AI News, Week 21 – Anthropic focuses on SMEs and legal, Google reworks Android, and AI finds a zero-day vulnerability for the first time

1. Anthropic: Claude for Small Business, Gates Foundation and Claude for Legal

Anthropic had a busy week – three separate launches, each for a different audience.

Claude for Small Business is a preconfigured package of connectors and workflows that integrates Claude directly into the tools SMEs use every day. The key integrations:

• QuickBooks: Payroll planning, month-end close, cash flow analysis and tax prep directly from chat

• PayPal: Billing, invoicing, disputes and refunds without switching systems

• HubSpot: Lead triage, customer feedback analysis and campaign attribution

• Canva: AI-assisted creation of marketing materials

• DocuSign: Contract workflows, approval processes, document review

• Google Workspace & Microsoft 365: Email drafts, calendar management, document editing

The package can be activated with a toggle install – no IT department needed. Anthropic is launching a free roadshow format in several US cities: half-day workshops for local SME owners with one month of Claude Max included.

Assessment: This skills bundle is a solid foundation. For Swiss SMEs, it can be adapted directly – to specific industries, compliance requirements or internal workflows. That is exactly where we help: we work with SMEs to build their own AI skill set – not a ready-made framework imposed from above, but capabilities the team can continue to develop itself. Those who start today will have a structural advantage in six months over competitors still using copy-and-paste AI.

Gates Foundation Partnership: The Bill & Melinda Gates Foundation is one of the world's largest philanthropic foundations. Its focus is global health – fighting malaria, tuberculosis and polio in countries that cannot afford Western healthcare infrastructure – as well as education and economic participation in developing regions. The foundation has assets of over USD 50 billion and co-funds projects where public health systems are missing or overstretched.

Anthropic and the Gates Foundation are investing USD 200 million over four years in AI-supported diagnostics in regions without specialists, education programmes and economic mobility. The Gates Foundation only deploys capital where measurable impact is created.

Separately: PwC is rolling Claude Code and Cowork out across its entire global workforce – several hundred thousand employees. The largest enterprise rollout of a frontier model to date.

Claude for Legal (12 May): 12 practical plugins for legal teams – from corporate legal and M&A diligence to privacy legal and IP legal. Each plugin learns the team's playbooks, escalation paths and house style via a setup interview. Over 20 MCP connectors for DocuSign, Box, Thomson Reuters (Westlaw via CoCounsel), Harvey, Relativity and Microsoft 365. Legal was the most used profession in Claude Cowork – three times more than the second-largest function.

2. Google: Gemini Intelligence and I/O today

On 12 May, Google showed the 'Android Show: I/O Edition' – a pre-announcement before the I/O keynote (today, 19 May, 19:00 Swiss time, live on YouTube).

Gemini Intelligence is an agentic layer under Android that completes tasks across multiple apps on its own. Demos: take a photo of an event flyer → Gemini searches Expedia and books it. Shopping list on screen → Gemini fills the basket in the required app. Further features: 'Magic Pointer', AI-generated widgets, Gboard dictation with automatic clean-up.

Available first on Samsung Galaxy and Pixel (from summer), then Wear OS, Android Auto and Android XR. It only runs in apps that the user has explicitly approved.

Caution: anyone who activates the system gives Gemini access to app content, calendar and communications. Opt-out details for company devices have not yet been communicated.

Today's I/O: Gemini 4.0 announcement expected, Android XR glasses (preview), Veo upgrade.

3. OpenAI: consulting firm at the expense of traditional advisory – and a cyber model for Europe

OpenAI Deployment Co. launched on 11 May. Licensing models is no longer enough if companies do not know how to use AI. The firm helps build and operate AI systems – with OpenAI staff and over USD 4 billion of initial capital.

That makes it a direct competitor to classic management consultancies. The difference: traditional consultancies optimise for billable hours – the incentive to deliver value fast is structurally weaker. OpenAI Deployment Co. has a direct interest in systems that work.

At the same time: full dependence on one vendor gives that vendor control. If prices rise, performance drops or compliance requirements shift – for example EU/Swiss data residency – there is no easy exit. A modular setup is more robust in the long run.

GPT-5.5-Cyber: a version of GPT-5.5 specialised in cybersecurity that identifies vulnerabilities, analyses attack vectors and supports threat hunting. Since 30 April in the US, since 13 May EU access for vetted security teams from companies, public authorities and EU institutions (including the EU AI Office).

Anthropic's Mythos – according to the UK AISI at a comparable cyber benchmark level – is not accessible in Europe. OpenAI therefore has a significant first-mover advantage in European security organisations.

4. Open Source: Four Chinese labs in twelve days – and why the combination matters

Within twelve days, four Chinese labs released open-weight coding models: Z.ai GLM-5.1, MiniMax M2.7, Kimi K2.6 (Moonshot AI), DeepSeek V4. Plus Qwen 3.6 from Alibaba. All can be deployed locally, with operating costs below a third of Claude Opus 4.7.

Kimi K2.6 Thinking: 78.57 LiveBench, 58.33 Agentic Coding. Qwen 3.6 27B: 77.2% SWE-Bench Verified.

The interesting question is not 'open source or cloud', but when each makes sense. For companies that must keep data on site – compliance, banking secrecy, health data – these models on their own infrastructure are a serious option. For tasks without high reasoning depth, they are economically worthwhile.

Claude is the right choice where depth of judgment matters: complex analyses, multi-step agentic workflows. Those who route 60–70% of traffic through inexpensive open-source models and use Claude selectively where the quality gap is measurable cut costs sharply.

5. AI and cybersecurity: the first zero-day discovered by AI

On 13 May, Google's Threat Intelligence Group (TIG) confirmed: an attacker group used AI to discover an unknown zero-day vulnerability and weaponise it for a mass attack. Google stopped the attack.

Zero-day vulnerabilities – flaws with no patch – used to be the most valuable asset in cybercrime. Finding them required highly specialised researchers and a lot of time. Frontier models accelerate this process dramatically. The fact that this has now been documented changes the threat landscape fundamentally.

That explains why OpenAI and Anthropic are pushing ahead on cybersecurity models at the same time. GPT-5.5-Cyber has been accessible in the EU since 13 May. Anthropic's Mythos remains locked – Anthropic says the risks of uncontrolled access are too high. The UK AISI confirms that Mythos matches GPT-5.5-Cyber on cyber benchmarks.

For businesses: AI-enabled attacks are coming faster than expected. The best defensive tools are not yet available everywhere.

3 things worth doing this week

1. Identify the top three use cases in your own business – Which three tasks repeat every week, take time and do not need deep judgement? Those are the first candidates for automation – and the right starting point for checking where open-source models are enough and where a frontier model like Claude makes the difference.

2. Weigh open source against frontier models – For tasks with confidential data or low requirements: test Kimi K2.6 or Qwen 3.6 locally. For complex decisions and multi-step workflows: use Claude. The combination lowers costs and at the same time increases control over your own data infrastructure.

3. Watch the Google I/O keynote today – 19:00 Swiss time, live on YouTube. A Gemini 4.0 announcement is expected. io.google