The Anthropic Institute's When AI builds itself, by Marina Favaro and Jack Clark, argues the world needs a verifiable, coordinated option to slow or pause frontier development, and backs it with internal data: as of May 2026 more than 80% of code merged into Anthropic's codebase was authored by Claude, engineers ship roughly 8x as much code per quarter as they did from 2021 to 2025, and AI task horizons are now doubling about every four months, according to METR, up from seven.

Anthropic is explicit that a pause only makes sense if multiple frontier labs across countries agree and can verify compliance, and that it is not pausing unilaterally. Clark took the argument to BBC Newsnight, saying the AI industry has a gas pedal but no brake pedal.

The Competition and Markets Authority imposed a binding conduct requirement on Google under the UK's digital markets regime, ordering it to give publishers effective controls over whether their content powers generative AI search features — AI Overviews, AI Mode and AI Overviews in Discover — an opt-out from fine-tuning, clear attribution links in AI answers and engagement metrics.

Google announced its compliance the same day with a toggle in Search Console, initially tested with a subset of UK publishers before global rollout, and said opting out is not a ranking signal for ordinary search. The CMA called it a world first. Full obligations land by December 2026, with potential fines up to 10% of turnover behind them. Google noted that AI Overviews have 2.5 billion monthly users and AI Mode has passed one billion.

At Build 2026 in San Francisco, Microsoft AI announced a family of seven models built in-house: MAI-Thinking-1, its first flagship reasoning model, and MAI-Code-1-Flash, an inference-efficient agentic coding model with 5 billion active parameters that is deeply integrated into GitHub Copilot and VS Code, alongside image, voice and transcription models.

Microsoft says the family was trained from the ground up on clean, commercially licensed data with no distillation from third-party models, that MAI-Thinking-1 is preferred to Sonnet 4.6 in blind human side-by-side evaluations, and that it is co-designed with its own Maia 200 silicon. MAI-Thinking-1 entered private preview in Microsoft Foundry the same day. Mustafa Suleyman framed the announcement as a repeatable hill-climbing machine — an internal pipeline from accelerator co-design through RL — reducing Microsoft's reliance on a single external lab.

Executive Order 14409, signed June 2 and published in the Federal Register June 5, directs agencies to set up a voluntary framework in which developers submit their most cyber-capable covered frontier models for government review up to 30 days before release, with a classified benchmarking process to define that threshold and the NSA director deciding which models qualify and which trusted partners get early access.

It orders an AI cybersecurity clearinghouse run through Treasury, NSA and CISA to coordinate vulnerability scanning and patching, and states explicitly that nothing in it authorizes mandatory licensing, preclearance or permitting. A day later OpenAI published a blueprint for democratic governance of frontier AI, backing a durable federal framework, a strengthened CAISI as the primary US frontier-safety institution, and preemption of duplicative state frontier-safety laws.

Anthropic said on June 1 that it had confidentially submitted a draft S-1 to the SEC, moving ahead of rival OpenAI in the race to public markets. It disclosed no share count or price and said any offering depends on market conditions and regulatory review.

Its most recent round was $65 billion raised in late May at a $965 billion post-money valuation, ahead of OpenAI. Reuters called it potentially one of the most consequential debuts in years and noted the counter-read from analysts: by filing first, Anthropic absorbs the disclosure risk of being the first frontier lab to show audited financials, while OpenAI gets to watch how institutional investors price it.