This term refers to the most advanced and capable AI models at any given time, situated at the frontier of technical knowledge. They are typically large-scale systems (with billions or trillions of parameters) trained on massive amounts of data and computing resources, and developed by a small number of organizations with the capacity to bear their high cost. Representative examples have been successive generations of models such as GPT, Gemini or Claude.
Their relevance is twofold. On the one hand, they set the state of the art: they define which tasks are technically possible and serve as a reference for the rest of the sector. On the other hand, they concentrate much of the debate on safety and regulation, as their emerging capabilities can entail risks that are difficult to anticipate.
Two nuances should be kept in mind:
- The term is relative and changing: what is a frontier model today will be surpassed in a few months.
- Not every practical use requires these models; in many cases, smaller and cheaper systems prove to be more efficient.