Washington Locks Europe Out, Frankfurt Rethinks Money

Highlight

Availability is the new dependency

On 12 June, the US government ordered Anthropic to block worldwide access to Claude Fable 5 and Mythos 5. Overnight, the strongest model on the market was unreachable – not because of a technical outage, but because of a geopolitical decision taken in a foreign capital. The real lesson for Europe's banks is not what such a model can do, but that its availability sits under foreign state power.

Just how real that is becomes clear from a sober look at the leaderboards: Artificial Analysis ranks Fable 5 number one – yet it is currently blocked globally. As close as fourth place, just behind the closed front-runners, sits an open-weights model you can run yourself: GLM-5.2. The frontier you cannot switch on is, in case of doubt, worth less than the good-enough model you fully control.

For regulated institutions, this is not a technology question but a procurement and resilience one. A multi-vendor strategy, exit readiness and an honest answer to "what keeps running if a provider goes dark tomorrow?" belong on the agenda now – not after the next shock. That is precisely what this issue's lead story is about.

LinkedIn Featured

Frontier AI under State Power: Why Europe's Banks Now Need AI Sovereignty

On 12 June 2026, the US government ordered Anthropic to block worldwide access to Claude Fable 5 and Mythos 5. The risk for Europe's banks is not the capability, but the availability under foreign state power. A commentary on why AI sovereignty and a multi-vendor strategy now belong on the agenda – and what that means in concrete terms for procurement, architecture and exit readiness.

Read the full article: Frontier AI under State Power →

Agentic AI

tools, skills & what's trending

Anthropic launches Fable 5 – and walks back the invisible guardrails

On 9 June, Anthropic released Claude Fable 5, the first publicly available Mythos-class model. The system card contained a silent anti-distillation guardrail: requests the model judged to be training attempts for competing models were rerouted without any notice. After the backlash, Anthropic apologised, called it the "wrong trade-off" and now makes refusals visible. On top of that come a mandatory 30-day retention of all inference traffic on Mythos models (safety monitoring, explicitly not training) and a dedicated data-sharing mode on AWS Bedrock – before access was blocked globally on 12 June owing to US export controls anyway.

OpenAI builds the enterprise flywheel: BBVA, Partner Network, Ona

Three moves in a single week. BBVA is rolling out ChatGPT Enterprise to more than 100,000 employees (target: all 120,000 across 25 countries, around three hours of time saved per week) – one of the largest GenAI deployments in financial services. In parallel, the OpenAI Partner Network launches with 150 million US dollars and launch partners such as McKinsey, BCG, Accenture and PwC (target: 300,000 certified AI consultants by year-end). And with the acquisition of Ona (formerly Gitpod), Codex gains persistent, sandboxed cloud environments for long-running agents.

GLM-5.2 leads the open models – and that is the point

Zhipu's GLM-5.2 is, according to Artificial Analysis, the new strongest open-weights model (Intelligence Index 51, fourth place overall). Above it sit only three closed models – including Claude Fable 5 at number one, which is currently unavailable globally because of export controls. This is exactly where the strategic punchline lies: an open model you can run yourself beats, in case of doubt, the best model you cannot reliably access. For regulated institutions, "self-hostable" is becoming a hard selection criterion.

Banking & Regulation

what actually matters now

ECB: 25 basis points up – and a plan for the money of tomorrow

The ECB Governing Council raised key interest rates by 25 basis points on 11 June, taking the deposit rate to 2.25 per cent (May inflation: 3.2 per cent, driven by the Middle East conflict). Four days later, Christine Lagarde sketched the strategic line at the "Money in transition" conference: tokenisation against market fragmentation (projects Pontes and Appia), the digital euro as an answer to international card schemes holding more than 60 per cent market share, and a clear warning against "digital dollarisation" through USD stablecoins. Monetary policy and money infrastructure are converging.

Germany pays predominantly cashless for the first time

The Bundesbank study "Payment behaviour in Germany 2025" marks a turning point: for the first time since the study began in 2008, everyday purchases were settled predominantly cashless – 55 per cent by card or smartphone, 45 per cent in cash (in 2023, cash still stood at 51 per cent). The girocard dominates, mobile payments are growing, and PayPal leads online. For strategy, this means the cash narrative no longer holds – investment in instant payments, wallet integration and the digital euro is no longer optional.

Signal & Noise

what your time is worth

• Frank Elderson: "Deregulation is a no-go" – European Central Bank
  The supervisor draws the line: simplification yes (securitisation approvals cut from three months to seven days, around 100 publications to be scrapped), lowering capital requirements no. Essential reading to separate modernisation rhetoric from genuine deregulation.
• ECB sets out the roadmap for the Integrated Reporting Framework – European Central Bank
  Harmonised reporting for all euro-area banks, first mandatory submission in Q2 2031, built on EU cloud solutions. Anyone planning reporting architecture should align with the milestones now.
• Apple will not roll out Siri AI in the EU – after a rejected DMA exemption request – Reuters
  The European Commission counters that this is "Apple's and Apple's only". A case study in platform control versus regulation – and another data point on the question of which AI features are available in Europe at all.
• DiffusionGemma: 4x faster text generation with open weights – Google DeepMind
  Diffusion-based rather than autoregressive decoding, 1,000+ tokens per second on a single H100, running quantised on an RTX 4090. The open-weights front keeps delivering arguments for self-hosting.
• An AI agent bankrupts its operator while scanning DN42 – Hacker News
  An autonomous agent burns through the budget unsupervised. The entertaining version of a serious lesson: agents need hard cost and permission limits before they run in production.

Personal Note

currently in IT-Finanzmagazin

Claude Fable 5 vs. Mythos – ein Modell, zwei Klassen: Anthropic rationiert KI nach Vertrauen, und Banken bleiben außen vor

Anthropic has made Fable 5 public, while the stronger Mythos 5 variant remains reserved for a partner circle (project Glasswing) – and for sensitive cyber queries, Fable silently reroutes to the older Opus 4.8. My piece for IT-Finanzmagazin (11 June 2026) shows why the advertised peak performance is effectively unreachable for regular users – and why the gap between advertised and actually available capability becomes a procurement risk for banks. (Article in German)

Read the article in IT-Finanzmagazin →

"Everything fails, all the time."