AI News Week 25: Anthropic disables Claude 5 after three days, powerful AI on mini-servers, Apertus 1.5 delayed

1. Claude Fable 5: Live for Three Days, then Pulled Off the Net

Announced last week, now here: on 9 June, Anthropic released Claude Fable 5 – the most powerful model the company has ever made public. It writes code, analyses, performs scientific tasks, and completes long workflows autonomously, topping multiple benchmarks. Using it, Stripe migrated a 50-million-line codebase in a single day instead of two months. Three days later, it was gone.

What happened. One day after the launch, a developer reported that he had bypassed the safety filters. His trick (jailbreak) was simple: instead of asking a forbidden question directly, he broke it down into many harmless sub-questions. He had the model answer each one individually, and then reassembled the replies. Each sub-question flew under the radar of the filters – the final result did not.

Who pulled the plug. According to Axios, Amazon called senior government officials on Thursday evening and presented its own jailbreak report. At least five other companies also got in touch. On Friday, Anthropic received an export control order from the US government: Fable 5 and the more powerful Mythos 5 must no longer be accessible to any foreign national – globally. The only way to comply was immediate shutdown for all customers. A sensitive point: Amazon is Anthropic's largest investor and hosts its models in its own cloud.

Anthropic disagrees. The company calls the incident a misunderstanding. They state that the same information is available via other public models like GPT-5.5 without any tricks. A security expert consulted by Anthropic called the government's reaction "completely excessive". As of 15 June, Fable 5 and Mythos 5 remain offline, with no return date set. All other Claude models are running normally.

Context: The real issue here is availability, not security. A model was the best on the market for three days – then vanished by decree. Anyone who had built it deeply into their workflow was left without a replacement overnight. The lesson: never hitch your operations to a single model. You always need a plan B.

2. Plan B in Practice: Open Models on Your Own Hardware

The same day Fable 5 was shut down, Kimi K2.7 Code was released – with open weights on Hugging Face, scoring 81.1 per cent in the MCPMark tool-use benchmark. The percentage is not the point; the principle is. A government cannot shut down an open model running on your own hardware.

This was long considered expensive and impractical, but that is no longer true. While massive flagship models like Kimi (around 1 trillion parameters) still require a data centre, powerful models in the 70B to 120B class now run on a single mini PC. A device with an AMD Ryzen AI Max+ 395 ("Strix Halo") and 128 GB of memory runs 70B to 120B models at under 120 watts. This is more efficient and quieter than a gaming PC, costing just a few thousand Swiss francs. Nvidia unveiled the RTX Spark at Computex on 1 June. This chip for mini PCs serves models up to 120 billion parameters and 1 million tokens of context locally, according to Nvidia. Nvidia's DGX Spark (128 GB, around $4,700) is already on the market.

For SMEs, sovereignty is no longer just a choice between a US cloud and your own server hall. A box on the desk is enough for mid-sized models. For those not (yet) using their own hardware, routing across multiple providers – using Claude, GPT, Gemini, and an open model in parallel – is the most pragmatic approach. During the Fable shutdown, this secured operations within minutes. What was once "nice to have" has become "standard equipment".

3. Apertus: The Open Swiss Option – Version 1.5 is Coming

When it comes to independence from US providers, Apertus is the Swiss answer. It is the open language model from ETH, EPFL, and the CSCS supercomputing centre. Version 1.0 has been live since September 2025 and has been downloaded over 420,000 times. It comes in two sizes (8B and 70B parameters) under an Apache 2.0 licence, available via Hugging Face, Swisscom's sovereign AI platform, Phoenix Technologies, and Amazon SageMaker.

Its defining feature is radical openness. Not only are the model weights public, but so are the training data, code, and methodology. It was trained on the Swiss supercomputer "Alps" using around 15 trillion tokens in over 1,800 languages. 40 per cent of these are non-English – including Swiss German and Romansh.

What is next: Apertus 1.5. At the AMLD conference in Lausanne in February, the creators announced version 1.5. It will process images and audio as well as text, and it gains advanced reasoning capabilities, working through problems step-by-step rather than responding instantly. The launch was planned for mid-to-late May. As of mid-June, 1.5 is still not out, meaning it has been delayed.

From the engine room: CSCS Director Thomas Schulthess openly admits that the "Alps" supercomputer is already reaching its limits, as is Europe's new "Jupiter" computer in Jülich. His proposal: Switzerland should not go it alone, but cooperate internationally and make training data a public good. "We are very well positioned globally. But that could change if we fall asleep."

Regarding performance, Apertus 1.0 is solid but does not play in the same league as Claude, GPT, or Gemini. Version 1.5 should be different. According to internal benchmarks, the new version is expected to rank on par with Google's Gemma 4. That would be a significant leap, turning Apertus into a serious open alternative. Added to this is its real trump card: radical openness and sovereignty. It is a model on your own infrastructure, with fully traceable origins, immune to any US decree.

4. Mistral: The European Alternative Becomes Reality

Mistral is bringing its data centre near Paris online this month with around 13,800 Nvidia GPUs. The company employs 1,000 people and aims for €1 billion in revenue by 2026. This means the "European alternative to OpenAI and Anthropic" is no longer a promise. It runs on its own infrastructure within the EU. For anyone unwilling to send data to the US, Mistral is the second serious sovereignty option alongside open models. The difference is that it comes as a finished service, without requiring your own hardware.

5. In Brief: OpenAI and Google

OpenAI builds on memory. ChatGPT has received an improved memory. It now remembers context and preferences more reliably over time, and there is a new, viewable overview page of stored information (initially for Plus and Pro in the US). In parallel, OpenAI has simplified model selection (Instant, Medium, High, Pro) and deactivated the old GPT-5.2 models on 12 June. Active chats migration to GPT-5.5 is automatic. A leaked GPT-5.6 checkpoint is also circulating in developer channels.

Google keeps us waiting – as usual. Gemini 3.5 Pro (2 million tokens of context, "Deep Think" mode), announced at I/O, had not launched by 15 June. Sundar Pichai said "give us another month" in May, so it is now expected at the end of June. I know this pattern all too well. Features like project folders in Gemini were announced months ago, but nothing has arrived. Google has top developers and a top model, but suffers from a chronically slow rollout. No company is slower, especially regarding availability across different countries. OpenAI and Anthropic present a feature and roll it out globally on the same day. Google must learn this, or it will permanently damage its credibility. Personally, I no longer expect much and only use Gemini for consumer use cases like flight planning or quick queries. Tellingly, the World Cup feature (scores, highlights, tables, optionally as a morning briefing) has been live for a long time, but the professional model is not.

Market figures: According to the Ramp AI Index (mapping over 50,000 US companies), 41 per cent of US businesses with a paid AI subscription used Anthropic in June. This is more than any other provider, including OpenAI, for the first time. A year ago, Anthropic stood at around 8 per cent. Note that this measures US corporate usage, not global reach.

Three Things to Keep in Mind This Week

1. Build a fallback. The Fable 5 shutdown showed how quickly a model can disappear. If a process in your company relies on a single AI model, set up a second provider. Test whether your most critical tasks run smoothly there.

2. Treat "the best model" as a snapshot in time. This week, the most powerful public model was live for three days, then gone. Select for reliability and availability, not just benchmarks.

3. Check sovereignty options if you have sensitive data. Public authorities, hospitals, or data-sensitive companies should look at Apertus or Mistral. Not for everything, but for the workflows where a US decree would become a genuine problem.