AI News Week 26: Google loses talent, MCP goes enterprise, EU AI Act finalised

1. Talent war: Google bleeds, Anthropic and OpenAI win

Within one week, Google has lost two of its most prominent AI researchers – to the very two competitors it is fighting.

John Jumper joins Anthropic. Jumper is Vice President at Google DeepMind and won the 2024 Nobel Prize in Chemistry for AlphaFold, the AI system for predicting protein structures. On Thursday, 19 June, he announced his move on X to Anthropic – after nearly nine years at DeepMind. He plans to take a break first, then start. His exact role remains open. The move aligns with Anthropic's expansion into life sciences and computational biology.

Noam Shazeer joins OpenAI. One day earlier, on Wednesday, 18 June, Shazeer announced his departure. He was Vice President of Engineering at Google and co-lead of the Gemini models. Crucially, he is co-author of the 2017 paper "Attention Is All You Need" – the work that introduced the transformer architecture on which virtually every large language model is built today. Google had brought him back from Character.AI less than two years ago for around $2.7 billion. At OpenAI, he will become "Lead for AI Architecture Research", heading the core design of future models.

Analysis: Google has the top model and the top researchers – but it cannot keep them. Both departures went to the two pure-play AI labs working on frontier models and an IPO. Talent goes where the most exciting work and the greatest upside are. For Google, this is a double signal: it is losing ground to Anthropic and OpenAI not just in rollouts (see Section 6), but also in retaining key personnel.

2. Fable 5: still offline a week later

Last week, I reported how the US government shut down Anthropic's strongest public model via an export control order. The status this week: nothing has moved.

The order came on 12 June from Commerce Secretary Howard Lutnick. Since then, Claude Fable 5 and the more powerful Mythos 5 have been offline globally. Anthropic engineers have been in Washington for the first in-person talks with the Commerce Department since 12 June. As of 18 June: no reactivation date, no formal lift of the order. A circulating rumour that access would return "within 48 hours" did not originate from Anthropic and is unconfirmed.

A spicy detail: Fable 5 was supposed to be included free in subscriptions until 22 June, and only for credits from 23 June. This deadline is now expiring – whilst the model is unavailable to anyone.

Analysis: This confirms the lesson from week 25. A model was the best on the market for three days, then completely gone for ten days and counting. Anyone who tightly coupled a process to it is still left without a replacement. Never link your operations to a single model.

3. MCP goes enterprise: centrally managed connectors

On 18 June, Anthropic introduced an update that affects daily operations more than many a model launch: Enterprise-Managed Authorisation for MCP connectors.

Previously, every employee had to individually authorise each tool they wanted to connect to Claude – Jira, Figma, the CRM. Now, IT administrators can approve connectors once centrally. Upon their first login, employees automatically gain access to the tools designated for their group, without any setup – across Claude Chat, Claude Code, and Cowork. This approval follows the groups defined in the identity system.

The first supported provider is Okta. At launch, seven connectors are available: Asana, Atlassian (Jira, Confluence, Rovo), Canva, Figma, Granola, Linear, and Supabase. Slack will follow. The setup is based on an open extension of the Model Context Protocol – meaning it also works with custom-built connectors and identically for every Claude customer. It is available as a beta for Team and Enterprise plans.

Analysis: This removes the biggest hurdle when rolling out AI agents in a company – individual tool linking by every single employee. For SMEs, this means onboarding and governance lie with IT, not with every person at their desk. Limitation: still in beta, and initially only with Okta as the identity provider.

4. EU AI Act: the "AI Omnibus" is passed

The European Parliament gave its final approval to the "AI Omnibus" on 16 June. The package amends key deadlines and rules of the EU AI Act.

Obligations for high-risk AI are postponed. This refers to AI that makes decisions about people – software that filters job applications, checks creditworthiness, or identifies individuals biometrically. For such standalone systems, the obligations now apply from 2 December 2027. If the AI is built into a regulated product as a safety component – such as a medical device, a machine, or a car – the deadline is 2 August 2028.

The labelling requirement begins on 2 December 2026. From then on, AI-generated content must be marked as such in a machine-readable format.

Two new bans also take effect on 2 December 2026: the creation of non-consensual intimate depictions ("nudifiers") and child sexual abuse material.

