AI Product Unit Economics in 2026: What a Query Costs, What Margins Look Like and How to Price

1. Reports

By Saroj Jha · 10 min read

What a query costs at the three largest model providers, how AI product margins compare, how to budget for a usage spike and how software companies price AI, from official price pages, filings and surveys with stated methods.

Most AI products pay a model provider for every answer. This report sets out what one query costs at the three largest providers, how the margins of software companies building AI products compare, what a usage spike can do to a bill, and how software companies are pricing AI. Prices are from each provider’s official price page, read on October 22, 2026; margins and pricing trends are from surveys and filings whose methods are stated.

A typical query costs from a twentieth of a cent to about five cents, depending on the model tier: for 3,000 tokens in and 500 out, a small model costs about $0.00055 and a mid-tier model about $0.011 at all three providers. Caching and batch processing cut that further, but two Gemini models are scheduled to double in price, and spending caps can lag: OpenAI says enforcement is not instantaneous and Google’s billing data can trail by about 10 minutes. The number to track is cost per completed task, not cost per query.

What a query costs

The table behind the figure gives the standard price per million tokens, input and output, for three representative models at each provider, and the cost of one illustrative query: 3,000 tokens in (instructions, context and the question) and 500 tokens out (the answer). The query size is an assumption, not a measurement; use your own logs for yours.

Two things the list prices do not show. First, long prompts cost more at some providers: Anthropic charges $0.50 per million input tokens and $2.50 output for Claude Haiku 5.5 above 100,000 tokens of prompt, and OpenAI’s long-context rates apply above 272,000 input tokens^[1],[2]. Second, keeping data in a chosen region adds about 10%: Anthropic applies a 1.1 multiplier to its US-only setting, and OpenAI charges a 10% uplift on regional processing endpoints for models released on or after March 5, 2026. Anthropic also notes that Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, so the same text is not always the same number of tokens across models.

Three ways the same query gets cheaper or dearer

Anthropic’s Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. Caching, batching and retries change what one answer costs.

The last line is why cost per completed task, not cost per query, is the number to track. The caching figure ignores the one-time cost of writing the cache, which Anthropic prices at 1.25 times the input price for a five-minute cache. A cache read costs 5% of the input price on Claude Opus 5.5 and Claude Sonnet 5.5^[1]. Batch processing gets a 50% discount on both input and output tokens at Anthropic and at OpenAI, and Google lists a 50% cost reduction for its Batch API^[2],[3].

Prices fall on average, but individual prices can rise

Stanford’s AI Index 2025 found that the cost of querying a model that scored the equivalent of GPT-3.5 on the MMLU benchmark fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024, “a more than 280-fold reduction” in about 18 months^[4]. That is capability getting cheaper, not a rule that every price falls.

Google’s price page shows the opposite for two current models. The gemini-4-argon price of $2.00 input and $10.00 output per million tokens is listed through November 30, 2026 and $4.00 and $20.00 starting December 1, 2026. The gemini-3.8-flash price of $0.75 and $3.75 runs through December 31, 2026, then $1.50 and $7.50 starting January 1, 2027^[3]. OpenAI’s page says GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026^[2]. A falling price per token does not guarantee a falling cost per customer: users ask more, products call models more often, and some prices are rising.

Margins: lower than classic software, improving

ICONIQ’s State of AI 2026 report draws on a Q2 2026 survey of about 300 executives at software companies building AI products, 87% of them in North America. It reports that “Gross margins are improving, jumping from 45% in 2025 to a projected 53% in 2026, and 59% in 2027,” and that two thirds of respondents report improved per-query unit economics^[5]. The figures are what executives reported and projected, and ICONIQ is an investor. The same report gives one builder’s warning: a workflow “budgeted at ten cents a run” drifted “past a dollar fifty as agents retried and corrected themselves.” That is one anecdote, not a measured rate, but it shows how retries multiply cost.

Public filings show what AI costs do to a margin. Duolingo’s fourth-quarter 2024 shareholder letter says gross margin fell about 120 basis points to 71.9% “due to lower subscription margins from increased generative AI costs due to the increased adoption of Duolingo Max, and lower advertising gross margin”^[6]. Its third-quarter 2025 letter says gross margin declined about 40 basis points to 72.5% “as a result of increases in generative AI and hosting costs”^[7]. Duolingo’s margins stayed far above the survey’s, so one company is a signal of direction, not a benchmark.

Spikes, and the limits that catch them

A launch, a viral post or one heavy customer can multiply usage in a day. At $0.011 a query, 100,000 queries a month cost $1,100; a month of ten times the usual traffic would cost $11,000. Each provider offers a way to cap spending, with different behavior.

In practice: keep free and trial traffic in its own workspace or project with its own cap, set daily limits per user inside your own product, and alert well below the hard cap, because a provider cap that stops everything also stops paying customers.

How software companies price AI

Three sources with stated methods point the same way.

The direction is consistent: away from flat fees and pure seats, toward pricing that rises with use. The three samples differ in who they cover, so the percentages are not comparable with each other.

What this means for a startup

1. Measure cost per completed task from your logs, including retries, tool calls and hosting, not cost per query from a price page.
2. Use the cheapest model that passes your own evaluation, and cache the instructions that repeat. On the example above, moving from a mid-tier to a small model cuts the cost of one query by a factor of 20 before caching or batching.
3. Price on a unit that rises with use, with a base fee that includes an allowance. Do not sell unlimited use at a fixed price.
4. Budget for the spike before the launch: ten times normal usage for a month, a separate capped workspace for free users, and daily limits per user.
5. Re-check prices every quarter. They fall on average, but individual models can double.

Method and limits

Prices are the standard rates on each provider’s official price page on October 22, 2026, for three representative models each; other models, regions and contract discounts differ, and prices change often. The 3,000-in, 500-out query and the 1.4 retry rate are illustrative assumptions. ICONIQ’s margin figures come from an executive survey and projection; Poyar’s survey was open to readers of his newsletter, so it is not a random sample; Bain’s review covers public pricing pages. Duolingo is one company. AAJ did not use media reports of providers’ internal margins, because no filing states them. This report is general information, not financial advice.