AI Startup Ideas Judged by Real Margins

Last updated: 2 November 2026

Judge an AI startup idea by asking whether revenue covers inference, hosting and support, then leaves enough gross margin after those costs.

Why judge an AI idea by gross margin first?

Gross margin is revenue minus cost of goods sold, divided by revenue. For an AI product, cost of goods sold includes inference, hosting and support. That makes gross margin a first test because a clever product can still fail if every request costs too much.

Typical software as a service businesses show gross margins of 65 to 70 per cent, according to Bessemer. AI companies are often in the 50 to 60 per cent range, according to a16z in 2020. More recently, ICONIQ’s State of AI 2026 report says the average AI product margin was 45 per cent in 2025, with projections of 53 per cent in 2026 and 59 per cent in 2027. Those figures are averages and projections, not promises for any single company.

a16z also notes that AI firms often spend 25 per cent or more of revenue on cloud resources. That matters because cloud and inference costs can rise with usage, not just with customer count. A startup with flat pricing and heavy users may see its margin fall as adoption grows.

What does one request cost?

The cost per request is input tokens divided by one million, multiplied by the input price, plus output tokens divided by one million, multiplied by the output price. The formula is simple, but the result changes sharply with model choice and output length.

Anthropic’s prices on 2 November 2026, per million tokens input and output, are: Claude Sonnet 5.5 at $2 and $10; Opus 5.5 at $4 and $20; Haiku 5.5 at $0.10 and $0.50 up to 100,000 tokens; and Fable 5.1 at $10 and $50. OpenAI’s prices on 31 October 2026 are gpt-6.1-sol at $2 and $10, and gpt-6-luna at $0.10 and $0.50. Gemini’s price on 31 October 2026 is Gemini 3.5 Flash-Lite at $0.30 and $2.50.

A worked example uses 3,000 input tokens and 800 output tokens on a model priced at $2 per million input tokens and $10 per million output tokens. That is 3,000 divided by 1,000,000 times $2, which equals $0.006, plus 800 divided by 1,000,000 times $10, which equals $0.008. The total is $0.014 per request.

Model	Input price per million tokens	Output price per million tokens
Claude Sonnet 5.5	$2	$10
Claude Opus 5.5	$4	$20
Claude Haiku 5.5	$0.10	$0.50 up to 100,000 tokens
Claude Fable 5.1	$10	$50
gpt-6.1-sol	$2	$10
gpt-6-luna	$0.10	$0.50
Gemini 3.5 Flash-Lite	$0.30	$2.50

Margin test for three idea types (all computed examples, ours)

Idea A is support-ticket drafting priced at $0.50 per resolved ticket. Using the same 3,000 input and 800 output tokens at $2 and $10 per million tokens, one model call costs $0.014. If the workflow needs five model calls per ticket, the model cost is $0.07. On model cost alone, the margin is 86 per cent. Hosting and human review should be added before quoting a final margin.

Idea B is a chat assistant priced at $20 a month per user. A heavy user sends 100 requests a day at $0.014, which is $1.40 a day and $42 a month. That is a loss on model cost alone. A light user with five requests a day costs $2.10 a month. Flat pricing is risky when usage varies so much between customers.

Idea C is long-document analysis. On Claude Opus 5.5 at $4 and $20 per million tokens, 150,000 input tokens and 3,000 output tokens cost $0.60 plus $0.06, which is $0.66 per document. On Haiku 5.5, 150,000 tokens exceeds the 100,000 token threshold, so the $0.50 and $2.50 rates apply: 150,000 divided by 1,000,000 times $0.50 is $0.075, and 3,000 divided by 1,000,000 times $2.50 is $0.0075, giving $0.0825. Model choice moves cost eightfold.

Which structures protect margin?

Structures that can protect margin include usage-based or outcome-based pricing, routing easy queries to cheaper models, prompt caching and capped output. These are editorial suggestions, not measured guarantees. They help align revenue with cost, especially where one customer uses far more tokens than another.

Our teardown of 12 startups shows how this looks in real companies. It is useful as a pattern library, not as proof that any single approach will work for a new idea.

Questions to ask before building

What does one unit of value cost in tokens? Who pays for heavy users? Can a cheaper model do 80 per cent of calls? Does revenue grow faster than inference? These questions turn a product idea into a margin test. Linked unit economics and runway guides can help with the wider financial picture.

Frequently asked questions

What is gross margin for an AI startup?

Gross margin is revenue minus cost of goods sold, divided by revenue. For AI, cost of goods sold includes inference, hosting and support.

How much does one API request cost?

It depends on model and token counts. The formula is input tokens divided by one million times input price, plus output tokens divided by one million times output price. In the worked example, 3,000 input and 800 output tokens at $2 and $10 cost $0.014.

Why is flat pricing risky for AI products?

Flat pricing is risky when usage varies. In the worked example, a heavy user costs $42 a month against a $20 price, while a light user costs $2.10 a month.

How much can model choice change cost?

Model choice can move cost eightfold in the long-document example. Claude Opus 5.5 costs $0.66 per document, while Haiku 5.5 costs $0.0825.

What should I add before quoting a final margin?

Add hosting and human review before quoting a final margin. The worked examples show model cost only, so they are illustrations rather than final margins.

Sources

Figures and prices were checked on 31 October and 2 November 2026; check the linked sources before relying on them. This article is general information, not tax, legal or investment advice.