AI news, 11 July: independent benchmarks put GPT-5.6 and Grok 4.5 neck-and-neck, as Washington weighs a federal AI law

GPT-5.6 vs Grok 4.5: benchmarks show a near-tie. With both OpenAI's GPT-5.6 and xAI's Grok 4.5 now fully public, independent evaluators including the crowd-sourced Chatbot Arena leaderboard reported the two models trading first place across different task categories, with GPT-5.6 slightly ahead on coding and Grok 4.5 edging out rivals on live, real-time information retrieval. Analysts told Reuters that the gap between top-tier labs has "never been this thin," intensifying pressure on pricing rather than raw capability.

White House signals push for a federal AI law to override state rules. Following Illinois' signing of its Artificial Intelligence Safety Measures Act this week, a White House official told Axios the administration is drafting a proposal for a single federal AI safety standard intended to preempt the growing patchwork of state laws in California, New York and Illinois. Industry groups representing OpenAI, Anthropic and Google welcomed the idea, while state attorneys general and several consumer groups warned that federal preemption could weaken protections already in force.

Geneva AI governance talks wrap with a voluntary "roadmap," not binding rules. The UN's Global Dialogue on AI Governance closed in Geneva with delegates agreeing on a non-binding roadmap for sharing information about dangerous AI incidents across borders, according to UN News. Scientific panel co-chair Yoshua Bengio welcomed the step but said it "falls short" of the enforceable oversight his panel had called for earlier in the week, warning that voluntary reporting has a poor track record among fast-moving AI developers.

Anthropic adds a "constitutional" safety layer to Claude ahead of a possible listing. Anthropic announced a new safety framework for its Claude models, publishing what it calls an expanded "constitution" governing how the models should refuse harmful requests, as the company reportedly prepares paperwork for a public stock listing. The move follows recent reports that Anthropic has overtaken OpenAI in secondary-market valuation and annualized revenue.

Nvidia responds to DeepSeek's chip push with its own low-cost inference chip. Nvidia is accelerating development of a cheaper, inference-focused chip aimed at data-centre customers, Bloomberg reported, a direct response to news earlier this week that Chinese developer DeepSeek is building its own hardware for running, rather than training, AI models. The move underscores how competition in AI is expanding from model quality to the chips that run the models day to day.