Frontier model
- Training
- Industry term
A frontier model is a model near the leading edge of what AI can do at a particular time. In everyday use the label is informal and goes out of date as newer models arrive.
Most models near the edge come from a handful of companies that train at very large scale, though a few have reached it with far less. A model's place near it is read from public tests and from rankings where people compare two answers without knowing which model wrote each, so the label reports broad skill on those measures.
Where you’ll meet it
Your writing tool at work adds a model menu. The top option, marked “frontier” and priced at three times the standard one, sits above the model your team has used all year.
What to do
Try both on a few things your team writes every week, such as a client update and a meeting summary, and read the results side by side. If the standard model's versions read as well, keep it, and save the frontier option for the tasks where it does better.
Easy to mix up
The newest model loses that title at the next release. A frontier model keeps its label while it still ranks close to the leaders, even a year after it came out.
Check yourself
A news story says a new small model “matches frontier models” on a well-known math test. What does that tell you?
Go deeperLawmakers, labs, and leaderboards draw the frontier in different places
Some recent laws give the word a formal meaning, defining a frontier model by the computing power used to train it, a fixed line that a model crosses once. Labs and reporters use the word for the models near the lead at the moment, so a model can stay frontier under a law long after it has fallen behind.
Rankings have their own limits. A model can score well on a public test because questions from it leaked into its training data, which is benchmark contamination, and a test that every leading model aces no longer separates them.
Sources
- OpenAI model guidance, OpenAI
- Prompting Claude Fable 5, Anthropic