America's largest AI developers last week unveiled new models with a twist: their greatest selling point was not higher performance but lower cost per query. OpenAI's GPT-5.6, Meta's Muse Spark 1.1 and xAI's Grok 4.5 each promise to use fewer tokens, responding to enterprise clients who have been stunned by runaway bills. Senior executives in Europe report that a single month's usage of a frontier model can cost millions of dollars, prompting a hurried revaluation of how AI is budgeted. The push toward efficiency marks a pivot from the land-grab era and exposes a pricing war that is reshaping the competitive landscape in Silicon Valley.

As the cost conversation intensifies, a parallel privacy storm is gathering. A quiet settings update from Google, now rolling out globally, activates a 'Search Services History' that can save images users upload, voice searches and transcripts, and even audio from Translate speaking practice. The saved media, unless a user explicitly opts out, may be used to train Google's AI models. Privacy advocates in Europe note that the mechanism is set by default for many, and previously collected data can linger in training sets for up to four years, even after deletion. The episode fuels a wider consumer backlash over opaque data-collection practices for AI training.

Far from the spotlight, the insurance industry is trying to size a moving target. As businesses embed AI into everything from customer chatbots to supply-chain decisions, cyber underwriters in Nairobi and London are discovering that traditional policies do not cover AI-specific failures. Model hallucinations, data-poisoning attacks and unauthorised disclosures through AI systems create liabilities that sit outside standard cyber cover. Insurers are beginning to add endorsements for AI-related incidents and are demanding evidence of governance frameworks before quoting. The uncertainty itself is becoming a line-item cost that companies must manage.

Underpinning these frictions is a bitter dispute over data provenance. The distillation practices that allow one lab to train on another's outputs – decried by Anthropic as illicit 'attacks' – mirror the web-scraping that built the very models now being raided. The public row between Elon Musk and Sam Altman, trading accusations of scamming, is merely the most visible sign of a sector where intellectual property lines are blurred. Legal experts on both sides of the Atlantic are watching a tangle of lawsuits that will test the limits of fair use and terms-of-service enforcement for scraped and distilled training data. The coming quarter's court rulings and the European Union's forthcoming guidance on consent for AI training will set the guardrails for the next phase of expansion.