirt: none, light, heavy)

# one real answer from results/rec2/calls.jsonl, photo duck-01
{"type": "choice", "name": "condition", "choice": "dirty", "probabilities": [{"value": "clean", "probability": 0.11}, {"value": "dirty", "probability": 0.84}, {"value": "cracked", "probability": 0.02}, {"value": "unclear", "probability": 0.03}], "confidence": 0.79}

The request the lab sent, and one real answer from the saved responses.

VS Code terminal showing one row per egg photo with four probabilities, the crack probability, the dirt score, latency in milliseconds and tokens — The classify run in VS Code. Row tray-06: clean 0.44, dirty 0.47, cracked 0.02, unclear 0.07, 197 ms.

Why the top answer is the wrong thing to act on

Egg tray-06 is clean. The API gave it clean 0.44 and dirty 0.47. If your code takes the top answer, that egg is rejected. A model that writes one word would have written dirty and never shown you how close it was.

The fix is two cut-offs on one number: the chance the egg is not clean, which is 1 minus P(clean). Below the low cut-off, pass. Above the high one, reject. In between, a person looks. Tray-06 scores 0.56, so it goes to a person.

With cut-offs at 0.30 and 0.70, none of my 33 photos went the wrong way, and 5 went to a person. With a single cut at 0.50, one good egg was rejected.

The honest part: I chose 0.30 and 0.70 by sweeping the same 33 photos, so 0 wrong is an in-sample result. With 33 photos it says little about new eggs. Pick your cut-offs on labelled examples, then check them on photos you did not use to pick them. If binning a good egg costs more than letting a dirty one through, raise the reject line.

$ python3 egg_lab.py sweep --run rec2
33 labelled photos, 15 not clean. score = 1 - P(clean)
pass below reject above passed bad rejected good to a person
      0.50         0.50          0             1           0
      0.20         0.80          0             0          12
      0.10         0.90          0             0          15
      0.05         0.95          0             0          16
      0.30         0.70          0             0           5

sent to a person with cut-offs 0.30 / 0.70: tray-06 0.56, box10-01 0.35, box10-08 0.31, box10-09 0.45, box10-10 0.48

The sweep, run offline over the saved responses and my labels.

An axis from 0% to 100% labelled chance it isn't clean, with cut-off lines at 30% and 70%, three bands pass, a person checks and reject, the clean egg tray-06 at 0.56 in the middle band, and 32 other dots — The 33 real scores on one axis. Green dots are photos I labelled clean, red are dirty or broken.

Against the same model writing its answer

On 15 of the photos, one question each, I called the Decisions API and then asked gpt-6-luna through the Responses API to write the answer as JSON, at reasoning effort none and at medium. The calls ran back to back from the same laptop, so the times include the network.

	Decisions API	Responses, effort none	Responses, effort medium
median time per photo	201 ms	1262 ms	1878 ms
median input tokens	226	127	127
median output tokens	0	12	47
cost per 1,000 photos	$0.0234	$0.0197	$0.0488
matched my label	15 of 15	14 of 15	13 of 15
what came back	a probability per answer	one word	one word

Costs are tokens times OpenAI's list prices, not a bill. 15 photos is far too few to compare accuracy; the useful difference is the probabilities.

$ python3 egg_lab.py summary --run rec2
102 saved calls in rec2/calls.jsonl
classify           calls  33  median    204 ms  (min 188, max 256)  median input tokens 523  median output tokens 0
repeat             calls  24  median    205 ms  (min 194, max 255)  median input tokens 523  median output tokens 0
compare-decisions  calls  15  median    201 ms  (min 197, max 264)  median input tokens 226  median output tokens 0
compare-none       calls  15  median   1262 ms  (min 1160, max 1736)  median input tokens 127  median output tokens 12
compare-medium     calls  15  median   1878 ms  (min 1425, max 3458)  median input tokens 127  median output tokens 47
repeat             8 photos x 3 tries, largest gap in any probability 0.000
compare-decisions  matched my label 15/15  $0.0234 per 1,000 photos
compare-none       matched my label 14/15  $0.0197 per 1,000 photos
compare-medium     matched my label 13/15  $0.0488 per 1,000 photos

Read back from the saved responses.

