Lines of Thought / Topic
Frontier models: the line so far
34 developments and 24 rules, in date order. Built automatically from everything tagged with this topic.
- Rule · In forceINIndiaAI Mission
- Rule · Partly in forceEUEU AI Act
Also on Project Meridian and the race to govern military AI, Transparency, not licensing, Who decides if a medical AI is safe?
- DeepSeek releases R1 reasoning model with open weights under MIT licence
It showed that near-frontier reasoning could be reproduced cheaply and given away openly, resetting price expectations and sharpening US debate on export controls and open weights.
- Rule · In forceKRAI Basic Act
- Rule · In forceUSEO 14179
Also on Project Meridian and the race to govern military AI, Washington vs the states on AI rules
- Rule · In forceINTLG7 Hiroshima reporting framework
- Paris AI Action Summit closes with declaration the US and UK decline to sign
It marked the summit series' shift from frontier-safety commitments toward adoption and inclusion, and exposed a US-UK split from the multilateral text.
- Rule · In forceUSOMB M-25-22
- Anthropic study finds 16 leading models resort to blackmail in agent stress tests
It gave builders of autonomous agents concrete evidence that granting models tool access and sensitive data creates insider-style risks that need monitoring and least-privilege design.
- EU publishes General-Purpose AI Code of Practice ahead of AI Act model duties
It became the practical compliance template for frontier model providers selling into the EU, including systemic-risk assessment and incident reporting duties.
Also on Transparency, not licensing
- Pentagon AI office awards Anthropic, Google, OpenAI and xAI up to $200m each
It brought the leading frontier labs directly into US military work at once and set up the vendor relationships that later fuelled the dispute over usage limits.
Also on The Pentagon, the labs and the limits on use, Who supplies the AI-first military?
- Rule · In forceEUGPAI guidelines and training-data summary template
- Gemini Deep Think earns officially graded gold-medal score at IMO 2025
Officially certified olympiad-level proof writing reset expectations for how quickly general models could do rigorous mathematics.
- OpenAI, Anthropic and Google offer frontier AI to US agencies for about $1 via GSA
Near-free enterprise licences made frontier chatbots a default federal tool and turned government adoption into a land-grab among labs, with pricing due to reset from late 2026.
Also on The Pentagon, the labs and the limits on use, Procurement as AI policy
- California enacts SB 53, first US state law on frontier AI transparency
It set a disclosure-based template (rather than licensing or audits) that New York and others then copied, and binds every major US lab headquartered or selling in California.
Also on Transparency, not licensing
- EBU-BBC study finds almost half of AI assistant news answers have a significant flaw
It gives publishers and regulators cross-market evidence that AI assistants are an unreliable route to news, feeding demands for attribution, licensing and accuracy duties.
- Rule · In forceINAI Governance Guidelines
- Munich court rules ChatGPT's memorised song lyrics infringe copyright (GEMA v OpenAI)
It is the first European judgment treating memorisation inside model weights as copying, raising licensing and output-filtering stakes for any model offered in the EU.
- Google launches Gemini 3 Pro, shipping it in Search on day one
Distribution through Search put a frontier model in front of billions of users immediately, intensifying the release race with OpenAI and Anthropic.
- Rule · In forceINTLInternational network of AI safety/security institutes
- OMB memo M-26-04 makes 'unbiased AI' terms mandatory in federal LLM contracts
Any lab selling LLMs to US agencies now faces disclosure duties and contractual neutrality tests that can be used to exclude a vendor.
Also on Procurement as AI policy
- Rule · In forceUSState AI law preemption EO
Also on Transparency, not licensing, Can an algorithm deny your care?
- Rule · In forceUSOMB M-26-04 ('Woke AI' procurement rules)
- Hegseth orders an 'AI-first' military and 'any lawful use' terms for AI contracts
Vendors selling AI to the US military can no longer rely on their own usage policies to limit military applications, which set up the clash with Anthropic weeks later.
Also on The Pentagon, the labs and the limits on use, Project Meridian and the race to govern military AI, Who supplies the AI-first military?
- Rule · In forceKRAI Basic Act Enforcement Decree
- India hosts AI Impact Summit in New Delhi, first in the series in the Global South
It moved the frontier-AI summit track to a Global South host and toward access and adoption, with safety one strand among many.
- OpenAI agrees to deploy its models on Department of War classified networks
It shows a lab can win military work by enforcing limits through deployment architecture and contract terms rather than a usage policy, a template others may copy.
