Frontier model

The current most-capable AI models — typically training-compute-frontier LLMs from OpenAI, Anthropic, Google DeepMind, with broad capability and high cost.

"Frontier model" is industry shorthand for the current most-capable AI systems — typically trained with billion-dollar compute budgets, hundreds of billions of parameters, and broad general capability across reasoning, coding, vision, and language. As of 2026 it generally means GPT-5, Claude 4 (Opus / Sonnet tiers), Gemini 2.5 Pro / Ultra, and the most-capable Chinese models like Qwen3-Max and DeepSeek V4. It matters because frontier models are where new capabilities first emerge. They set the ceiling on what's possible; smaller open-source models follow 6-12 months later. Regulators and AI labs use "frontier model" as a category for safety policy — "frontier AI" gets extra scrutiny because of its capability and cost. A concrete example: when Claude 3 Opus launched, it set a new bar on legal, medical, and complex coding tasks. Within a year, similar capabilities appeared in cheaper Sonnet, Haiku, GPT-4o-mini tiers, and various open-source models. The frontier moves; the rest of the field follows. The term has policy weight: the UK and US have proposed frontier model evaluation requirements, the Frontier Model Forum (Anthropic + Google + Microsoft + OpenAI) coordinates voluntary safety practices, and "frontier AI" is the focus of most existential-risk discourse. Related: scaling laws, AGI, ASI, alignment.