VS Code terminal with the compare table for 15 photos and three summary lines for decisions, written effort none and written effort medium — The compare run in VS Code.

after “Against the same model writing its answer”

Curious what sits underneath routing and thresholds for production AI? This lesson takes it apart:

When the Data Answers Back: A Feedback Loop, Simulated on Real Demand: Some models decide what the world can do, and then learn from what it did. I simulated that on two years of real bike-rental counts: a model forecasts demand, bikes are placed to match, riders can only rent the bikes that are there, and the model is retrained on what was rented. The real counts are the demand; the rule for placing bikes is invented. With no margin the model missed true demand by 90.7 bikes an hour over 2012, against 43.6 for a model that learned from every rider, and turned away 769,320 riders. Scored on its own records it missed by 2.6. A 20% margin brought the error to 60.0 and a 50% margin to 46.4, for 820,304 idle bike-hours. What a feedback loop is, why the records hide it, what a margin gives and costs, and what to log.

68 min · diagrams · code you run · quiz

Get the AI Engineering course ₹999 in India · $49 elsewhere · pay onceor open the lesson

Continue in your course

What it costs

Decisions bills input tokens only. Plain gpt-6-luna has the same input price and also charges $0.50 per million output tokens, so the saving is the output and the wait, not cheaper reading.

Plans and limits

Decisions input, gpt-6-luna: $0.10 / 1M tokens
Decisions output: not billed
gpt-6-luna output, normal API: $0.50 / 1M tokens
My run, per photo, 3 questions: 523 tokens, about $0.00005
My whole lab, 102 calls: under 1 US cent

How it compares with Jev

TypeSafe released Jev on 15 September 2026 with the same idea: ask typed questions, get probabilities back, pay for input only. Its question types are Noul (yes or no), Choice and Score, and it costs $0.042 per million input tokens, 2.4 times less per token than Decisions.

The big difference is the input. Jev takes text only, so it cannot look at an egg photo. Decisions takes up to 128 images per request. The two models also count tokens differently, so a lower price per token does not mean a lower price per request.

When to use it, and when not to

Use it. when the answer comes from a fixed list: route a ticket, flag a photo, score a severity, check a policy.
Use it. when you have no training set yet. The egg demo's author says the same: start here, collect the hard cases, train a small model later.
Not for. anything you need written: an explanation, an extracted field, a reply. That is Structured Outputs or function calling.
Not at huge volume. if you already have labels. A small trained vision model is faster and has no per-call bill.
Always. set cut-offs from labelled examples and send the uncertain middle to a person. Asking the same photo again gave identical odds, so retries do not break a tie.

The beta has one model, and the docs disagree on images: the guide says base64 only, the API reference also allows public URLs. I used base64.

Get the free AI Engineering Cheat Sheet

What our labs measured about tokens, prompting, RAG, evals, agents, serving, cost and fine-tuning, and the rule each result teaches.

the question most readers have next

What would it take to put routing and thresholds for production AI into production?

A demo works once. Production has to work every time, under load and on bad input. The lessons below cover what that takes, starting with "Model Routing and Cascades: Running Three Models Without Chaos".

The lessons that teach the next part

Cheap models for easy traffic, strong models for hard cases. How to route, when a cascade stops paying, and the one thing your gateway must never do.
Model Routing and Cascades: Running Three Models Without ChaosAI Engineering course85 min
A cascade lets a cheap model answer first and passes some questions to a strong model. Everything depends on the clue that picks which ones. On 175 maths problems with known answers, the cheap model said it was 100 out of 100 sure on almost every answer, so asking how sure did exactly as well as random. Asking it three times beat random clearly, but mostly by spotting hard questions, a free rule on the question did about as well, and in tokens it cost as much as the strong model alone.
Was the Cheap Model Sure? Choosing Which Questions a Cascade Passes OnAI Engineering course38 min
Most of an LLM bill is tokens you resend, not tokens users type. Decompose the spend, then cut it with caching, slimming, output caps, and batching.
Token Cost Engineering: Where the LLM Bill Actually GoesAI Engineering course80 min

