Costs, speed and limits

The price

Decide with a Model is billed on the tokens it reads — the content, the questions and the pictures — and nothing else: a decision model writes nothing, so there is no output, reasoning or cache to pay for.

	Credits
Per 1,000 tokens read	0.00625
Per 1,000 tokens read, above 272,000 tokens in one request	0.0125
The smallest charge of a call	0.001
A step that fails — before the call, on a provider error, or with On Uncertain fail	0

The charge of each step is what it read, rounded up to the thousandth of a credit; the history of the run records the tokens and the credits of every call. A refused question is answered by the same call as the others: it costs the tokens it adds, which are few.

What counts

Measured on the live endpoint:

• The content is read once, however many questions you ask: a 4,600-token text with one question read 4,656 tokens; with two, 4,808.
• Each question adds a fixed part and its words: about 150 tokens for a yes/no, about 100 for a choice or a score, plus a few for each option or level and the words of the instructions and descriptions.
• Each picture costs about 200 tokens at low detail, whatever its size; at high, about 1,000 for a 1200-pixel photo and about 2,600 for a 2400 × 3200 one. Very small pictures cost a few tokens: 128 thumbnails of 32 pixels read 414 tokens in all.
• The envelope around the content is about 40 tokens.

So asking several questions in one step is much cheaper than one step per question: the content and the pictures are read once.

What typical decisions cost

From the labs and examples of this section:

Decision	Tokens read	Credits
A customer message, three questions	≈ 500	0.003–0.004
A document, two questions	≈ 250–450	0.002–0.003
A product photo, five checks, low	≈ 960	0.006–0.007
The same photo, high	≈ 1,650	0.009
Four pictures and a brief, two questions	≈ 900–1,000	0.006–0.009
1,000 customer messages	≈ 500,000	≈ 3.2
1,000 product photos, low	≈ 960,000	≈ 6

Next to the alternatives: judging a photo in JSON with Image Analysis costs a few hundredths of a credit, and its answer must then be validated and read field by field; a language step costs at least 0.01 credits a call. Decide with a Model is the cheapest way to judge content with a model — and still more expensive than the free nodes, which should decide everything that can be measured or compared.

The estimate before a run

The estimate counts what can be known before the run: the questions exactly; the content when it is written on the step, otherwise 1,000 tokens assumed; and one picture at the step’s detail, because how many pictures arrive is only known at run time. The estimate says so (“content assumed (~1000 tokens); one image assumed”). It reserves a little more than a short decision costs — the difference is released at the end of the run — and less than a step that receives 20 pictures at high detail.

Speed

Request	Median	95th percentile
A message or a document	190–200 ms	210–250 ms
One photo, low	295 ms	530 ms
One photo, high	340 ms	600 ms
Four pictures	720 ms	2.7 s

That is the model’s answer; a step in a running workflow takes a few tenths of a second more. The first call after a quiet period takes about two seconds.

Limits

Limit	Value	When it is broken
Questions per step	1 to 16	refused at validation (invalid_question_count)
Options of a choice	2 to 50	refused at validation (invalid_option_count)
Levels of a score	2 to 10	refused at validation (invalid_level_count)
Name of a question	lowercase letters, digits and _, at most 40	refused at validation (invalid_port_name, reserved_port_name)
Instructions	up to 2,000 characters	refused at validation
Value of an option, label of a level	up to 100 characters; descriptions up to 500	refused at validation
Pictures	up to 20 per step, wired and from the image set together	above 20, Image Set Limit decides (TOO_MANY_IMAGES with fail)
Content	the model’s context, about a million tokens; above 272,000 the rate doubles

A decision model answers questions about one content at a time. To decide on many items, put the step inside an iteration: one call per item, in parallel.