Scoring without output tokens: the cost of classifying millions of documents

When a language model classifies by generating an answer, you pay for the text it writes, wait while it writes it, and then parse it. For a label, a yes/no or a score, the generated text is overhead. spinf classifies by scoring instead: it returns the probability of each answer you list and generates nothing. There are no output tokens, and only the input is billed.

What generation costs in a classification job

A generated answer adds three costs that grow with every document:

• Output tokens. Even a short reply ("Category: billing. Confidence: high.") is several tokens, and explanations run to hundreds. Output tokens are billed on every item.
• Parsing. Free text has to be mapped back to your labels, with rules for replies that do not match any of them.
• Waiting. Text is produced one token at a time after the input has been read.

A scored answer has none of these. The response lists your answers with their probabilities; the choices of a chat completion are empty because nothing is written.

What is billed

Only processed input tokens, at $0.09 per million on spinf-12b (our optimized Gemma 4 12B):

• the content, once per call, however many questions you ask about it;
• each question's own tokens: the filled template and its answers;
• one empty floor per question per call: the same question scored with no content, which gives the question's bias (Calibration). You can skip it with "empty_floor": false once you have it.

The scores themselves are free. There is a minimum of 1,000 tokens per call, counted on the whole call, so batch work sends many documents per call with inputs: the minimum and the floors are then shared. The details are in Limits & pricing.

The cost of a million documents

Take documents of about 300 tokens, each asked 10 questions of about 20 tokens:

content	300 tokens
+ 10 questions × 20 tokens	200 tokens
= tokens per document	about 500
× 1,000,000 documents	500M tokens
× $0.09 per million	about $45

With many documents per call, the floors add almost nothing per document. The same arithmetic on the use-case pages gives about $117 for a million 800-token news articles with 20 questions each, and about $32 for a million 120-token survey answers with 8 questions.

Many questions cost little

Because the content is read once per call, the cost per question is only its own tokens. Asking 10 questions of a document costs far less than 10 separate classifications of it: the 300 tokens of content are paid once, not ten times. This changes how jobs are designed:

• ask the questions you would otherwise skip, such as a policy check next to the topic and sentiment;
• ask the same question for every company, product or aspect with a placeholder, up to 1,000 values per question;
• re-score an archive when the codebook or the questions change, instead of living with the first version.

Labelled examples are input too

Few-shot examples placed in the prompt raise accuracy the most (zero-shot and few-shot classification), and they are part of the content, so they are billed with every document. In our ticket benchmark, 8 short labelled examples added about 1,100 tokens per ticket: about $0.12 per 1,000 tickets on spinf-12b. For content moderation, the use-case figures are about $0.01 per 1,000 posts with no examples and about $0.08 with 8 examples.

Batch LLM processing at scale

The on-demand API is built for throughput: backfills, archives and periodic batches, self-serve. The Thinking Text news indexes were built this way on 15 million articles and about a trillion tokens across research iterations. For projects over a trillion tokens we also offer dedicated deployments and file-based batch; for decisions in live flows, a reserved endpoint.

New accounts start with 50M free tokens: enough to score about 100,000 documents of the size above, or to test several question wordings on a few hundred of your own items many times over.

Try it

The classification API page runs the cost example and a live call; the news analysis and survey & review analysis pages show it per use case. Related: Limits & pricing and probabilities from a language model.