Stuck on a step? Every lesson has an AI tutor beside it. Ask it your question and it answers from the lesson you are reading. Try the tutor

Get the free AI Engineering cheat sheet (PDF)

Where this is taught

AI Engineering

206 lessons · about 213 hours · 10 free to read

₹999 in India by UPI$49 elsewhere by PayPal, once, for life

Get the AI Engineering course

You own this course

Also by the same author

System Design Masterclass

770 lessons · about 283 hours · 18 free to read

₹499 in India by UPI$49 elsewhere by PayPal, once, for life

Get the System Design course

You own this course

Questions

What is the OpenAI Decisions API?

An endpoint, POST /v1/decisions, that takes text or images plus questions with a fixed set of answers, and returns a probability for each allowed answer instead of writing a reply. It runs on gpt-6-luna and has been in public beta since 6 October 2026.

How much does it cost?

$0.10 per million input tokens with gpt-6-luna, and no charge for output tokens. In my run one egg photo with three questions was 523 input tokens, about $0.00005.

Is it faster than asking a chat model?

In my run, yes: a median 0.20 s per photo against 1.26 s for the same model writing a one-word JSON answer at reasoning effort none, on 15 photos. OpenAI's own figure is about 10 times faster.

Is it cheaper than the Responses API?

Not in my run. Per 1,000 photos it cost $0.0234 against $0.0197 for the Responses API at effort none, because the Decisions call counted the image as more input tokens.

How do I choose a threshold?

Use two cut-offs on one score, for example the chance an item is not clean. Pick them on labelled examples by what each kind of mistake costs, send the middle band to a person, and check the cut-offs on examples you did not tune on.

How is it different from Jev?

Same idea, typed questions and probabilities. Jev is text only and costs $0.042 per million input tokens. Decisions takes images and costs $0.10 per million input tokens.

Sources

What this explainer is based on, so you can check it.

Decisions guide OpenAI
Create a decision, API reference OpenAI
gpt-6-luna model page OpenAI
API changelog, 6 October 2026 OpenAI
DevDay 2026 recap OpenAI
Egg demo with Roboflow RF-DETR Erik Kokalj
Jev models TypeSafe

Learn the parts behind it

Model Routing and Cascades: Running Three Models Without ChaosCheap models for easy traffic, strong models for hard cases. How to route, when a cascade stops paying, and the one thing your gateway must never do.85 min lesson Was the Cheap Model Sure? Choosing Which Questions a Cascade Passes OnA cascade lets a cheap model answer first and passes some questions to a strong model. Everything depends on the clue that picks which ones. On 175 maths problems with known answers, the cheap model said it was 100 out of 100 sure on almost every answer, so asking how sure did exactly as well as random. Asking it three times beat random clearly, but mostly by spotting hard questions, a free rule on the question did about as well, and in tokens it cost as much as the strong model alone.38 min lesson Token Cost Engineering: Where the LLM Bill Actually GoesMost of an LLM bill is tokens you resend, not tokens users type. Decompose the spend, then cut it with caching, slimming, output caps, and batching.80 min lesson

Go deeper than the news

This explainer sits on top of our AI Engineering course: 206 lessons on RAG, evals, agents, serving, security and MLOps, many built around a real experiment. 10 lessons are free to read, with no card needed.

course 2

AI Engineering

Take models from notebook to production, with labs on real models.

