Anthropic has released new AI models – and one of them was switched off again within days. Here is what happened, and what executives should take away from it for choosing and running AI in their organisation.
Two new AI models from Anthropic: Fable 5 and Mythos 5
Anthropic – the company behind Claude – introduced a new model called Mythos 5. It is so capable that the company initially did not release it to the general public: it had found security vulnerabilities in many established IT systems. Anthropic first made it available only to companies that can close those gaps. Alongside it, Anthropic released a smaller model from the same family for broader use: Fable 5.
The newest model was switched off
Last week a new model suddenly appeared – even stronger on security and on recognising visual content, including analysing what actually happens inside videos rather than just the subtitles. Then it was gone again. On Friday the US government issued an order citing a risk to national security and directed Anthropic to withdraw the model for non-US citizens. Anthropic could not technically separate US from non-US users, so it switched the newest model off completely.
Should you always jump to the newest model?
Not necessarily. The chance of a model being withdrawn is low, but it is a residual risk – and after just two or three days hardly anyone has built productive workflows on a brand-new model. More importantly: if a model you depend on were switched off (imagine Opus 4.7 or 4.8 disappearing), many services and automations built on it would suddenly break. The lesson: do not always run on the very latest model, and use solutions where you can swap the underlying model.
Digital sovereignty and swappable models
The shutdown made digital sovereignty a live topic. Some companies are now considering exit scenarios from US services – what if the US government decided to cut off Europe tomorrow? It is wise not to become fully dependent on a single provider. Options include vendors that host on German or European servers, such as Langdock, or – a further step – corporate, in-house LLMs running locally, which make you more independent of cloud services. The downside: locally installed models are still noticeably less capable than the large data-centre models. The good news: models keep getting smaller and more capable and already run well on five- or six-year-old Macs or with an external graphics card.
Bigger is not always better
There is a constant temptation to reach for the biggest, newest model. For demanding tasks that makes sense – but for small jobs a smaller (and often older) model is plenty, and cheaper. Newer top models can cost roughly twice as much per token as the previous generation, so you might save on staff costs while paying more for token usage. With Anthropic in particular, the strongest models are genuinely expensive.
Watch out: older models get retired
Assuming older models are never switched off is too optimistic: when new models arrive, older ones are often retired (for example Opus 4.6 or 4.7 becoming unavailable once 4.8 ships). Establish a process that checks, whenever a new model appears, how long the older one will remain available. The good news: Anthropic focuses on enterprise customers and tends to keep models available longer – but you cannot always control it, as the abrupt restriction of the newest model showed.
Run more than one system in parallel
It pays not to rely on a single track. Even if you mainly work in the cloud, consider using Copilot or ChatGPT in parallel – if you have used up your tokens in one system, you may still have some left in another.
Microsoft’s smart strategy
Microsoft’s approach is clever: instead of building its own models, it integrates the best models on the market. In Copilot you can choose Claude’s Opus models, for example. By default this is not enabled for companies in Microsoft 365 and has to be configured – so route data pragmatically: public information (for example in marketing) can run through Claude, while sensitive personal data is better processed in Copilot or, if you are not on Microsoft 365, in Langdock. Microsoft also goes beyond the chatbot, integrating Claude to execute tasks on the computer and working towards autonomous agents inside Microsoft 365. In short, Microsoft is strategically well positioned – and it is a good idea to start using Claude in selected areas in parallel.
Key Takeaways
- Anthropic introduced two new models: Mythos 5 (very capable, initially limited to selected companies due to discovered security risks) and Fable 5 (a smaller model for broader use).
- The newest model was briefly switched off for non-US users after a US government order (national security); Anthropic could not technically separate US and non-US users.
- Do not jump to the newest model reflexively: it can be withdrawn or become significantly more expensive.
- Digital sovereignty: avoid becoming fully dependent on US services and choose solutions where the AI model can be swapped.
- Options: EU/German providers like Langdock, or in-house/corporate LLMs (local, more independent, currently less capable but improving).
- Bigger is not always better: smaller, often older and cheaper models are enough for small tasks; new top models can cost twice as much.
- Watch out: when new models arrive, older ones are often retired (e.g. Opus 4.6/4.7) – set up a process to check availability windows.
- Run more than one system: use Copilot and ChatGPT in parallel (token budget, resilience).
- Microsoft’s strategy: it builds no models of its own but integrates the best (Claude/Opus selectable in Copilot, configurable in Microsoft 365); route data pragmatically.
Frequently Asked Questions
Which new AI models did Anthropic introduce?
Anthropic introduced two new models: Mythos 5, a very capable model that was initially made available only to selected companies because of discovered security risks, and Fable 5, a smaller model for broader use.
Why was the new AI model switched off again?
The US government saw a risk to national security and directed Anthropic to withdraw the model for non-US citizens. Because US and non-US users could not be separated technically, Anthropic temporarily switched the newest model off entirely.
Should you always switch to the newest AI model immediately?
Not necessarily. New models can be withdrawn, are often significantly more expensive, and are barely embedded in production in the first few days. It is wiser not to rely solely on the newest model and to use solutions where the underlying model can be swapped.
What does digital sovereignty mean for AI?
Not becoming fully dependent on individual (for example US) providers. Options include vendors with EU or German hosting such as Langdock, or local in-house models (corporate LLMs) that are more independent – currently less capable, but increasingly smaller and better.
Are older AI models switched off?
Yes. When new models appear, older ones are partly retired (for example Opus 4.6/4.7 when 4.8 ships). Companies should establish a process that checks, for every new model, how long the older one remains available.
How should Microsoft’s AI strategy be assessed?
Microsoft builds no models of its own but integrates the best on the market – such as Claude/Opus models, which can be selected in Copilot and configured in Microsoft 365. A pragmatic data-routing approach is recommended: public content can run through Claude, while sensitive personal data is better handled in Copilot or Langdock.
