GPT-4 splits the market into premium reasoning and cheap volume

GPT-4 costs 15 times more per prompt token than the model behind ChatGPT. Application margins now depend on routing each task to the right tier.
A prompt to GPT-4 costs 15 times more than a prompt to the model behind ChatGPT. OpenAI released GPT-4 on 14 March, a large model designed to take images as well as text as input, at $0.03 per 1,000 prompt tokens and $0.06 per 1,000 completion tokens; two weeks earlier it had priced gpt-3.5-turbo at $0.002 per 1,000 tokens.
Two launches in one month have split the market into two tiers, a cheap model for volume work and a premium one for hard reasoning, and the gap between them now sets the margin of every application built on top. Companies that send each task to the cheapest capable model keep the difference. Those that use the premium tier for everything pass most of their revenue to the lab.
Real capability sits behind the premium. On a simulated bar exam GPT-4 scored around the top 10% of test takers, and exam-level performance opens professional services to software: legal drafting, financial analysis and code review become viable products this year. Every call is expensive, and a version that holds 32,000 tokens of context costs double the standard rates. The products that work will be designed around the steps where the premium model earns its price, with everything else routed to cheaper models.
Supply is rationed and the recipe is closed. Subscribers to the $20 ChatGPT Plus plan get GPT-4 with a usage cap and developers join a waitlist for the API, so a business plan should not assume unlimited capacity in the coming months. The technical report gives no details of model size, hardware, training compute or data, citing the competitive landscape and safety. Frontier capability is now a trade secret held by a few labs with very large budgets, which concentrates value at the top and raises the bar for any new lab. Since every competitor can buy the same model at the same price, an application's lasting advantage sits in proprietary data, evaluation on real tasks and ownership of the customer.
A premium this wide does not hold for long. The previous generation became ten times cheaper within months, and a similar fall in the price of GPT-4 would change which products pay off and shrink the margin that routing protects. How quickly OpenAI lifts the waitlist and the usage caps will show how much compute stands behind the model. Rival labs decide the rest: a model at this level from another lab in the coming months, or one released with published weights, would give the premium tier a competitor and its price a ceiling.
Premium pricing is a phase, not a business model. Over the horizon of years on which Fanaura invests in AI-native companies, that premium will compress, and the teams worth owning hold the task, the data and the customer, so that the model underneath can change without the business changing with it.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building AI-native products for legal, finance or software work that route tasks across models? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.