First, a quick question: how good are these models?
It is tempting to dismiss this as marketing. That would be a mistake, because the jump in capability is huge. I used Fable for the full three days it was available and could hardly believe what I was seeing. Yes, I hit my token limits very quickly; it is expensive. But it is also insanely good. The model found improvements that Opus 4.8 never found.
I was able to solve things in fifteen minutes that I had spent hours on before without success. It was fantastic.
Anthropic released Mythos in April 2026 as its most powerful model to date. During internal testing, it proved so strong in code and security research that it found thousands of unknown vulnerabilities in nearly every major operating system and browser. Some of those flaws had been sitting in code reviewed by countless developers for years. Mythos found them in weeks. Anthropic found the model so sensitive that it first limited access to a small group of partners, under the name Project Glasswing. Fable 5 is the public version of that same technology, with strict limits built in: in high-risk areas such as cybersecurity and biology, the model blocks and falls back to the existing Claude Opus 4.8.
OpenAI is at the same level. The GPT-5.6 series consists of three models: Sol as the heaviest, Terra for everyday work, and Luna for speed and low cost. OpenAI claims that Sol is slightly better than Mythos 5 on some coding tasks, and that it achieves comparable capability with a fraction of the token usage. The U.S. government and OpenAI internally regard the model as equivalent to Mythos.
Independent rankings confirm that picture. On Artificial Analysis’s Intelligence Index, Fable 5 is at the top, with Opus 4.8 and GPT-5.5 just below it. The models do what is claimed, and they get better every few months.
The new reality: the emergency brake is in Washington
What happened in June matters more than which model is exactly on top.
The order for Anthropic to take Fable offline came from the U.S. government. The Department of Commerce prohibited access to Fable and Mythos for everyforeign national, inside and outside the U.S., including Anthropic’s own non-American employees. Because Anthropic could not separate users by nationality quickly enough, the only technically possible move was: everything off, for everyone, that same evening. The trigger was a reported jailbreak that could have unlocked Mythos’s cyber capabilities. Anthropic disputed how serious it was and called it a misunderstanding, but complied with the order.
Beneath that lies a broader structure. In early June, President Trump signed an executive order asking AI companies to make their heaviest models available for review up to thirty days before release, with government input on whichtrusted partnersget early access. On paper, voluntary. In practice, a licensing regime, as a former White House AI adviser put it. The GPT-5.6 story shows how that works out: OpenAI was allowed to start with about twenty U.S. partners whose participation had been shared with the government, and immediately said it does not see that as a sustainable standard.
Notice the word missing from that executive order: allies. There is not a single clause saying Europe belongs on the list. Whether we get access to a new frontier model quickly, then, is decided in Washington case by case..
By now, the situation seems less severe than first presented. On June 26, the government allowed Anthropic to make Mythos 5 available again to about a hundred U.S. organizations that run critical infrastructure. Fable 5, the model for the general public, is still switched off at the time of writing, and it may well return soon. We only think Fable’s return does not change the lesson. The mechanism remains. The switch sits with a government on the other side of the ocean.
Why this hits Europe harder than we like to admit
Europe has almost no homegrown frontier models and little computing power to train them.
A group of European researchers and policymakers published the scenarioEurope 2031in June, written under the leadership of Daan Juijn of the Arq Foundation. It is a fictional story, but rooted in existing trends, and it sharpens the balance of power. The United States hosts around 80 percent of global AI computing power, Europe about five. The largest AI supercomputer in the U.S. runs on more than a thousand megawatts, the largest in Europe on tens. The European Gigafactories meant to close that gap are years behind schedule. The scenario describes exactly the dynamic we are now seeing on a small scale: an American government discovering that access to models is leverage, and having no reason to give that leverage up.
You do not need to believe the darkest ending of that story to take the core seriously. Our dependence has become double. We rely on American providers, and those providers now rely on their own government for permission to serve Americans and, even more so, anyone outside the country.
Mistral deserves our attention
If Europe has one card to play, it is Mistral. The Paris-based company was founded in 2023 and has very clearly positioned itself as the sovereign option: models that governments and companies can run on their own infrastructure, beyond the reach of American export rules.
That position is gaining weight. In September 2025, Mistral raised a €1.7 billion Series C led by ASML, and is reportedly in talks over a new round of about $3 billion at a valuation around $20 billion. Revenue has grown fast, and the customer list now includes more than a hundred major organizations, among them the governments of France, Germany and Greece. The company is working on its own cybersecurity model as an alternative to Mythos, and founder Arthur Mensch warned the French National Assembly that Europe has a short window to prevent deeper dependence.
At the same time, we need to be honest about scale. Mistral has raised about four billion dollars in total. The big American labs sit on multiples of that. If we keep going like this, Mistral will not win the general arms race for the very best model. And that is something we in Europe should do something about. We are not only lagging when it comes to computing power. Many of Europe’s best AI experts also work in America, at companies such as Alphabet, OpenAI and Anthropic. While there does not seem to be much wrong with these companies from Europe’s direct point of view, the danger is that they are mainly focused on giving American companies better AI models.
Open source as a viable alternative
The second answer to dependence lies in open models, and a lot has changed in just one year.
Take GLM-5.2 from the ChineseZ.ai, released in mid-June and available for free download. Independent benchmark Artificial Analysis calls it the best open model in the world and ranks it fourth overall, behind only the closed top tier of Fable 5, Opus 4.8 and GPT-5.5 (at that time, OpenAI’s newest models had not yet launched). It runs at about one-sixth of the cost of GPT-5.5. And most tellingly: the model is said to match Mythos in automatically finding security bugs, precisely the capability cited as a national security risk when Mythos was kept behind closed doors.
There is a catch, though.Z.aiis on the U.S. Entity List, and if you use the model through their hosted API, your data passes through China. For sensitive workloads, that is not an option. The solution lies in the nature of open source itself: you run the full model on your own infrastructure and keep the data inside. That way, you use the capability without handing the keys to the provider.
Not every problem requires a frontier model. For much everyday work — summarizing, translating, simple analysis, document processing — a small model is more than enough.
Google released Gemma 4 in April under a free Apache 2.0 license, in formats ranging from a few gigabytes to a model that fits on a powerful laptop. The smallest versions, for example, run perfectly well on my iPhone 16 Pro, fully offline, with nothing sent to Google. No API key is needed and no subscription either. And there is no government that can shut off access. For a large share of the work organizations actually do, that is enough, and no special hardware is needed.
The wake-up call
It is tempting to feel relieved when Fable switches back on and GPT-5.6 becomes more widely available. Enjoy it, but stay critical. The events of June show what the world looks like now, and that situation does not disappear when the models return.
As far as we are concerned, draw a few sober conclusions from it. Do not build your work so that it depends entirely on one American API that can be shut off from outside. Use the best American model where it truly adds value, while also making sure you can run open models yourself, so you have a fallback option that nobody can revoke. Keep sensitive data within your own walls. And when buying, weigh who owns the model and who can turn it off, not just what it costs and how highly it scores.
For Europe, the message is even simpler. It is time to wake up. The dependence is real, it is double, and it cannot be regulated away with a fine speech about sovereignty. It calls for computing power on home soil, for European providers that generate real revenue, and for the competence inside organizations to keep open models running. That is work that needs to be done now.
The models are impressive. That is not where the snake is hiding. The snake is in the question of who decides whether you may still use them tomorrow. Right now, that answer is not with us, and we should not simply accept that.
