What are today's frontier AI models? The current lineup, how they compare, and why the strongest are harder to get

In the seven weeks from September 3 to October 23, 2026, OpenAI, Anthropic, Google, Meta, Mistral and Moonshot AI all released new top-tier models. Each launch came with a chart showing the new model ahead of its rivals on some test. If you are trying to choose a tool, explain AI to a colleague, or simply follow the news, the result is a fog of version numbers and conflicting claims.

This article clears some of that fog. It explains what "frontier model" means today, lists the current leading models from each major lab, compares them using independent measurements rather than marketing charts, and looks at a newer trend that matters as much as raw capability: the most powerful models are now often withheld, restricted or released with built-in safeguards. Every figure here carries a date, because this snapshot will age quickly.

What "frontier model" means in late 2026

A frontier model is one of the most capable AI systems available at a given moment. The term is relative: today's frontier is next year's mid-tier. Most frontier models are large language models (LLMs), systems trained on huge amounts of text, and increasingly images, audio and video, to predict and generate content.

A few terms help when reading launch announcements:

• Model family and tiers. Labs now release families: a large flagship plus cheaper, faster versions. OpenAI's GPT-6 line includes Astra (its most capable), Sol and Luna. Anthropic's includes Fable, Opus and Sonnet.
• Reasoning effort. Most frontier models "think" step by step before answering, and users can set how long. Benchmark scores are often reported at "max" effort, which costs more than everyday use.
• Tokens and context window. A token is a chunk of text, roughly three-quarters of a word. Prices are quoted per million tokens, and the context window is how many tokens a model can consider at once. One million tokens is now common at the top.
• Open-weight model. A model whose trained parameters are published so anyone can download and run it. Closed models are available only through the developer's app or paid interface.
• System card. A technical report a lab publishes at launch describing a model's capabilities, safety testing and known risks.

The current lineup, lab by lab

Here is where the main labs stood as of October 24, 2026. Benchmark figures in this section are the developers' own unless stated.

OpenAI. GPT-6 Astra, released September 3, is OpenAI's most capable model and, in its words, its "first model to reach the Critical level of cybersecurity capability" under its internal risk framework, the Preparedness Framework⁵. On October 7, OpenAI made GPT-6 Sol (paid tiers) and GPT-6 Luna (free tiers) the default models in ChatGPT, which it says reaches more than 1.2 billion people a week⁷. OpenAI rates both as "High" capability in cybersecurity and in biology and chemistry, but below its Critical cyber threshold⁶.

Anthropic. Anthropic's line has an unusual shape. Its Claude Mythos model was first disclosed in April 2026 but not made generally available because of its ability to find software vulnerabilities. Access went to a group of partners through an initiative called Project Glasswing¹⁰. Claude Fable 5.1 is a more widely available sibling, deployed with safeguards in high-risk areas⁹. Claude Opus 5.5, released September 22, is described by Anthropic as performing "at the level of Claude Fable 5.1 on most work" while costing 40% less to run than its predecessor, Opus 5⁸. Its list price is US$4 per million input tokens and US$20 per million output tokens, down from US$5 and US$25⁸.

Google. Google's newest model, Gemini 4 Argon, was announced September 30 but rolled out only to a select group of cybersecurity partners. TechCrunch reports it was "trained specifically for defensive cyber work"¹¹. A week earlier, DeepMind's Koray Kavukcuoglu said the broader Gemini 4 had entered post-training and safety testing, and that Google wanted to release it "as soon as possible"¹⁸. Its most recent widely available update at that point was Gemini 3.8 Flash¹⁸.

Meta. On October 12, Meta released Muse Wonder through its paid API, describing it as its first frontier model. A more capable Muse Wonder Full is limited to subscribers of its coding product¹². Meta's model card lists a one-million-token context window and compares it against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5¹³.

Challengers outside the US. Paris-based Mistral released a preview of Mistral Large 4 on October 6, a model with about one trillion parameters (the internal numbers a model learns), trained on 3,800 Nvidia chips in European data centres, with weights promised by month's end¹⁴. China's Moonshot AI launched Kimi K3.1 Pro on October 23. Artificial Analysis, an independent benchmarking firm, ranks it fourth on its index and expects it to become the highest-scoring open-weight model once Moonshot publishes the weights⁴.