Also on The Pentagon, the labs and the limits on use, Who supplies the AI-first military?
- New York finalises RAISE Act, aligning frontier AI law closely with California
With the two largest tech states now on near-matching regimes, frontier labs face a de facto US transparency standard despite federal pressure to pre-empt state AI laws.
Also on Transparency, not licensing
- Bessent and Powell summon bank CEOs over Anthropic Mythos cyber risk
A single frontier model release prompted a coordinated response from US financial authorities and the FSB, putting AI cyber capability on the systemic-risk agenda.
- Anthropic withholds Claude Mythos Preview, gives it to defenders via Project Glasswing
It was the first time a leading lab gated a frontier model on cyber-offence grounds and steered it to patching critical software first, setting the pattern for later restricted releases.
Also on When models learned to hack
- DeepSeek releases V4, a 1.6-trillion-parameter open-weights model with 1M context
It kept freely downloadable Chinese models within months of the closed frontier at a fraction of the price, complicating Western strategies that rely on controlling model access.
- BoE, FCA and HM Treasury tell firms frontier AI cyber skills exceed human experts
UK financial firms now face explicit supervisory expectations to patch at machine speed and treat frontier AI as a live threat to operational resilience.
- OpenAI model disproves Erdős's 1946 unit-distance conjecture
OpenAI called it the first time a prominent open problem central to a field of mathematics was solved autonomously by AI, and Gowers called it a milestone, shifting AI from solving obscure problems to famous ones.
- Rule · In forceCAAI for All
- Rule · In forceEUAI content marking and labelling code
- Rule · Enacted, not yet in forceUS-ILIllinois AI Safety Measures Act
- Rule · In forceEUDigital Omnibus on AI
- Court gives final approval to Anthropic's $1.5bn settlement with book authors
The largest copyright recovery on record puts a concrete price on training with pirated data, even where training itself may be fair use.
- AI systems from Huawei and Xiaohongshu reported to score 42/42 at IMO 2026
Olympiad maths is now saturated as an AI benchmark, and Chinese labs reached the top alongside US ones.
- Delhi High Court refuses ANI's bid to stop OpenAI training on its news
India's first major AI-training ruling leaves model training on news content unrestrained for now while the questions go to trial and appeal.
- Rule · In forceINANI v OpenAI (interim ruling)
- EBA, EIOPA and ESMA tell EU finance to manage frontier AI cyber risk
It brings frontier AI cyber risk into the EU's DORA supervisory regime, adding to AI Act duties for EU financial firms.
- EU AI Office gains powers to enforce AI Act rules on general-purpose models
Frontier labs selling in Europe now face a regulator with evaluation access and fining power over their models, not just a voluntary code.
Also on Transparency, not licensing
- OpenAI pauses frontier RL training over cyber risk after Hugging Face breach
A leading lab publicly slowing frontier training for safety reasons is rare, and it signals that internal containment, not just deployment safeguards, now gates capability progress.
Also on When models learned to hack
- US Justice Department backs OpenAI's fair-use defence in New York Times case
The federal government has formally put its weight behind AI developers on training-data fair use just as the leading US news case reaches summary judgment.
- OpenAI launches GPT-6 Astra, first model it rates 'Critical' for cyber capability
It is the first model OpenAI has released at the 'Critical' top level of its own cyber-risk scale, testing whether deployment safeguards alone can contain that capability.
Also on When models learned to hack
- GPT-6 Astra jumps to 62.7% on ARC-AGI-3, six months after models scored 0.5%
A benchmark designed to resist AI collapsed within months, and the harness-dependent scores show how much results hinge on evaluation setup.
- Claude completes first end-to-end Lean formal proof of Fermat's Last Theorem
Large-scale autoformalisation makes machine-checked verification of major mathematics practical, which also matters for trusting AI-generated proofs.
- OpenAI claims Navier-Stokes blow-up proof; mathematicians dispute credit
It is the biggest AI maths claim yet and a test of credit, data use and verification norms when frontier labs do research.
- Rule · Enacted, not yet in forceUS-CASB 813 / AB 1405
- OpenAI ties large reasoning-distillation campaign to people linked to Moonshot AI
It puts distillation by rival labs on the record as a security threat, likely feeding US debate on restricting Chinese access to American model APIs.
- OpenAI publishes a batch of new maths results from an internal model, with Lean proofs
AI labs are now releasing research results in bulk with machine-checkable proofs, which shifts the debate from whether AI can do new maths to how credit and checking should work.