AI News Week 17 | What executives need to know now

Last week was unusually packed. Anthropic launched two new products in quick succession. OpenAI is embedding ChatGPT directly into Excel. Google is merging Gemini and NotebookLM. Mistral is releasing an open-source coding agent. And Switzerland leads the global AI talent ranking. Here is what matters - and where caution is needed.

Anthropic: New model, new design tool

Claude Opus 4.7 has been available since 16 April and is Anthropic's most powerful model:

• Solves three times more real-world software tasks than its predecessor

• Follows instructions more precisely, which matters for automated workflows

• Analyses higher-resolution images (3x more than before) - for document analysis, screenshot review, automated reporting

• Reasoning logic adapts automatically: simple requests return quickly, complex ones get more thinking time

Budget holders should note: A new tokenizer means the same amount of text uses up to 35% more tokens - at the same price per token. Anthropic has not communicated this as a price increase, but it is one. Anyone planning to use Opus 4.7 in production workflows should calculate the cost first. The criticisms.

Alongside this, Anthropic has launched Claude Design - available at claude.ai/design:

• Collaborative work on designs, prototypes, presentations and one-pagers directly with Claude

• During onboarding, Claude reads your own design files and builds a team design system from them

• Export to Canva, PDF, PPTX or HTML is possible

For marketing teams that need to produce visual communication quickly, this is a real workflow gain. The product comes from Anthropic Labs and is experimental - critical processes still need more testing time.

OpenAI: ChatGPT comes to Excel

GPT-5.4 is OpenAI's most powerful model right now. The more relevant company announcement: ChatGPT is being embedded directly into Excel and Google Sheets - in beta. Anyone who works daily with financial models, budgets or reporting can now build, analyse and update spreadsheets using natural language.

GPT-5.4 outperforms specialists in the GDPval benchmark (real knowledge work) in 83% of comparisons. It makes 33% fewer errors on single claims than GPT-5.2 - relevant when the model is used for presentations, analyses or customer communication.

The GPT-5.4 Thinking variant shows a thinking plan before answering. That sounds technical, but it is practical: you can correct the approach before the model runs - which saves iteration loops on complex tasks.

Also reported: The Sora video API can now generate videos of up to 20 seconds in 1080p, with reusable characters. For content teams that want to scale video production, that opens up new automation paths - although quality and controllability on longer clips still need improvement. All current model updates from OpenAI.

Google: Gemini and NotebookLM grow together

Google has introduced Notebooks in Gemini. The Gemini app and NotebookLM now share the same workspace - no more manual file transfer between the tools. Anyone starting a project in Gemini can open it directly in NotebookLM and use features such as Audio Overviews or infographics there.

NotebookLM has also received an engine upgrade: the full 1 million token context window is now active in all plans, and multi-conversation capacity has been increased sixfold. For research, strategy or knowledge management processes with large volumes of documents, this is a real step up in quality.

New: chat personas can be defined. Gemini 3.1 Pro is available exclusively in NotebookLM for Pro and Ultra users.

One note: the growing merging of Google products makes the ecosystem more powerful, but ties users more closely to Google - a classic lock-in dynamic that should be considered in strategic tool decisions.

Mistral: Open-source coding agent from Europe

Devstral 2 has been released by Mistral - a coding agent that, according to benchmarks, is up to 7 times more cost-efficient than Claude Sonnet. The model is open source and can be self-hosted. Anyone who relies on on-premise solutions for regulatory or data protection reasons now has a serious option. API price: from $0.10 per million tokens.

There is also Mistral Vibe CLI, an agent that navigates and changes codebases using natural language. Also new: Mistral 3 with a large flagship model (675 billion parameters, Apache 2.0), which takes second place in the open-source ranking. Europe now has a competitive alternative to US models - important for anyone prioritising data sovereignty.

Switzerland & Europe: Talent, regulation, medical LLM

Switzerland tops the 2026 Stanford AI Index in AI talent density: 110.5 AI researchers per 100,000 inhabitants. The Swiss National AI Institute (ETH + EPFL) has signed a cooperation agreement with Stanford HAI.

The CHUV university hospital in Lausanne will begin pilot testing Meditron in May, a Swiss medical LLM for clinical decision support - starting in the emergency room.

EU AI Act: Full applicability takes effect from 2 August 2026. Open-source models under 10 billion parameters have easier compliance requirements; larger models do not. Anyone planning AI systems today should be aware of the current logging requirements for AI agents - they apply sooner than many think.

Three things worth doing this week

Test Claude Design - if your team regularly produces presentations, decks or one-pagers. The prototype character is known, but the workflow gain is real. claude.ai/design

ChatGPT in Excel beta - sign up if financial models or data analysis are part of daily work.

NotebookLM with the new 1M-token window - upload larger document sets and use the extended multi-turn capacity. This can significantly speed up research and strategy analysis.