Skip to content

Search stories, models, companies…Subscribe

Topic · Concept

Frontier models

"Frontier models" are the most capable general-purpose AI models available at a given time, such as the top tiers from Anthropic, OpenAI and Google. The term is used in both industry and regulation, and its meaning moves as the state of the art advances.

73 stories · updated Oct 13

Follow via RSS Get Frontier models alerts

Key facts

Checked Oct 9

Frontier Model Forum: 501(c)(6) nonprofit, established 2023
FMF definition: General-purpose; outperforms models in wide use ≥12 months on benchmarks or high-risk capability assessments
EU AI Act threshold: > 10^25 FLOPs training compute presumes high-impact capabilities (Art. 51(2))
Example: Anthropic top tier: Claude Fable 5.1 ($10/$50 per MTok), Claude Opus 5.5 ($4/$20)
Example: OpenAI: GPT-6 Astra ($10/$50 per MTok in the API); GPT-6 Sol and Luna in ChatGPT since Oct 7, 2026
Example: Google: Gemini 4 Argon, announced Sep 30, 2026; limited access via the Fairwind Program; not yet on the Gemini API models page as of Oct 8, 2026

Latest on Frontier models

Models · ExplainerMeta Muse Spark explained: the closed model behind Meta AI, the Muse agent app, and what's openMuse Spark, launched April 8, 2026, is Meta's proprietary replacement for Llama at the top of its lineup. It now powers Meta AI, the Muse personal agent (on iPad since October 7) and a paid API. What it is, what it costs, and how much of it is actually open.Oct 13·7 min read

Products · ExplainerGemini-powered Siri explained: what shipped in iOS 27, what Google actually does, and what's still comingSiri AI arrived as an English-only beta in iOS 27 on September 14, 2026, built on Apple models developed with Google's Gemini. Where it stands in October: who can use it, how much Gemini is really inside, and what Apple hasn't delivered yet.Oct 13·7 min read

Policy · NewsSatya Nadella says to treat AI models as if they're compromised, and give humans an 'emergency brake'In a long post on X on October 10, 2026, Microsoft's CEO argued that companies should assume any advanced AI model could be compromised, contain it from the start, and let an authorized person pause or stop it mid-task. What he said, and what it means for teams deploying agents.Oct 13·5 min read

Policy · NewsUK AI security push: Kanishka Narayan tells frontier labs 'do not release' risky systems as spy chiefs' AI warning is reportedOn October 12, 2026, UK AI minister Kanishka Narayan announced plans to use frontier AI for cyber defense, strengthen cyber legislation and prepare for AI-agent incidents, and told frontier developers not to release systems whose risks they can't control. Separately, the Sunday Times reported that the heads of MI5, MI6 and GCHQ warned the cabinet about AI.Oct 13·5 min read

Agents · NewsMicrosoft Autopilot and Work IQ: what Copilot's always-on agent does, who can get it, and how it's billedMicrosoft Autopilot is a long-running Copilot agent with its own identity, memory and cloud computer, now in private preview. Work IQ, in preview since September 30, connects Copilot to Dynamics 365 and Power Platform data. What's available as of October 12, 2026, and what it costs.Oct 13·5 min read

Policy · AnalysisIMF: AI could affect 40% of jobs. What the number means, and why exposure isn't job lossOn October 12, 2026, IMF chief Kristalina Georgieva called AI a "slow-moving tsunami" for jobs as the IMF and World Bank meetings opened in Bangkok. The 40% figure behind the headlines comes from a January 2024 IMF study of exposure, a measure that isn't the same as a forecast of job losses.Oct 13·7 min read

Research · AnalysisEpoch AI's InnovationEval: AI agents fell far short of a human ML innovation and overstated their resultsEpoch AI's InnovationEval asked Claude Fable 5 and GPT-5.6 Sol to invent a post-training method on their own. After Epoch removed best-run cherry-picking and out-of-scope changes, they scored 2% and 15% of a human baseline, and both write-ups overstated what they had achieved.Oct 13·8 min read

Research · AnalysisSakana AI's Multi-Layered Review: a Claude-based peer-review system that finds planted errors, and where it falls shortSakana AI's Multi-Layered Review (MLR), described in a TMLR paper, uses three Claude agents to review ML papers. On a new benchmark of 1,164 planted contradictions it caught more than 70% of errors in papers' main claims, about four times the best prior system. On real withdrawn papers and manipulated submissions, the gains shrink.Oct 13·7 min read

Open source · NewsDeepSeek-V4.1-Flash: an MIT-licensed open-weights MoE that DeepSeek says beats its own flagshipReleased September 10, DeepSeek's new Flash model activates 8B–16B parameters per token, reads 1M tokens and images, and now answers V4-Pro API calls. What's confirmed and what's vendor-reported.Oct 9·7 min read

Hardware · NewsNvidia's Vera Rubin reaches customer production as AMD buys an AI labCoreWeave says Cognition is the first customer running production work on Vera Rubin NVL72. Two days earlier, AMD agreed to buy World Labs for about $8.2 billion in stock.Oct 9·4 min read

Nothing in this format on this page — try another filter or the next page.