AI News Week 18 – Anthropic admits error, OpenAI speeds up, Gemini becomes Siri

This week, four topics dominated: Anthropic admits a self-inflicted drop in quality in Claude Code and publishes the postmortem. OpenAI follows up with GPT-5.5 just six weeks after the last model. Google confirms: Gemini replaces OpenAI as the Siri engine – and has already launched a native Mac app. Elon Musk is quietly forging an alliance against Anthropic and OpenAI.

1. Anthropic: Self-criticism instead of silence

For weeks, developers complained that Claude Code had become worse. Anthropic said nothing – until 23 April. Then the company published a public postmortem and named three specific errors:

First: On 4 March, Anthropic reduced the default reasoning effort from "high" to "medium" – to cut latency. Quality suffered. Rolled back on 7 April.

Second: A caching bug from 26 March caused the model to keep discarding its own reasoning trace. Claude seemed forgetful and unreliable. Fix: 10 April.

Third: On 16 April, a system prompt was introduced that limited answers to 25 words between tool calls. Rolled back four days later.

All three errors were fixed by 20 April. On 23 April, Anthropic reset the usage limits of all subscribers. The company announced stricter internal tests, an improved code review tool and a new @ClaudeDevs account for product communications.

Assessment: This is a rare example of genuine corporate transparency in the AI industry. At the same time, the case shows how fragile quality control is in fast-moving labs. Anyone using Claude Code in production should follow release changelogs more closely – or at least plan for a downgrade scenario.

2. OpenAI: GPT-5.5 – six weeks, new model

On 23 April, OpenAI launched GPT-5.5, internally called "Spud". The model arrived just six weeks after GPT-5.4 – a pace that has made the industry take notice.

GPT-5.5 works at a much higher level of intelligence at the same token throughput. OpenAI describes it as its most intuitive model so far: it understands tasks faster, needs fewer tokens for Codex tasks and works independently across multiple tools until the job is done. Context figures: ChatGPT has over 900 million weekly users, 50 million subscribers and 9 million paying businesses. 4 million Codex users are active.

GPT-5.5 and GPT-5.5 Pro have been available via the API since 24 April.

Assessment: Six weeks between frontier models is not a sign of technical maturity – it is a competitive instrument. For teams without dedicated AI expertise, evaluating models becomes a permanent task. Anyone relying on fixed model versions must build in evaluation cycles.

3. Google: Gemini on the Mac – and soon on the iPhone

Gemini for Mac (15 April, not in the last newsletter): Google has launched a native macOS app for Gemini. With Option+Space, Gemini can be called up from anywhere on the Mac. The app lets users share the current screen content directly – for example, to have a chart analysed. Available from macOS 15, free to use.

Gemini becomes Siri (22 April): Google Cloud chief Thomas Kurian confirmed at Google Cloud Next 2026 in Las Vegas: Gemini replaces OpenAI as the AI backbone of Siri. The rollout happens in two phases:

• Phase 1 (active now): iOS 26.4 already uses Gemini for context-aware understanding and on-screen recognition.

• Phase 2 (September 2026): With iOS 27 and iPhone 18 comes «Full Conversational Siri» – fully based on Gemini.

Apple keeps the privacy infrastructure: processing on Apple devices and Private Cloud Compute. In addition, Google Cloud announced a $750 million fund for the partner ecosystem and agentic AI.

Assessment: Apple has made a strategic choice – better the best third-party AI than a mediocre in-house one. For companies using Apple devices, Siri will be significantly more capable by September. Privacy questions remain relevant, especially for regulated industries.

4. xAI / Mistral / Cursor: Musk forges an alliance

Business Insider reported on 22 April: xAI, the French AI lab Mistral and the coding tool Cursor are in talks about a three-way partnership – with the aim of directly challenging Anthropic and OpenAI.

Background: Mistral co-founder Devendra Chaplot moved to xAI in March and now leads pre-training there. Cursor is already training models on xAI's Colossus supercomputer in Memphis. SpaceX has secured a purchase option on Cursor worth $60 billion – if it does not exercise the option, SpaceX pays Cursor $10 billion for joint development work.

None of the three companies has officially confirmed the talks.

Assessment: Mistral is the only European alternative of consequence to US models. A link with xAI's infrastructure and Cursor's developer tooling would be strategically significant – especially for organisations that prioritise AI sovereignty. At the same time: unconfirmed Musk projects tend to leave a wide gap between announcement and reality. Wait and see.

5. Switzerland: AI use clearly above the EU average

The Federal Statistical Office has published new figures: 47 per cent of the Swiss population aged 16 to 74 use generative AI – third place in Europe, behind Norway (56%) and Denmark (48%). The EU average is 33 per cent.

73 per cent of AI users also use the technology at work – no other European country reaches this figure. Among 16- to 24-year-olds in Switzerland, 79 per cent use generative AI (EU: 64%). The survey is based on around 3,000 respondents in spring 2025.

Assessment: Switzerland is further ahead with AI adoption than many think. High usage without competence building leads to uncritical dependence. With 53 per cent non-users, there is also significant potential – and corresponding advisory demand for companies.

Three recommendations this week

1. Claude Code users: Since 20 April, performance is back to normal. Anyone who had problems over the last few weeks should re-evaluate now – and follow release changelogs more closely in future.

2. Test Gemini for Mac: The native app, with a global shortcut (Option+Space) and screen sharing, is a serious everyday assistant. Free to use – worth a short test.

3. Apple devices in the company: The Gemini-Siri integration is coming in two phases. Now is the right time to clarify with Legal and IT which data may go through Siri – before Phase 2 goes live in September 2026.