For Switzerland: There is no Swiss AI law modelled on the EU. Instead, the Federal Council intends to ratify the Council of Europe's AI Convention. However, Swiss companies doing business in the EU are bound by the EU deadlines.

Analysis: For most companies, the Omnibus provides breathing space – the expensive high-risk obligations have slipped back by over a year. Conversely, the labelling requirement in December 2026 is concrete and close. Anyone publishing AI content should plan for machine-readable marking now. Civil rights organisations criticise the package as deregulation that weakens protective rights – an issue that will continue in the political debate.

5. Europe wants its own AI model – and already has one that nobody knows

On 19 June, the EU Commission designated the "EUROPA" consortium. Its mission: to develop an open AI model for the continent, independent of the US and China.

The catch: an open European model has existed for a long time. Teuken 7B, built in the German research project OpenGPT-X (Fraunhofer, DFKI, Forschungszentrum Jülich, TU Dresden). Trained in all 24 official EU languages, over half the data non-English, freely available on Hugging Face. Sounds good – but has two problems. Firstly, performance: with 7 billion parameters, Teuken plays in the league of Meta's Llama 3.1 8B and Mistral 7B – entry-level models, far from Claude, GPT, or GLM. Secondly, and more seriously: hardly anyone knows it exists.

Analysis: This is Europe's AI dilemma in one image. The Commission announces it will build something that already exists on a smaller scale – whilst the existing model model goes unused due to lack of performance and visibility. Funding and announcements replace neither computing power nor reach. A European alternative is not created by another consortium, but by a model that people actually use.

6. In brief

GLM-5.2: the open model comes closer. On 17 June, Z.ai (formerly Zhipu) released GLM-5.2 under an MIT licence with open weights – the same day Fable 5 and Mythos 5 were blocked for foreigners. It is a so-called Mixture-of-Experts model: instead of using the entire model for every query, it only activates the relevant sub-networks. This means 744 billion parameters (the stored knowledge) are available, but only 40 billion compute per query. The result: the knowledge of a massive model at the speed and price of a small one. On SWE-bench Pro, a test for real-world programming tasks, it scores 62.1 points – ahead of GPT-5.5 (58.6), making it the strongest open model; on long tasks, it is just behind Opus 4.8. This is Plan B from Section 2 in its purest form: no government can turn off an open model on your own hardware. Note: this applies to the open weights. Anyone using GLM via the China API introduces a data privacy issue.

GPT-5.6 – expected but unconfirmed. Several sources expect OpenAI's next model between 22 and 28 June; on Polymarket, the probability is around 83 to 89 per cent. Chief Scientist Jakub Pachocki internally calls it a "significant improvement". Circulating specs (1.5 million token context, aggressive pricing below Claude) stem from developer leaks – OpenAI has confirmed nothing officially as of 22 June. Treat this as an outlook, not a fact.

Google, the king of announcements. Gemini Notebooks – Google's response to Projects in Claude, ChatGPT, and Mistral – are finally available in Europe and Switzerland. But only in the free version. Paying Google AI subscribers in the EEA and Switzerland were skipped during the rollout and must wait. Those who pay are effectively penalised. Furthermore, Gemini 3.5 Pro, announced at I/O, remains in limited Vertex preview as of 19 June. OpenAI and Anthropic launch globally on the day of announcement. Google announces – and delivers months later, in stages, with gaps.

xAI leads in video. Grok Imagine Video 1.5 has been generally available since 16 June and leads the Image-to-Video-Arena ranking. Price: $4.20 per minute in 720p – compared to $30 per minute for Sora 2 Pro, making it around 86 per cent cheaper. The model turns a still image plus text prompt into clips of up to 15 seconds, with sound generated in the same pass.

Three things to watch this week

1. Plan for the labelling requirement. From 2 December 2026, AI-generated content in the EU must be marked in a machine-readable format. If you publish such content, clarify now how your tools implement this.

2. Maintain a Plan B. Fable 5 is still gone after more than ten days. If a process depends on a single model, set up and test a second provider.

3. Keep an eye on Enterprise-Managed Auth. If you roll out AI agents in your company, central, IT-managed tool approval will greatly simplify onboarding – initially with Okta, soon more broadly.