What a frontier model is, and when it's not worth paying

A frontier model is a large AI model that exceeds the capabilities of the most advanced existing models and can perform a wide variety of tasks. That is the definition given by Google, Microsoft, OpenAI and Anthropic when they founded the Frontier Model Forum ¹.

In practice, these are each provider's flagship models: the best at reasoning, writing code and reading long documents. And the most expensive.

What it costs to stay on the frontier

An example from OpenAI's price list: GPT-6 Astra costs 10 dollars per million input tokens and 50 for output, while GPT-6.1 Sol costs 2 and 10 ³. Five times as much, for a model that the provider itself describes as close to the most powerful one ⁴.

When you don't need it

Answering customer questions about opening hours, sorting emails, summarising meeting minutes, writing a draft product description: a mid-range model does these tasks well. Paying for a frontier model to do them is like sending the chief engineer to answer the phones.

The flagship model makes sense where mistakes are costly: contract analysis, decisions with many steps, agents that work alone for long stretches. Even then, it's worth testing it against a cheaper model, on your own cases, before you decide.

What about the AI Act?

The European regulation doesn't use the word "frontier". It speaks of general-purpose models with systemic risk, and presumes a model is one if it was trained with more than 10^25 floating-point operations of compute ². The obligations for these models fall on those who develop them, not on the small business that uses them.