# Frontier Model A **frontier model** is a [[Foundation Model]] at the cutting edge of capability — the most powerful AI systems in existence at any given time. The term entered policy discourse through the UK AI Safety Summit (November 2023) and the Biden-era AI Executive Order (October 2023), which used training compute thresholds (10^26 FLOP) to define which models trigger government reporting and safety requirements. As of mid-2026, the frontier includes: OpenAI's GPT-4.5/o3, Anthropic's Claude Opus, [[Google DeepMind]]'s Gemini Ultra, and — on the open-weight side — Meta's Llama 4. Chinese labs (DeepSeek, Alibaba's Qwen, Baidu's ERNIE) are approaching or matching the frontier on specific benchmarks. ## Why the Category Exists "Frontier model" is not a technical classification — it is a **governance category**. It exists to answer a regulatory question: which AI systems are potentially dangerous enough to warrant government oversight? The category creates a dividing line: - **Above the frontier line:** A small number of models built by a small number of organizations, subject to (varying degrees of) safety testing, government reporting, and responsible-deployment commitments. - **Below the frontier line:** Everything else — smaller models, fine-tunes, open-weight derivatives — operating with minimal or no oversight. The problem is that the frontier moves. Models that were frontier-class two years ago (GPT-4, March 2023) are now matched or exceeded by open-weight models that anyone can download and run. Regulation designed for today's frontier is perpetually chasing yesterday's. ## The Geopolitical Dimension Frontier models are the object of an explicit AI arms race between the United States and China. U.S. export controls on advanced GPUs (NVIDIA A100/H100, October 2022; tightened October 2023 and 2024) are specifically designed to prevent Chinese labs from training frontier models. China's response — DeepSeek's R1 (January 2025, trained on older hardware with novel efficiency techniques) — demonstrated that compute restrictions slow but do not prevent frontier capability development. The frontier-model arms race is the technological substrate of the [[The Quest for 90 Gigawatts]]: every frontier model requires more compute, more energy, more capital. The race to stay at the frontier is the economic engine driving the data-center buildout, the nuclear renaissance, and the trillion-dollar capital expenditures documented in the 90 Gigawatts tracker. ## KB Significance 1. **The governance bottleneck.** Who decides what counts as a "frontier model" determines who is subject to safety requirements. The Biden-era 10^26 FLOP threshold was a first attempt; the Trump administration's approach has been to deregulate rather than classify. The frontier-model category is the site of the regulatory-capture dynamics documented in the KB's defense-tech and TESCREAL sections. 2. **The concentration indicator.** The number of organizations capable of building frontier models is shrinking, not growing. Training costs for frontier models have risen from ~$10M (GPT-3, 2020) to ~$1B+ (current frontier, 2026). This cost escalation consolidates power in the hands of the organizations documented across the KB's tech-oligarchy coverage. 3. **The dual-use apex.** Frontier models are the most capable and therefore the most dangerous dual-use technology. Every capability advance at the frontier — better reasoning, longer context, multimodal understanding — simultaneously advances both beneficial applications and military/surveillance/propaganda applications. ## See Also - [[Foundation Model]] — The broader category - [[Scaling Laws]] — Why frontier models keep getting more expensive - [[AGI]] — The hypothetical endpoint of the frontier trajectory - [[Compute Sovereignty]] — The geopolitics of who can build frontier models - [[AI Safety]] — The governance framework for frontier models