Frontier model

There is no single definition. The most common ones:

• Relative: the Frontier Model Forum calls frontier a general-purpose model that outperforms the models widely deployed for at least the previous 12 months.
• By compute (European Union): the AI Act presumes “systemic risk” for general-purpose models trained with more than 10^25 floating-point operations, with extra obligations for whoever offers them.
• By compute (California): SB 53, in force since January 2026, defines a frontier model as one trained with more than 10^26 operations.

As of November 2026

The labs considered frontier are American: Anthropic (Claude Fable 5.2), OpenAI (GPT-6), Google DeepMind (Gemini 4 Carbon), xAI (Grok 4.7) and Meta (the Muse family). On the Artificial Analysis index, the top model on November 12, 2026 was Claude Fable 5.2. The best Chinese models follow, with a gap that depends on the test (see Chinese and US models).

Why it matters at work

Frontier models are usually the most expensive. For many everyday tasks, a smaller or open-weight model does the job for a fraction of the price. The frontier makes a difference in long tasks, hard reasoning and agents with many steps.