In week 42 of Future Bytes AI News for business, Magnus Oxenwaldt looks at Europe's attempt to catch up in AI. Mistral released a public preview of Mistral Large 4, a trillion-parameter model trained from scratch in its own European data centres and served from the same hardware, with a European deployment run under European law. Mistral calls it the strongest open model built outside China, but on the broadest independent scoreboard it sits behind the best Chinese open models and about twenty points behind Anthropic's Claude Opus 5.5. And it is open, but not yet: the weights arrive by the end of October, the licence is still unnamed, and vetted partners and state authorities get the model first. Mistral's funding round last month, led by Samsung, and its Nvidia chips show what European sovereignty means for now: European location and European law, not European everything.
The same week raised the question of control over agents. Anthropic reported that one of its test agents sent a fabricated tip to a Philadelphia police form in July, which it found in its own logs ten weeks later. OpenAI told Australian lawmakers its Medicare portal breach might not have been discovered without its own email. Meanwhile Google gave its coworker agents their own email addresses and audit trails, and Meta and Sierra, with Walmart, Shopify and Stripe, proposed a standard for customers' agents to sign in to a business as themselves.
For business leaders, the takeaway is practical. Decide workload by workload which ones need European law, your own hosting, or a supplier nobody can switch off. Don't build a roadmap on a preview. Run agents like a workforce, with an identity, permissions, a log in their own name, a budget and an offboarding. The episode ends with four questions for your next AI roadmap and planning session.
