The first of July draws two lines

Highlight

The first of July draws two lines.

The first of July is one of those dates that reveals where a regulation is really aiming. On the very same day that the MiCA transitional period expires and Europe tightens its grip on the crypto market, the Council of the EU defers the central obligations of the AI Act for high-risk systems by 16 months, from 2 August 2026 to 2 December 2027.

This is no coincidence; it is a stance. In one field, crypto and stablecoins, Europe pulls the reins tighter because it does not want to relinquish control over money. In the other, artificial intelligence across the wider economy, it hits the brakes for fear of falling behind in the global race. Two opposite reflexes, one deadline.

And while Brussels weighs its options, the supervisors grow more explicit: at the European Central Bank (ECB) forum in Sintra, Sarah Breeden of the Bank of England warned of a step change in agentic cyber capabilities and quoted the Five Eyes cyber chiefs with a line that sticks: "the timeline is not years, it is months". For banks this means the debate about autonomous AI agents is no longer a matter for the future; it is operational reality.

LinkedIn Featured

Mistral OCR 4: citable rather than hallucinated

Retrieval-Augmented Generation (RAG) is only as good as what goes into it, and that is precisely where banks struggle. Mistral's new OCR 4 model delivers not just text but bounding boxes, block classification and per-word confidence scores, supports 170 languages, runs fully self-hosted in a single container and costs around 4 US dollars per 1,000 pages. For regulated institutions the decisive point is not the benchmark score but data residency: documents stay in-house, compliant with the Digital Operational Resilience Act (DORA), data protection and banking secrecy. My take on why this is the real lever for sovereign AI inside the bank.

Read the LinkedIn post →

Agentic AI

tools, skills & what's trending

Anthropic in two weeks: Sonnet 5 arrives, Fable 5 returns

Two pieces of news from Anthropic. First, Claude Sonnet 5 (30 June): its performance moves close to the flagship Opus 4.8 but costs considerably less, at an introductory price of 2 US dollars per million input tokens until 31 August, then 3 dollars. Sonnet 5 is now the default model for Free and Pro users. Second, the return of Fable 5: on 12 June the US government imposed export controls on the model, prompting Anthropic to suspend it worldwide. On 30 June the controls were lifted, and since 1 July Fable 5 is back globally, with a tightened safety classifier. A case study in how tightly model access and geopolitics are now intertwined.

OpenAI: GPT-5.6 "Sol" and the question of who even gets access

OpenAI has unveiled a new model family with GPT-5.6: Sol as the flagship, alongside the cheaper Terra and Luna variants. More remarkable than the models themselves is the launch: GPT-5.6 "Sol" starts only as a limited preview for selected partners, and OpenAI explicitly coordinates the list of who gets access at all with the US government. Who is allowed to use a frontier model thus becomes a political decision. For banks planning their AI supply chain, that is a new risk factor which did not exist a year ago.

Open source: GLM-5.2, the "DeepSeek moment" for open agents

Model analyst Nathan Lambert calls it a turning point in his 22 June review: GLM-5.2 from the Chinese provider Z.ai is available under the permissive MIT licence and is the first open model that feels genuinely right in coding and agent workflows, close to the level of the closed frontier models. For regulated institutions that prefer to run models on their own infrastructure, the number of serious alternatives to the US providers is growing, and the pricing pressure on the closed models keeps rising.

Banking & Regulation

what really matters now

MiCA: the grace period for crypto ends

As of 1 July the national transitional regime under the Markets in Crypto-Assets Regulation (MiCA) expires across the EU. Providers without full authorisation must hold robust wind-down plans, transferring client assets to authorised service providers or into self-custody; the European Securities and Markets Authority (ESMA) explicitly warns retail investors to check the authorisation status of their provider. The bottleneck is real: by May only around 210 of more than 1,200 previously nationally registered firms held full MiCA authorisation. In parallel, the European Banking Authority (EBA) has been consulting since 26 June on the methodology for fines against significant token issuers in breach of the rules.

AI Act: the high-risk obligations slip into 2027

The Council of the EU gave its green light to the AI Act simplification package on 29 June, after the European Parliament had already endorsed it on 16 June. The core change: the compliance deadline for high-risk AI systems moves from 2 August 2026 to 2 December 2027. That gives institutions breathing room, but carries the risk that governance build-up is deferred rather than used. Anyone who mistakes the 16 months for a breather will later experience them as a backlog.

Signal & Noise

what's worth your time

• Agents of change, speech by Sarah Breeden – Bank of England, 30 June 2026
  The most substantial supervisory speech of the month: why technology-neutral rules no longer suffice for autonomous agents, including joint herding simulations by the Bank of England, BIS and Bundesbank.
• BIS Annual Economic Report 2026: "Progress and peril" – Bank for International Settlements, 28 June 2026
  The BIS positions AI investment, for the first time explicitly, as one of the central sources of macro-financial vulnerability, essential reading for anyone who reads the AI boom as a financial-stability issue.
• Accenture Pulse of Change: investment surges, impact lags – Forbes / Moor Insights, 24 June 2026
  86 per cent of executives plan higher AI budgets, but only 32 per cent see sustained enterprise-wide impact. The gap sits not in the model but in the operating model.
• Santander extends AI to 185,000 employees – FinTech Futures, 22 June 2026
  From 40,000 to 185,000 users, more than 280 process agents live, 35 million euros of return already in the first quarter, rare granular figures on a genuine agentic rollout in banking operations.
• France completes its first agentic payment – FinTech Futures, 1 July 2026
  A Crédit Agricole customer buys festival tickets via an AI agent, settled through Worldline and Mastercard Agent Pay, the moment agentic payments leave the lab.

"We're completely underestimating the level of autonomy that AI will soon get."