Frontier models: the current list and what counts

A frontier model is a general-purpose AI model at the current state of the art: in the Frontier Model Forum's words, "general purpose AI models that constitute the state of the art". On 22 October 2026 the list is Gemini 4 Argon, Muse Wonder, Grok 4.7, GPT-6 Astra and Claude Fable 5.1.

Updated 22 October 2026

A frontier model is the current best general-purpose AI

The term is relative. The Frontier Model Forum, an industry body founded in 2023, describes frontier AI as "general purpose AI models that constitute the state of the art, a collection which will shift over time as the field progresses". So this page is a list, not a fixed definition: a model joins when it reaches the state of the art and leaves when newer models pass it. On 22 October 2026 the list below has five models, one from each of OpenAI, Anthropic, Google, Meta and SpaceXSI. It covers these five labs only; models from other labs, such as DeepSeek, Alibaba's Qwen team and Microsoft AI, have not been checked against the rule below. The labs behind them, and the words each uses for its goal, are compared on the labs page.

The current frontier models

List last updated 22 October 2026. Each row links to the lab's own announcement.

The rule for a row: the lab's own page must call the model frontier or its most capable, the model must be available to the public, and each lab gets one row, for the model its own pages rank highest. Anthropic's Claude Opus 5.5 (22 September 2026) is newer, but Anthropic's model guide says to use Claude Fable 5.1 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still don't meet your requirements". Cheaper models from the same labs, such as OpenAI's GPT-6.1 Sol, are left out because their makers rank them below the flagship. Moonshot AI's open-weight Kimi K3 is left out because its maker said in July 2026 that its "overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol"; both have since been replaced by newer models in the table.

Capability, compute and a lab's own label decide what counts

The first test is capability compared with the best existing models. For the 2023 AI Safety Summit the UK government defined frontier AI as "highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today’s most advanced models". The same paper left out "powerful narrow AI" such as AlphaFold, which cannot do as wide a range of tasks.

The Frontier Model Forum uses a looser test for its membership: "a general-purpose model that outperforms, based on a range of conventional performance benchmarks or high-risk capability assessments, all other models that have been widely deployed for at least 12 months". It says this is "intentionally more expansive than only “current state of the art” models".

The second test is training compute, which two laws use. California's Transparency in Frontier Artificial Intelligence Act defines a frontier model as "a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations", counting fine-tuning and reinforcement learning on top of the original run. The EU AI Act presumes a general-purpose AI model has high-impact capabilities, and so systemic risk, "when the cumulative amount of computation used for its training measured in floating point operations is greater than 10²⁵". The European Commission can change that number "for these thresholds to reflect the state of the art".

The third signal is the lab's own claim. Labs label their releases: Google calls Gemini 4 Argon "our latest frontier model", and SpaceXSI calls Grok 4.7 "SpaceXSI's frontier model for coding, agentic tasks, and knowledge work". A label alone is not proof, which is why the list above also checks how each lab ranks its own models.

The tests can disagree. None of the five announcements linked in the table states the model's training compute, so the announcements do not show whether each model passes the legal thresholds.

Every frontier model is a foundation model, not the reverse

Stanford researchers coined the term foundation model in 2021 for "any model that is trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks". California's law builds on it: a foundation model there is one "Trained on a broad data set", "Designed for generality of output" and "Adaptable to a wide range of distinctive tasks", and a frontier model is a foundation model above the compute line.

The UK paper draws the same line from the other side. Its scope diagram puts "Frontier AI (e.g. cutting-edge LLMs)" in scope and "Sub-frontier foundation models" out of scope. Foundation describes how a model is built and used. Frontier describes where it ranks today. A model remains a foundation model after it is overtaken; its frontier status does not. Smaller open models are often measured against the frontier, as in the Sage-A1.5 article.

Frontier ranks models; LLM and AGI describe them

A large language model (LLM) is a type of model. The UK paper said in 2023 that frontier AI "primarily includes large language models (LLMs)" but that frontier systems "may not be underpinned by LLMs, and could be underpinned by another technology".

Artificial general intelligence (AGI) is a capability level, not a ranking. In Google DeepMind's levels framework, written in September 2023, the authors placed the frontier models of that time low on the ladder: "current frontier language models would therefore be considered a Level 1 General AI (“Emerging AGI”)". Being at the frontier says a model is the best so far, not that it has reached any level. How AGI differs from the step above it is set out on the AGI vs ASI page, and that top step is explained on the what is ASI page. The wider idea of an intellect beyond humans is on the superintelligence page.

More compute and better algorithms move the frontier

The UK paper names three inputs: "computing power, data, and improvements in the underlying algorithms". It says training compute "has expanded over the past decade by a factor of 55 million", and that "better algorithms roughly halved compute requirements each year for vision and language models". It also notes that post-training changes such as better tools and prompts "can significantly improve performance in specific domains at a small fraction of the original training cost".

The definitions allow for this: the EU AI Act lets its threshold change "in light of evolving technological developments, such as algorithmic improvements or increased hardware efficiency".

Former frontier models

Models that have left the table, oldest first. Entries are added, never removed.

• Claude Fable 5 (Anthropic), released 9 June 2026. Replaced by Claude Fable 5.1 on 1 September 2026, which Anthropic says "is capable of much higher performance than its predecessor, Fable 5".
• GPT-5.6 Sol (OpenAI), released 9 July 2026. Replaced by GPT-6 Astra on 3 September 2026. OpenAI now calls it "our previous frontier cyber-capable model".

What changed in this list

• 22 October 2026: List started with Gemini 4 Argon, Muse Wonder, Grok 4.7, GPT-6 Astra and Claude Fable 5.1.