206 interactive lessons
Serving, pipelines, feature stores and monitoring
RAG, fine-tuning, evals and agents
Step-by-step diagrams and quizzes
Lifetime access and all future lessons
A separate purchase from System Design

Read 10 lessons free first

In India

₹999

Pay with UPI

Everywhere else

$49

Pay with PayPal

once · yours for life

ExplainedOpenAIDecisions APIclassificationconfidence thresholdsgpt-6-luna

OpenAI's Decisions API, explained: I ran it on 33 egg photos, and the top answer nearly binned a clean egg

Free PDF: the AI Engineering cheat sheet, 12 pages Get it free

Roni Das

8 min read

Published9 October 2026, 08:00 UTC

Two courses by the author of this page:

AI Engineering

206 lessons · 10 free to read

₹999 in India · $49 elsewhere, once

Get AI Engineering

You own this course

System Design

770 lessons · 18 free to read

₹499 in India · $49 elsewhere, once

Get System Design

You own this course

Every lesson in both courses has an AI tutor beside it. It reads the same lesson you are reading and answers your questions from it. Try the tutor

Egg tray-06 from the run. The Decisions API gave it clean 0.44 and dirty 0.47.

The photo is clean, and the API's top answer is dirty. The two numbers are three points apart, so the API was close to a coin toss on this egg. What your code does with that is the whole lesson.

In short

The Decisions API returns a probability for each answer you allowed. It does not write a reply, and you pay only for input tokens.
On 33 egg photos it answered in a median 0.20 s. The same model writing a JSON answer took 1.26 s.
It was not cheaper per photo in my run: $0.0234 against $0.0197 per 1,000, because it counted the image as more tokens.
Taking the top answer rejected one clean egg. Two cut-offs (pass below 30%, reject above 70%) sent it to a person and got none of the 33 wrong.
I picked those cut-offs on the same 33 photos. Test yours on photos you did not tune on.

Released: 6 Oct 2026, public beta
Endpoint: POST /v1/decisions
Model: gpt-6-luna (only one)
Price: $0.10 per 1M input tokens, no output charge
My run: 33 photos, 102 calls, 9 Oct 2026

This page is the free part.

The course goes deeper on routing and thresholds for production AI

₹999 in India$49 everywhere elseonce, for the whole course

Model Routing and Cascades: Running Three Models Without Chaos
Cheap models for easy traffic, strong models for hard cases. How to route, when a cascade stops paying, and the one thing your gateway must never do.
85 min · diagrams · code you run · quiz
Was the Cheap Model Sure? Choosing Which Questions a Cascade Passes On
A cascade lets a cheap model answer first and passes some questions to a strong model. Everything depends on the clue that picks which ones. On 175 maths problems with known answers, the cheap model said it was 100 out of 100 sure on almost every answer, so asking how sure did exactly as well as random. Asking it three times beat random clearly, but mostly by spotting hard questions, a free rule on the question did about as well, and in tokens it cost as much as the strong model alone.
38 min · diagrams · code you run · quiz
Token Cost Engineering: Where the LLM Bill Actually Goes
Most of an LLM bill is tokens you resend, not tokens users type. Decompose the spend, then cut it with caching, slimming, output caps, and batching.
80 min · diagrams · code you run · quiz

all part of

AI Engineering

206 lessons · about 213 hours · 10 free to read

₹999in India, by UPI

$49everywhere else, by PayPal

Get the AI Engineering course Read 10 lessons free first

You own this course

Continue with Model Routing and Cascades: Running Three Models Without Chaos

pay once,
yours for life

What the Decisions API is, in plain words

“The Decisions API evaluates text, images, or both and returns typed answers about 10x faster than the Responses API.”

OpenAI, Decisions guide

Most AI APIs answer by writing. You ask a chat model whether a photo shows a clean egg, it writes a word or a sentence, and your code has to read that text and hope it is in the shape you asked for.

OpenAI says it is about 10 times faster than the same model through the Responses API. It runs only on gpt-6-luna for now, and it is a public beta.

The hands-on: 33 egg photos, from VS Code

Every response was saved raw. Every number on this page is read back from that file, not typed in.

What we measured: Median 204 ms per call with three questions, 523 input tokens, 0 output tokens, on 33 photos.

