AI News Week 23 – Anthropic overtakes OpenAI, Mistral targets industrial AI, EU deadlines collapse

1. Anthropic: Claude Opus 4.8, 965 Billion Valuation, Mythos Soon for All

Anthropic made three moves in a single week, shifting the balance of power in the AI market.

Claude Opus 4.8 has been live since 28 May. Anthropic describes the model as having "sharper judgment, more honesty about its own progress, and the ability to work autonomously longer than its predecessors". In benchmarks: agentic coding rises from 64.3% to 69.2%, multidisciplinary reasoning with tools from 54.7% to 57.9%, and agentic computer use from 82.8% to 83.4%. The knowledge work score jumps from 1753 to 1890. Early testers report that Opus 4.8 is roughly four times less likely to quietly let errors slip through in its own code. Pricing remains identical to 4.7. New in claude.ai: users control how much effort Claude should put into a task. Claude Code gets "dynamic workflows" for very large problems. Fast Mode runs at 2.5x speed and costs three times less than predecessors. (Anthropic News)

I have tested Claude 4.8 over the last few days and am very satisfied. What you should remember: before every chat, consider how complex the task actually is – and then choose the appropriate mode between low, medium, high, extra and, in absolute exceptional cases, max. A quick email does not need high mode. A strategic concept or a difficult analysis will benefit greatly from it. Max remains for those few cases where it really matters.

965 Billion Valuation. On the same day, Anthropic closed a $65 billion Series H, led by Altimeter, Dragoneer, Greenoaks, and Sequoia. Co-leads: Capital Group, Coatue, D1, GIC, ICONIQ, XN. Participants: Baillie Gifford, Blackstone, Fidelity, General Catalyst, Insight, Jane Street, Lightspeed, MGX, T. Rowe Price, Temasek. $15 billion of this consists of already committed investments from hyperscalers, including $5 billion from Amazon. Valuation after the round: $965 billion. This sees Anthropic overtake OpenAI for the first time. Annualised, Anthropic currently earns around $47 billion – more than an eightfold increase since the February round. (Anthropic Series H)

Claude Mythos is being made widely available. In April, Anthropic classified the model as too dangerous for public release – it finds software vulnerabilities autonomously better than practically any human. Following Project Glasswing (over 10,000 highly critical vulnerabilities identified in 30 days, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw), Anthropic is preparing the broad rollout. During Mozilla testing, Mythos found 271 distinct vulnerabilities in Firefox. Anthropic estimates 6–12 months until Mythos-class models are widely available.

Analysis: The valuation makes for a good story, but the real point lies elsewhere. Firstly, whilst OpenAI lost $1.22 per dollar of revenue in Q1 2026 (see CW22), Anthropic's numbers are significantly better – more revenue per customer, less loss per customer. Secondly, with Opus 4.8 and the concrete Mythos pipeline, Anthropic has the clearest path to useful agents – not just the biggest demo. Thirdly, anyone watching the past year will recognise the pattern. OpenAI launches agentic features more loudly; Anthropic simply builds them in. The ChatGPT experience and model benchmarks are two different topics.

2. Mistral: From Model Provider to Industrial Supplier

On 28 May, Mistral held its first proprietary conference, the AI Now Summit in Paris. Four announcements together define their strategy.

Mistral for Industrial Engineering is an AI solution specifically for heavy industry – the model understands physical contexts and can handle real engineering and simulation data. Launch customers: Airbus, BMW, EDF, CMA CGM. The Airbus deal runs over five years, covering civil aviation, helicopters, and space. It includes on-board AI systems for aircraft, automated technical documentation, faster design cycles, and edge AI for flight safety. BMW uses Mistral to accelerate crash simulations – trained on BMW's archive of crash data. (Bloomberg)

Emmi AI Acquisition. In May, Mistral acquired Emmi AI for around €300 million, bringing on board over 30 researchers specialising in physics-based AI models.

Proprietary Data Centre. In Les Ulis (south of Paris), Mistral is building a 10 MW facility dedicated solely to serving AI queries. Opening Q3 2026. This ends Mistral's dependency on US hyperscalers.

Vibe is the new name for Le Chat. Vibe is a single agent for multi-step tasks – integrating inbox and calendar, researching, drafting texts, orchestrating recurring processes, and executing coding tasks from brief to merged pull request. Available as a web app, editor, and terminal.

Mistral now employs 1,000 people and targets €1 billion in revenue for 2026. (VentureBeat)

Analysis: I have found Mistral highly impressive for some time. In my own tests, it has beaten ChatGPT 4.x repeatedly – particularly in precise data analysis. Mistral is a serious European option that fits directly into the sovereign stack we are building: data stays in Europe, the model comes from Europe, and the new data centre provides a robust infrastructure behind it. For Swiss industrial and engineering companies that want to keep their design data out of the US, Mistral is becoming the first choice.

