What Interviews into Case Studies Cost in Model Tokens

Two line items sit here. The speech-to-text call is priced per minute of audio, so a one-hour interview is a single charge you can read straight off the provider's page. Cleanup is a text call: roughly 10,000 input tokens of raw transcript plus a short instruction, with about 9,000 tokens back for punctuation, speaker attribution and removal of filler words.

Cleanup is the cheapest place to save money and a poor place to use a frontier model. Punctuation restoration is a pattern task with a correct answer, so a small model handles it. The decision rule: if the task has one objectively correct output and a wrong output is caught by a person skimming the text, use the smallest model that passes your spot check.

Ask whether you need the pass at all. A strong drafting model tolerates messy punctuation, so you can skip cleanup and save one full-transcript round trip per interview. Keep it when a human editor reads the raw transcript, which is common when the customer must approve quotes before publication as the testimonial workflow recommends.

This stage earns the case study its credibility and multiplies the token count. The evidence base is consistent that the story needs specifics: the problem, the solution and the measured outcome, with striking statistics placed inside the narrative rather than drowning it per Elliott Pearce's B2B case study best practices. Case studies built on named numbers and personal quotes persuade more than vague praise according to Weber Associates.

Split extraction into two or three focused calls rather than one giant prompt. Focused prompts produce cleaner structured output, and each call is easier to re-run alone:

1. Entity and context call. Send the transcript and ask for company, industry, team size, tools replaced, timeline and named stakeholders as JSON.
2. Outcome call. Ask for every quantitative claim with the exact quote and timestamp, plus a flag for claims that need verification before publication.
3. Theme and objection call. Ask for the problem, the alternatives considered, what nearly stopped the purchase and the decisive proof. These map to the buying-decision questions that yield the most useful testimonials per the 100tasks question set.

Each call re-sends the full transcript. Three calls at 10,000 input tokens each is 30,000 input tokens before the draft exists. That is the biggest structural cost in a naive pipeline, and the fix is prompt caching: cached input is billed at a fraction of base input on current Anthropic rates, and the batch API takes half off both input and output per the Metacto pricing breakdown. The equivalent cuts exist on OpenAI tiers, where batch prices run at half of standard and input above a long-context threshold moves to a higher schedule per Silicon Data.

Run all three extraction calls on the small model. A wrong tag here yields a bad draft you catch in review, and a re-run costs one call, not a correction in front of a customer. Structured output modes make small models reliable at JSON extraction, which is exactly the shape these calls need.

Quote selection is cheap in tokens and expensive in risk. The call sends the transcript plus the extraction JSON and asks for 8 to 12 candidate quotes, each tagged with the theme it supports, the timestamp and a flag for anything that reads as a superlative claim. Output is small, perhaps 600 to 1,000 tokens.

The compliance pass is a narrower call over that shortlist: check for claims needing substantiation, comparative statements about competitors, and anything the customer might regret in print. It is not optional if you follow advertising rules. In the United States, the FTC's endorsements and testimonials guidance is the place to check what a published customer statement must be able to support per the 100tasks guide. A small or mid-tier model handles this provided the prompt states the rule set explicitly.

Keep the customer's voice. The interview guides consistently warn against leading the customer or polishing away credibility, because a testimonial works when it is recognizable as the customer's own words per the t2d3 guide. Instruct the model to make light edits only, and log any quote you rewrite so the approval email shows the final text.

The draft is the first call with substantial output, and it is where quality shows. A 1,200-word case study is roughly 1,600 output tokens. The input is the transcript (or a condensed evidence pack), the extraction JSON, the chosen quotes and your house style instructions, typically 12,000 to 15,000 input tokens.

Budget generously on output. Models that reason before answering bill that reasoning as output, so a draft call can cost several times its visible word count. Output tokens still dominate many production bills, and context length is a pricing event rather than just a longer prompt per Silicon Data. OpenAI publishes the per-million rates for its current tiers including cached input and output on its API pricing page.

Structure the prompt around problem, solution and outcome, the format the case study literature recommends for keeping the narrative tight per Elliott Pearce, and keep the customer as the hero while the vendor stays the guide as the Content Marketing Institute advises. Instructions are nearly free in tokens; what you pay for in the draft call is judgment, which is why this is the stage for the large model.

Generate two drafts at different temperatures and let a human pick. The second draft costs about what the first did and saves an editorial round trip, which is worth more than the tokens.

Review is the quiet budget killer because it is iterative. A typical flow is a fact-check pass (does every number in the draft match the transcript, with a quote and timestamp), a style pass (house voice, reading level, banned phrases) and a final approval-ready pass after the customer's comments arrive. Three review calls at 3,000 to 5,000 input tokens each, with 300 to 600 output tokens, is normal.

The fact-check pass belongs on the large model. It reads the whole draft against the whole evidence pack and is the last line of defense before a named customer sees a wrong number attached to their name. The style pass can run on a mid-tier model. The post-comment pass goes back to the large model, because that is the version that ships.

Here is a worked estimate for one 60-minute interview at standard list rates, using a cheap tier at roughly $0.10 per million input and $0.50 per million output against a large tier at roughly $2 per million input and $10 per million output, the shape of the current Claude ladder as compiled from Anthropic's pricing docs per the Metacto breakdown.

Two numbers deserve attention. First, the total sits near the top of the "a few cents to 20 cents" band because the transcript is re-sent six or seven times; a leaner pipeline that condenses the transcript into a 3,000-token evidence pack after extraction halves the draft and review inputs and lands nearer 8 to 10 cents. Second, audio transcription is billed per minute and often costs more than every text call except the draft, which is why the headline attributes the bill to drafting and review rather than to audio.

Repeat that over 200 interviews a year and inference comes to $16 to $33 at list rates, before caching or batch discounts. The labor around it, scheduling, approval emails and editorial judgment, is two orders of magnitude larger. Inference is not the constraint on a case study program; the interview calendar is.

If you are running these calls across several model families to get the tier mix right, putting them behind a single endpoint with one key and one prepaid balance removes most of the plumbing: Quanta Inference serves text, image, video, audio and embedding models through an OpenAI-compatible API priced in US dollars.

The routing rule is not about cost per token; it is about the cost of being wrong.

• Small model: any task with one correct answer, or where a wrong answer is caught by a cheap check. Cleanup, entity extraction, theme tagging, quote shortlisting, style linting. A bad tag costs one re-run at a fraction of a cent.
• Mid-tier model: tasks that need judgment over short text with clear criteria. Style rewrites, headline options, meta descriptions, social variants.
• Large model: tasks whose output a customer or a buyer reads word for word. The draft, the fact-check pass, and the final version after customer comments. Here a wrong sentence is seen by a named customer and possibly by their legal team.

A second rule governs volume: any call that re-sends the full transcript should be cached, and any call whose output nobody reads within the hour should run in batch at half price. Read the discounts off each provider's own page before committing a budget: OpenAI publishes standard, batch and long-context rates per million tokens on its API pricing documentation, and Anthropic's ladder, caching multipliers and batch discount are summarized in the Metacto breakdown here.

Reading the draft against a published buyer's guide is also worth the tokens. Buyers weigh relevance, credibility and context, and journalistically written stories with clear metrics and independent quotes outperform product-dense advertorials per Prime Creative. Put those three criteria in the review prompt as an explicit checklist.

If you would rather not build and tune this pipeline yourself, describe the workflow you need (transcript in, structured evidence out, draft and review passes) and a specialist will build it with you at Quanta.