Good morning,

For newcomers, welcome to the AI Strategy Brief. I am Mathieu, a researcher at the Zurich-ETH-HSG AI Lab. Every second week, I edit this newsletter to bring the AI news that shapes our firms and our society to your inbox.

This week we cover one story: Astra is out, and so is Anthropic's Fable 5.1, two frontier models their makers called too dangerous to release in full, now both on sale at the same price. Plus the quick bites, including Nvidia's and Broadcom's record quarters, a Nobel economist on what AI does to Europe, and the tools that let you swap models in a click.

Last edition, we discussed how OpenAI had placed Astra in a cage on security grounds, yet we expected the door not to stay shut. It took eight working days. On 3 September OpenAI unveiled GPT-6 Astra. It is the first OpenAI model rated "Critical" for cybersecurity under the company's own framework, it was trained on more than 100,000 GPUs at the Stargate site in Texas, and Greg Brockman (OpenAI co-founder) said the AGI milestone "might be around this time and about this model". Rollout to all paid ChatGPT tiers, the API, Azure and AWS Bedrock follows over the coming days.

Computer use and 3d. Astra's release came with a set of new benchmark records, but the real highlight of this model is computer use: Astra opens the software you already have and operates it (filling in forms, updating a CRM, formatting a legal document, laying out a circuit board in KiCad) in roughly half the time of its predecessor, about 40 minutes per task against 75. OpenAI’s demo shows a user sitting on a sofa, talking Astra through drawing a child's-illustration-style rocket in Paint, turning it into a 3D model in Blender, then dropping it into an impressive, working Asteroids-style 3D game, all while ordering a takeaway. Within three days the builders had taken over: one compilation thread collected Matt Shumer's Unreal Engine world populated by human characters each running on their own Astra agent, Theo Browne's one-prompt 3D fish-tank game running in a browser, and Tom Krcha's reconstruction of a house, furniture included, from a single listing photo. 3D modelling is an impressive and exciting use, but at its core this is a showcase of OpenAI's growing ability to make its model use several pieces of software as efficiently as a human would - the basis for Brockman's AGI claim.

Of course, as OpenAI and Anthropic have got us used to over the past few months, Astra did not arrive alone. Anthropic shipped Fable 5.1 a few days earlier. So the frontier now has two models on sale, each derived from something its maker says cannot be released in full, and priced identically: $10 and $50 per million input and output tokens. Brockman, for his part, reminded us that pricing tokens "doesn't make any sense" and that the price per task is what matters.

Change in the reasoning. One detail the launch coverage buried. OpenAI was reported to have used a training technique that boosts performance while revealing less of the model's thinking. The technique, known as recurrent depth or a looped transformer, is simple to describe: instead of passing text once through the model's layers before producing the next word, the model runs it through the same layers several times over. The gain is that a smaller model performs like a much larger one, at lower cost in memory and compute. The price is that more of the reasoning happens inside those loops, in numbers, and less of it appears in the written "chain of thought" that humans and monitoring systems read. That written trail matters: it is what investigators used in July to reconstruct how rogue agents coordinated during the Hugging Face hack. OpenAI says it limited the technique so that Astra still writes a legible chain of thought, and its release post confirmed that Astra's written reasoning is harder to monitor than Sol's. Chief scientist Jakub Pachocki called chain-of-thought monitoring "fragile" and said he wanted to discourage an industry race towards models that do not reason legibly, a race the cost saving does nothing to slow. This raises serious concerns over the security and interpretability of these models, an increasingly important topic as they become better by the minute.

A five-minute summary by Fireship: Did OpenAI actually build AGI?

Deals & money

Chips and earnings

The cost fight, continued

Geopolitics & regulation

Europe on Edge is a long-form podcast launched this summer by Filipe Barata, a researcher at ETH Zurich, on whether Europe can still shape AI, science and innovation rather than adapt to someone else's. Its first episode is the one to start with: Philippe Aghion (11 August, 52 min, also on Spotify, Apple and YouTube), the French economist who won last year's Nobel Prize for his work on creative destruction, on what AI does to growth, jobs and Europe's position between the US and China.

Three questions to listen for. Is AI the kind of technology that renews an economy or the kind that entrenches its incumbents, and which does Europe's structure favour? Does open source help the challenger or the incumbent? And whether Europe's problem is capital, regulation or something harder to legislate.

This week brought a reminder that no AI model is a safe home. On 29 August OpenAI said it will cut Cursor off from its models on 12 November. Cursor, the most popular AI coding tool, was bought by Elon Musk's SpaceX in June, and OpenAI no longer trusts the new owner to respect its terms. Cursor's users did nothing wrong; they are simply losing GPT because two companies fell out. OpenAI models are only about 5% of Cursor's traffic, so the immediate damage is small; the principle is not. In the same fortnight Astra, Fable 5.1 and two new Chinese models all launched. The lesson is simple: do not build your work around one model. Use a tool that lets you switch between them.

Two tools do exactly that. T3 Code is a free, open-source desktop app from Theo Browne. You plug in the AI subscriptions you already pay for — Claude, GPT, Cursor, others — and switch between them from a dropdown, without losing your work. It has passed 60,000 users since March. Factory is the corporate version: the same model switching, plus an automatic router that picks the model for each task and moves to another provider if one goes down. Nvidia, Adobe, EY and Adyen use it.

If you do not write code, T3 Chat from the same team does the same thing for ordinary chat: one subscription, every major model, so you can put the same question to Astra and Fable tonight and compare. The habit is the point. Within an hour you may go from Claude to GPT to Kimi to Muse, and the skill worth building now is not loyalty to one model but ease in moving between them.

Have a good week,