3. EU AI Act: 16 Months More Time, a Few Harsh New Bans

The EU adjusted the AI Act for the first time on 7 May 2026. Three points you should know.

More time for most obligations. The strict rules for so-called high-risk AI will take effect later. High-risk refers specifically to AI in HR (cv screening, recruitment decisions), AI for credit decisions and insurance, AI in education (exams, admissions), AI in law enforcement, and AI in migration. Instead of August 2026, the deadline is now December 2027 – giving businesses roughly 16 months more time. For medical devices, lifts, and radio equipment, the deadline is extended by another year to August 2028. (Council of the EU)

Hard stop on deepfakes. Banned from 2 December 2026: AI that generates or manipulates nude images or sexual content of a person without consent. Likewise banned: AI that generates child abuse material. This is the first area where the AI Act becomes highly concrete, very quickly.

Clearer rules in consultation. On 19 May, the EU Commission published drafts describing more precisely what qualifies as high-risk. Businesses and associations have until 23 June to comment.

Analysis: For Swiss companies, this means concretely: those selling into the EU or working with EU customer data have more breathing space regarding HR tools, credit, or insurance AI. However, anyone offering or embedding image generation must implement safeguarding mechanisms within six months – meaning detection of real people, blocklists for sensitive content, and clear consent flows.

4. Switzerland: AI Creates More Jobs Than It Cuts – For Now

The EY Study from 28 May 2026 surveyed 604 individuals from Swiss companies. Key figures:

18% report that their company has created additional roles related to AI – such as in data science or AI engineering. 11% say open positions are no longer being filled because of AI. 7% are experiencing direct job cuts through AI automation. 42% cannot make a clear assessment yet.

Main hurdles to scaling: data quality and data silos (20%), security and data privacy (19%), and specialist skills shortage (18%). (Netzwoche)

Analysis: These figures align with the UBS study from CW22: 97% of companies have AI initiatives, but only 5% have a clean data foundation. The EY data now shows the other side – we are still seeing net job creation, but 42% do not know what to make of it. This is the real signal: Switzerland is not in an AI job crisis. It is in an AI orientation phase. The strategic question for 2026/27 is not "when will jobs disappear", but "what new roles does AI create in my business, and who is building them".

5. Rapid Fire: Apple, Meta, xAI, OpenAI

Apple WWDC starts on 8 June. Reports are hardening: Siri will run on Google Gemini, hosted on Apple's Private Cloud Compute. Siri replaces Spotlight; every search becomes a Gemini query. Additionally, a new "Siri" camera mode (5th mode after photo, video, portrait, panorama) is coming. The domain genai.apple.com has been registered. This is Apple's biggest Siri overhaul in 15 years – and an admission that Apple Intelligence did not work in its previous form.

Microsoft Build runs 2–3 June. Expected: General availability of Azure AI Foundry, a new framework for autonomous agents across Microsoft 365, Azure, and Windows, and the next generation of the GitHub Copilot coding agent. Windows Local AI gets its own track. We will report in detail in CW24.

Meta is testing two new subscriptions from June: Meta One Plus ($7.99/month) and Meta One Premium ($19.99/month). Premium delivers more compute power, deeper reasoning, and expanded image and video generation. Launching in Singapore, Guatemala, and Bolivia. Meanwhile, WhatsApp gets an "Incognito" mode for Meta AI – temporary conversations that, according to Meta, even their own engineers cannot read.

xAI had a productive May: Grok 4.3 (1 million token context, native video inputs), Imagine Agent Mode, connectors for SharePoint, Outlook, Drive, Workspace, Notion, GitHub, Linear, and the new Grok Build 0.1 as a pure coding model with 256k context and custom skills.

OpenAI published a playbook on 29 May for trusted third-party evaluations alongside its own governance framework. On the same day, OpenAI introduced "self-improving tax agents with Codex". Outlook email and calendar apps now support shared mailboxes in ChatGPT. Additionally, Gartner positions OpenAI as a leader in enterprise coding agents. According to Sam Altman, the company's IPO is still not ready for filing internally, even though September 2026 remains the target.

Three Things to Note This Week

1. Play with Claude 4.8 and compare the modes. Take a typical task of your own – a briefing, an analysis, a code snippet – and run it in low, medium, and high. Review the differences. Use Max mode with caution: it burns through tokens quickly. It is only needed in absolute exceptional cases.

2. Spend 20 minutes trying Mistral. Take your favourite prompt and send it to both ChatGPT (or whatever you currently use) and Mistral via chat.mistral.ai. Compare the responses. You will be surprised at how competitive the European option has become – particularly for data analysis.

3. Identify what is blocking effective AI adoption in your business. The EY study highlights data silos as the main hurdle. Is this the case for you too? What data resides in which systems? What is preventing integration – technology, ownership, or politics? Use your AI for exactly this: explain the situation and spend ten minutes brainstorming potential first steps. You do not need an external consultant for the initial thinking.