# egg_lab.py (shortened): one photo, one choice question, sent to POST /v1/decisions
body = {
    "model": "gpt-6-luna",
    "input": [{"role": "user", "content": [
        {"type": "input_text", "text": "One egg from a grading line."},
        {"type": "input_image", "image_url": data_url(photo)},   # base64 data URL
    ]}],
    "questions": [{
        "type": "choice",
        "name": "condition",
        "instructions": "What is the condition of the egg shell in this photo?",
        "choices": [
            {"value": "clean",   "description": "Shell is intact with no dirt, stains or droppings."},
            {"value": "dirty",   "description": "Shell is intact but has dirt, stains or droppings."},
            {"value": "cracked", "description": "Shell has a crack, hole, or is broken open."},
            {"value": "unclear", "description": "The photo does not show the shell well enough."},
        ],
    }],
}
# the run also sent a predicate ("is there a visible crack?") and a score (dirt: none, light, heavy)

# one real answer from results/rec2/calls.jsonl, photo duck-01
{"type": "choice", "name": "condition", "choice": "dirty", "probabilities": [{"value": "clean", "probability": 0.11}, {"value": "dirty", "probability": 0.84}, {"value": "cracked", "probability": 0.02}, {"value": "unclear", "probability": 0.03}], "confidence": 0.79}

The request the lab sent, and one real answer from the saved responses.

Why the top answer is the wrong thing to act on

With cut-offs at 0.30 and 0.70, none of my 33 photos went the wrong way, and 5 went to a person. With a single cut at 0.50, one good egg was rejected.

$ python3 egg_lab.py sweep --run rec2
33 labelled photos, 15 not clean. score = 1 - P(clean)
pass below reject above passed bad rejected good to a person
      0.50         0.50          0             1           0
      0.20         0.80          0             0          12
      0.10         0.90          0             0          15
      0.05         0.95          0             0          16
      0.30         0.70          0             0           5

sent to a person with cut-offs 0.30 / 0.70: tray-06 0.56, box10-01 0.35, box10-08 0.31, box10-09 0.45, box10-10 0.48

The sweep, run offline over the saved responses and my labels.

Against the same model writing its answer

	Decisions API	Responses, effort none	Responses, effort medium
median time per photo	201 ms	1262 ms	1878 ms
median input tokens	226	127	127
median output tokens	0	12	47
cost per 1,000 photos	$0.0234	$0.0197	$0.0488
matched my label	15 of 15	14 of 15	13 of 15
what came back	a probability per answer	one word	one word

Costs are tokens times OpenAI's list prices, not a bill. 15 photos is far too few to compare accuracy; the useful difference is the probabilities.

$ python3 egg_lab.py summary --run rec2
102 saved calls in rec2/calls.jsonl
classify           calls  33  median    204 ms  (min 188, max 256)  median input tokens 523  median output tokens 0
repeat             calls  24  median    205 ms  (min 194, max 255)  median input tokens 523  median output tokens 0
compare-decisions  calls  15  median    201 ms  (min 197, max 264)  median input tokens 226  median output tokens 0
compare-none       calls  15  median   1262 ms  (min 1160, max 1736)  median input tokens 127  median output tokens 12
compare-medium     calls  15  median   1878 ms  (min 1425, max 3458)  median input tokens 127  median output tokens 47
repeat             8 photos x 3 tries, largest gap in any probability 0.000
compare-decisions  matched my label 15/15  $0.0234 per 1,000 photos
compare-none       matched my label 14/15  $0.0197 per 1,000 photos
compare-medium     matched my label 13/15  $0.0488 per 1,000 photos

Read back from the saved responses.

after “Against the same model writing its answer”

Curious what sits underneath routing and thresholds for production AI? This lesson takes it apart:

68 min · diagrams · code you run · quiz

Get the AI Engineering course ₹999 in India · $49 elsewhere · pay onceor open the lesson

Continue in your course

What it costs

Decisions bills input tokens only. Plain gpt-6-luna has the same input price and also charges $0.50 per million output tokens, so the saving is the output and the wait, not cheaper reading.