How close are they? What independent scoreboards show

Developer charts are selective by design, so two independent trackers are more useful.

Epoch AI, a research group, combines many benchmarks into one score, the Epoch Capabilities Index (ECI). As of mid-September, GPT-6 Astra led with 166, ahead of Claude Fable 5.1 at 164 and GPT-5.6 Sol at 162¹. But the 90% confidence intervals overlap heavily (Astra 163–172, Fable 5.1 161–168), and the ranking flips by domain: Astra leads on math, while Fable 5.1 leads on software engineering¹. Epoch also notes that Astra's score fell from 169 at launch to 166 as more results arrived¹. In plain terms, the top models are statistically hard to separate.

Artificial Analysis runs its own tests and publishes a 10-evaluation Intelligence Index. At its September 7 relaunch, Fable 5.1 and GPT-6 Astra tied at 53³. By October 23, Claude Opus 5.5 led at 58, with Claude Sonnet 5.5 and Muse Wonder Full at 56 and Kimi K3.1 Pro at 55⁴. A tie on the headline number can hide big differences in cost: Artificial Analysis found that Astra's average cost per index task was US$3.26 against US$7.63 for Fable 5.1, 57% lower for similar scores³.

Even single benchmarks don't agree across sources. On Terminal-Bench 4.0, a test of completing tasks in a computer terminal, Anthropic reports Fable 5.1 at 55.8% and Astra at 57.9%⁸, while Artificial Analysis measured 52.0% and 59.1%³. Different settings, harnesses and run counts produce different numbers. Anthropic itself cautions that "benchmark margins have become a less reliable guide to real-world differences"⁸.

How fast the frontier is moving

The broader trend is clearer than any single ranking. Epoch estimates that since OpenAI's first reasoning models in September 2024, the top ECI score has risen by about 14 points a year, compared with 6 points a year before reasoning models². The UK's AI Security Institute (AISI), which has tested more than 30 frontier systems, reported in December 2025 that performance "in some areas is doubling every eight months"¹⁷.

The gap between closed and open-weight models has stayed modest. AISI put it at four to eight months in late 2025¹⁷. On cybersecurity specifically, Epoch's October 16 analysis found the best open-weight model, Z.ai's GLM-5.3, roughly where the closed frontier stood six to seven months earlier¹⁵.

One widely cited measure is losing its usefulness at the top. METR, a nonprofit that measures how long a task (in human working time) AI agents can complete, now says measurements above 16 hours "are unreliable with our current task suite," and as of September 8 its public tracking page is "no longer actively updated"¹⁶. When the tools for measuring progress saturate, claims about the newest models rest more heavily on developers' own reporting.

The new pattern: staged and restricted releases

The biggest change in 2026 isn't a benchmark score. It's how the most capable models reach the public.

• Withholding. Anthropic said it did "not plan to make Claude Mythos Preview generally available" and gave access first to partners including AWS, Apple, Google, Microsoft and Nvidia, plus more than 40 other organizations, to find and fix software vulnerabilities¹⁰.
• Limited partners. Google's Gemini 4 Argon went only to cyber partners in its first release¹¹. Meta keeps Muse Wonder Full off its public API¹².
• Safeguards with fallbacks. When Opus 5.5's classifiers detect requests in high-risk areas such as biology or cybersecurity, the request is handed to an older, less capable model rather than simply refused⁹.
• Formal risk ratings. Labs now publish where models sit against their own danger thresholds. OpenAI rates Astra "Critical" for cyber⁵. Meta's card says that without mitigations it does "not rule out" Muse Wonder reaching its threshold in chemistry and biology, but judges the risk "moderate or lower" with safeguards in place¹³.

This matters for ordinary users in two ways. The model you can buy may not be the most capable one a lab has built. And the model's behaviour on sensitive topics can change depending on what its safety systems detect.

Limits and open questions

Most evidence is self-reported. System cards are detailed, but they are written by the companies selling the models. Independent trackers help, yet they also depend on access the labs grant.

