Illinois enacts legislation requiring documentation, independent audits, and safety mechanisms for high-compute AI models, establishing the US's strictest state-level framework.

Microsoft's M-Dash project utilises over 100 AI agents. These agents autonomously identify, patch, and validate software vulnerabilities, marking an advance in agentic security systems.

Anthropic re-enables global access to Fable 5 and Mythos 5 following classified negotiations regarding US export controls, setting a precedent for formal safety evidence and governmental collaboration in frontier model deployment.

Meta develops affordable AI smart glasses while allocating ten times Llama 3 compute to its 'Watermelon' frontier AI model.

OpenAI explores granting the US government a 5% equity stake, potentially underpinning a sovereign wealth fund to distribute AI prosperity.

'Meta Compute' offers enterprises access to Meta's data centres and trained models, diversifying revenue beyond advertising.

DeepMind has delayed the public release of Gemini 3.5 Pro to mid-July. The decision follows the discard of its initial base model and implementation of architectural modifications.

OpenAI consolidates its core LLM offerings to 'GPT-5.5', integrating substantial multimodal enhancements across text, image, audio, and video processing, as detailed in recent release notes.

OpenAI has consolidated its core model offerings under 'GPT-5.5', significantly enhancing its multimodal integration, as detailed in recent release notes.

OpenAI chief executive Sam Altman suggested a 5% equity donation to a US sovereign wealth fund, aiming to distribute future AI-driven wealth.

Etched, an AI inference chip developer, secured a $5 billion valuation following a funding round, projecting $1 billion in annual sales. This signals a shift within the AI silicon market.

Cloudflare mandates AI companies identify training crawlers, enabling content licensing for publishers.

Elon Musk's two ventures merge to build orbital data centres and route public-market exposure to Grok through a SpaceX listing said to target as much as $1.5tn. Hyperscaler AI capex now runs past $725bn a year.

With MMLU long exhausted, evaluation has moved to GPQA Diamond, SWE-bench Verified and Humanity's Last Exam — where no single model wins every column.

The new 6,000-person division aims to solve the 'last mile' of AI adoption by letting customers build bespoke solutions while retaining full ownership of their intellectual property.

The Claude maker's new model is trained for long-form narrative reasoning — and, in a twist, beats Opus 4.8 at agentic tool use when the tools are cast as characters. Priced at a third of Opus.

A +154% jump in valuation, an ARR of $47bn, and a multiple of 21×. The S-1 filing reframes what "AI lab" means on Wall Street: less research kitchen, more software distribution at unprecedented gross margin.