Plans and limits

Decisions input, gpt-6-luna: $0.10 / 1M tokens
Decisions output: not billed
gpt-6-luna output, normal API: $0.50 / 1M tokens
My run, per photo, 3 questions: 523 tokens, about $0.00005
My whole lab, 102 calls: under 1 US cent

How it compares with Jev

When to use it, and when not to

Use it. when the answer comes from a fixed list: route a ticket, flag a photo, score a severity, check a policy.
Use it. when you have no training set yet. The egg demo's author says the same: start here, collect the hard cases, train a small model later.
Not for. anything you need written: an explanation, an extracted field, a reply. That is Structured Outputs or function calling.
Not at huge volume. if you already have labels. A small trained vision model is faster and has no per-call bill.
Always. set cut-offs from labelled examples and send the uncertain middle to a person. Asking the same photo again gave identical odds, so retries do not break a tie.

The beta has one model, and the docs disagree on images: the guide says base64 only, the API reference also allows public URLs. I used base64.

Get the free AI Engineering Cheat Sheet

What our labs measured about tokens, prompting, RAG, evals, agents, serving, cost and fine-tuning, and the rule each result teaches.

the question most readers have next

What would it take to put routing and thresholds for production AI into production?

The lessons that teach the next part

Cheap models for easy traffic, strong models for hard cases. How to route, when a cascade stops paying, and the one thing your gateway must never do.
Model Routing and Cascades: Running Three Models Without ChaosAI Engineering course85 min
A cascade lets a cheap model answer first and passes some questions to a strong model. Everything depends on the clue that picks which ones. On 175 maths problems with known answers, the cheap model said it was 100 out of 100 sure on almost every answer, so asking how sure did exactly as well as random. Asking it three times beat random clearly, but mostly by spotting hard questions, a free rule on the question did about as well, and in tokens it cost as much as the strong model alone.
Was the Cheap Model Sure? Choosing Which Questions a Cascade Passes OnAI Engineering course38 min
Most of an LLM bill is tokens you resend, not tokens users type. Decompose the spend, then cut it with caching, slimming, output caps, and batching.
Token Cost Engineering: Where the LLM Bill Actually GoesAI Engineering course80 min

Stuck on a step? Every lesson has an AI tutor beside it. Ask it your question and it answers from the lesson you are reading. Try the tutor

Get the free AI Engineering cheat sheet (PDF)

Where this is taught

AI Engineering

206 lessons · about 213 hours · 10 free to read

₹999 in India by UPI$49 elsewhere by PayPal, once, for life

Get the AI Engineering course

You own this course

Also by the same author

System Design Masterclass

770 lessons · about 283 hours · 18 free to read

₹499 in India by UPI$49 elsewhere by PayPal, once, for life

Get the System Design course

You own this course

Questions

What is the OpenAI Decisions API?

How much does it cost?

$0.10 per million input tokens with gpt-6-luna, and no charge for output tokens. In my run one egg photo with three questions was 523 input tokens, about $0.00005.

Is it faster than asking a chat model?

In my run, yes: a median 0.20 s per photo against 1.26 s for the same model writing a one-word JSON answer at reasoning effort none, on 15 photos. OpenAI's own figure is about 10 times faster.

Is it cheaper than the Responses API?

Not in my run. Per 1,000 photos it cost $0.0234 against $0.0197 for the Responses API at effort none, because the Decisions call counted the image as more input tokens.

How do I choose a threshold?

How is it different from Jev?

Same idea, typed questions and probabilities. Jev is text only and costs $0.042 per million input tokens. Decisions takes images and costs $0.10 per million input tokens.

Sources

What this explainer is based on, so you can check it.

Decisions guide OpenAI
Create a decision, API reference OpenAI
gpt-6-luna model page OpenAI
API changelog, 6 October 2026 OpenAI
DevDay 2026 recap OpenAI
Egg demo with Roboflow RF-DETR Erik Kokalj
Jev models TypeSafe

Learn the parts behind it

Go deeper than the news

course 2

AI Engineering

Take models from notebook to production, with labs on real models.

206 interactive lessons
Serving, pipelines, feature stores and monitoring
RAG, fine-tuning, evals and agents
Step-by-step diagrams and quizzes
Lifetime access and all future lessons
A separate purchase from System Design

Read 10 lessons free first

In India

₹999

Pay with UPI

Everywhere else

$49

Pay with PayPal

once · yours for life