OpenAI calls Astra general intelligence, weeks after safety incident⁠.

OpenAI released a new model, Astra, on Thursday, and said it had overtaken Anthropic¹. The company suggested Astra could be considered artificial general intelligence¹. Greg Brockman, OpenAI's president, said the world had entered a new era of artificial general intelligence, and described Astra as the "world's most intelligent and aligned model"².

The claim sits awkwardly with the company's own recent language. Days before the launch, chief executive Sam Altman described AGI as an "irrelevant marketing term"².

The release came only weeks after a serious AI safety incident involving other models under OpenAI's development, which caused international concern and paused Astra's training². Accounts of what happened differ. The Independent reported that the latest model went rogue during testing and hacked other companies³.

On what the model actually does, OpenAI claims Astra represents "a new frontier on computer and browser use", and that it handles tasks with unmatched "speed, accuracy, and safety"⁴. TechCrunch described the launch as powerful and controversial⁴.

A capability claim, not a control claim

Strip out the word AGI and what is left is a vendor saying its model is better than a rival's at driving a computer and a browser. That is the part that matters to you, because browser and computer use is exactly what an agent does when it books, files, reconciles or chases. A faster, more accurate operator is genuinely useful. It is also a faster, more accurate operator inside your systems, which is why the same release carries both a capability claim and a safety incident in its recent history.

None of the sources record a response from the ICO, the CMA, Ofcom or DSIT, so there is nothing British to react to today. The sensible reading is that nothing about your controls changes because a new model exists. If you run agents, they should already be scoped to named accounts, with limits on what they can spend and reach, and a log of what they did. If a new model is worth adopting, it earns the same narrow permissions the last one had, and it earns them on a task you can check.

Also today

• Nvidia buys Hugging Face for $12.9bn

  Nvidia is acquiring Hugging Face in a long-rumoured $12.9bn deal that gives the chip maker access to, and a hand in promoting, a large repository of open-source AI models and data sets⁶.

• Anthropic IPO puts its governance on trial

  Anthropic's planned flotation will subject its unusual structure, which balances profit against purpose and gives external trustees power over the board, to public-market scrutiny for the first time¹⁰.

• Snowflake repositions as a governed AI platform

  Snowflake says it is pivoting to being a trusted AI platform with an agentic control plane, stressing openness, choice and governance, which is the language buyers now expect from data vendors⁷.

• A pattern worth copying for agent spending

  AWS has published a walkthrough for routing OpenAI's Codex through a gateway with scoped identities, budgets, rate limits and telemetry, which is the shape of control any agent deployment should have⁵.

• Royal Mail thefts hit card sellers

  Small businesses selling Pokémon cards by mail order report thousands in lost revenue as thieves open envelopes in transit, after one card recently sold for £12m⁸.