LLM API Cost Comparison: What an AI Board Deck Really Costs

Price-per-token tables are easy to find and hard to use. We ran the same board deck on seven models and read the bill: what each step cost, and how far billed cost sat from list price.

By Ed · Published November 5, 2026 · 6 min read

Twenty-five years starting companies, consulting, owning a multi-million-dollar P&L and leading business intelligence for large enterprises.

The same board deck cost 15 cents on GPT-6 Luna and $11.10 on Claude Opus 5.5. On every model, 65% to 86% of the bill went to answering the data questions behind the slides, not to writing them. And list prices are a poor guide: six of seven models were billed 12% to 54% under list.

How this was tested. One prompt, “Northvale 2026 year-to-date board review, with an executive summary and recommendations”, run once per model in the Quiriz PowerPoint add-in on November 5, 2026. Northvale is a fictional US apparel company with stories planted in its data: Google and Meta ad exports and a HubSpot wholesale pipeline, through September 15, 2026. Every step ran on the model under test: plan, build (answering each slide’s question from the data), gap analysis, write and critic; see how a deck is built. Three small parts stayed fixed for every run (chart styling, question routing and a grounding check). Cost is what OpenRouter billed. Build time runs from pressing Build to the last slide written, and leaves out the plan, which waits for a person to approve it. One run per model, so treat close results as a tie. Full method on the benchmark page.

Cost by step

The steps are the five described in how a Quiriz deck is built: plan, build, gap analysis, write and critic.

Model	Plan	Build (answers)	Gap analysis	Write (with slide notes)	Critic	Fixed parts	Total	Build share
GPT-6 Luna	$0.006	$0.106	$0.005	$0.011	$0.004	$0.020	$0.15	70%
DeepSeek V4.1 Flash	$0.050	$0.196	$0.006	$0.018	$0.007	$0.023	$0.30	65%
Gemini 3.8 Flash	$0.059	$1.18	$0.063	$0.113	$0.042	$0.016	$1.47	80%
GPT-6.1 Sol	$0.076	$1.66	$0.072	$0.216	$0.043	$0.010	$2.08	80%
Gemini 4 Argon	$0.081	$2.37	$0.094	$0.334	$0.062	$0.033	$2.98	80%
Kimi K3.1 Pro	$0.654	$2.88	$0.104	$0.307	$0.042	$0.024	$4.01	72%
Claude Opus 5.5	$0.404	$9.54	$0.233	$0.895	—	$0.027	$11.10	86%

“Fixed parts” are the steps that stayed on the same small models for every run: chart styling, question routing and a grounding check. Together they cost 1 to 3 cents a deck. The critic runs only when writing finishes within two minutes; the Opus writer took 124 seconds, so its deck had none.

Why the answers are the bill

A board deck is 10 to 17 content slides, and each slide is a question answered from data: a total, a breakdown, a comparison with last year. Each answer is a short loop. The model reads the table schemas and the company’s metric definitions, writes a query, reads the result, and sometimes queries again. Those schemas and definitions are long and get sent on every round, so an answer step reads about 33,000 to 56,000 tokens and writes only a few hundred. Across a deck that is 1.4 to 2.7 million input tokens. The writer, by contrast, reads a summary of the finished answers once.

The practical consequence: if you want a cheaper deck, change the model that answers, not the one that writes. Putting a premium writer on top of cheap answers costs cents more; putting a premium model on the answers costs dollars.

Billed cost against list price

Model	List price per million tokens (in / out)	Answers at list price	Answers billed	Difference
GPT-6 Luna	$0.10 / $0.50	$0.18	$0.11	−41%
DeepSeek V4.1 Flash	$0.15 / $0.60	$0.43	$0.20	−54%
Gemini 3.8 Flash	$0.83 / $4.12	$2.43	$1.18	−52%
GPT-6.1 Sol	$2 / $10	$2.82	$1.66	−41%
Gemini 4 Argon	$2 / $10	$4.17	$2.37	−43%
Kimi K3.1 Pro	$2 / $10	$3.27	$2.88	−12%
Claude Opus 5.5	$4 / $20	$9.54	$9.54	0%

List prices are OpenRouter’s as of November 4, 2026. The gap is prompt caching: the same schemas and definitions go out on every round, and most providers bill a repeated prefix at a discount without being asked. Anthropic’s models cache only when the request marks the part to cache. Quiriz now marks it, so the Opus figure here is higher than an Opus deck costs from now on. If you compare models on price, compare billed cost on your own workload, not the rate card.

Cheaper per token is not cheaper per deck

Gemini 3.8 Flash charges about 60% less than GPT-6.1 Sol per input token. Its deck cost $1.47 against Sol’s $2.08, only 29% less, because Flash made 59 analysis calls to Sol’s 33 and read 2.7 million tokens to Sol’s 1.4 million. A model that takes more steps to answer can spend its discount. It was also the slowest build, at 13.8 minutes.

In perspective

Even the most expensive deck here cost less than ten minutes of an analyst’s time, for a first draft of a board review. The real difference between models is not the bill; it is how much each deck finds. On that measure the $11 deck found the most, and the 30-cent deck came close on the ad story. See what each deck found.

Download the data

Every run as a CSV: cost by step, tokens, analysis calls, build time and list prices.

See what your own deck costs

Build a deck from your own data in PowerPoint and see the source of every figure. Free to start.

Try Quiriz free →

Frequently asked questions

How much does it cost to make a board deck with AI?

In our test, from 15 cents (GPT-6 Luna) to $11.10 (Claude Opus 5.5) in model usage for a 13- to 22-slide deck built from ad and CRM data. Most models landed between $1.50 and $4. That is the API bill only, not a subscription.

Which part of an AI deck costs the most?

The Build step, answering the questions behind each slide: 65% to 86% of the bill on every model. Each answer re-reads the table schemas and the company’s metric definitions, so input tokens dominate. Writing the slides and their notes cost between 1 and 90 cents, and the critic under 7 cents.

Why is the billed cost lower than tokens times the list price?

Prompt caching. When the same long prefix is sent again, most providers bill the repeated part at a discount. Six of the seven models came in 12% to 54% under list. Anthropic models cache only what a request marks; Quiriz now marks it, so Claude decks cost less from here than the $11.10 measured in this run.

Is a cheaper model per token always cheaper per task?

No. Gemini 3.8 Flash costs about 60% less than GPT-6.1 Sol per input token, but its deck was only 29% cheaper, because it made 59 analysis calls to Sol’s 33 and read twice as many tokens.