Testing is getting harder. Anthropic writes that Opus 5.5 "often suspects it is being evaluated," and that catching every failure before deployment "remains an unsolved problem"⁸. OpenAI reports that Astra's "monitorability," how well its reasoning can be inspected for warning signs, decreased relative to GPT-5.6 Sol⁵.

New versions can regress. OpenAI's October report found GPT-6 Sol showed "a statistically significant regression" on self-harm safety tests⁶. Newer doesn't always mean better on every dimension.

Safeguards are imperfect. AISI found "universal jailbreaks," prompts that bypass a model's safety rules, in every system it tested in 2025, though the effort required was rising¹⁷.

What this means for you

• Don't choose by headline rank. The top four or five models sit within statistical noise of each other. Test two or three on your own work, including your hardest cases.
• Compare cost per task, not price per token. A model that "thinks" longer can cost more than twice as much for a similar score³. Ask vendors for task-level cost on your workload.
• Check which model you're actually getting. Read whether your plan uses the flagship, a smaller tier, or a fallback model for some requests. Under the hood, "GPT-6" or "Claude" can mean several different systems.
• Read the system card's limitations section. It's the most candid part of most launches and often names failure modes the marketing omits.
• Plan for churn. Six major launches in under two months means today's choice will likely be overtaken within months. Keep your prompts, test cases and data portable between providers.

Sources

1. Barry, A., & Lee, J. (2026, September 16). GPT-6 Astra leads on math benchmarks, but not on software engineering. Epoch AI, Data Insights. epoch.ai/data-insights/astra-eci-breakdown
2. Barry, A. (2026, September 1). The ECI frontier has advanced by 14 points per year since the introduction of reasoning models. Epoch AI, Data Insights. epoch.ai/data-insights/eci-frontier-trend
3. Artificial Analysis. (2026, September 7). Announcing the Artificial Analysis Intelligence Index v4.3. artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3
4. Artificial Analysis. (2026, October 23). Kimi K3.1 Pro launches at 4th on the Intelligence Index and will lead open weights models. artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks
5. OpenAI. (2026, September 3). GPT-6 Astra System Card. OpenAI Deployment Safety Hub. deploymentsafety.openai.com/gpt-6-astra
6. OpenAI. (2026, October 7). GPT-6 Sol and GPT-6 Luna: October 2026 update. cdn.openai.com/pdf/gpt-6-october.pdf
7. OpenAI. (2026, October 7). GPT-6 and Intelligent UI for everyone. openai.com/index/gpt-6-for-everyone
8. Anthropic. (2026, September 22). Introducing Claude Opus 5.5. anthropic.com/claude-opus-5-5
9. Anthropic. (2026, September 22). System Card: Claude Opus 5.5. anthropic.com, Claude Opus 5.5 system card (PDF)
10. Anthropic. (2026, April 7, updated October 6, 2026). Project Glasswing: Securing critical software for the AI era. anthropic.com/glasswing
11. Ropek, L. (2026, September 30). Google releases Gemini 4 Argon, called its most powerful model yet. TechCrunch. techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet
12. Brandom, R. (2026, October 12). Meta releases its first frontier model, Muse Wonder. TechCrunch. techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder
13. Meta. (2026, October 12, updated October 23, 2026). Muse Wonder model card. research.meta.ai/static/muse-wonder-model-card
14. Mistral AI. (2026, October 6). Introducing Mistral Large 4. mistral.ai/news/mistral-large-4
15. Lee, J., & Barry, A. (2026, October 16). Open-weight models trail closed models in cyber capabilities by 6–7 months. Epoch AI, Data Insights. epoch.ai/data-insights/open-closed-cyber-eci-gap
16. METR. (2026, last updated September 8). Task-completion time horizons of frontier AI models. metr.org/time-horizons
17. AI Security Institute. (2025, December 18). Frontier AI Trends Report. UK Department for Science, Innovation and Technology. aisi.gov.uk/research/aisi-frontier-ai-trends-report-2025
18. Schoon, B. (2026, September 24). Google says Gemini 4 release is coming "as soon as possible." 9to5Google. 9to5google.com/2026/09/24/google-says-gemini-4-release-is-coming-as-soon-as-possible