AI News Weekly 30/31 – Opus 3.5 at half price, Apertus 1.5 as Swiss response, and shared-Claude chat data leak

Double issue: CW 29 was cancelled due to the summer break. This edition covers two weeks.

1. Claude Opus 5 – Peak Intelligence at Half the Price

Anthropic released Claude Opus 5 on 24 July. The core message is simple: Opus 5 delivers intelligence close to the top-tier Fable 5, but costs half as much. Key pricing remains matching its predecessor, Opus 4.8 – $5 per million input tokens, $25 per million output tokens. In short: more performance for the same price.

Opus 5 is now the default model on Claude Max and the strongest model on Claude Pro. Built for daily use, it works more efficiently, requiring fewer computational steps and less time for the same task. An effort slider lets you choose between more intelligence and faster speed – depending on whether the task demands precision or pace.

The benchmarks are less interesting than the practical impact on office work. Anthropic shared data from real-world deployments:

• Zapier used Opus 5 to run a complete customer retention workflow autonomously – identifying at-risk clients, alerting the right account manager, and drafting a comprehensive briefing. Previous models failed; Opus 5 completed the task end-to-end.

• Box reports an 8% improvement in accuracy over Opus 4.8. This rises to 11% for data analysis and 17% for contract and document reviews.

• In financial modelling, accuracy rose by 9 percentage points, with one-third fewer steps and 60% less time required.

• When auditing and correcting contracts, Opus 5 performed almost twice as well as Opus 4.8.

• Presentation creation (tested using the Gamma tool) saw the clearest leap: better formatting and fewer errors on slides.

Two footnotes. Anthropic states that Opus 5 is its most controlled model yet, with the lowest rate of hallucinations and safe-policy violations. There is also a "Fast Mode" that runs roughly 2.5 times faster for double the base price.

Our take: The price cut is the real news here. Previously, those wanting top-tier quality had to pay premium rates for Fable 5. Opus 5 offers near-identical quality at half the price, whilst becoming the new standard. For most office tasks – research, writing, analysis, editing – it is fully sufficient. Fable 5 is now only worth using for the most complex edge cases.

Which model for what? Anthropic now has five tiers. Here is your quick guide:

• Mythos 5: The absolute peak. Built for niche fields like cybersecurity and biology. Rarely needed for daily business tasks.

• Fable 5: The most powerful all-rounder, but slow and expensive. Reserve this solely for the most complex reasoning tasks.

• Opus 5: The new daily standard. Near-Fable performance at half the price. Default on Pro and Max, ideal for knowledge work.

• Sonnet 5: Fast and cheap. A solid model for basic tasks if you frequently hit your usage limits on higher tiers.

• Haiku: The fastest, cheapest model. Best for simple, high-speed routing tasks.

Rule of thumb: Set Opus 5 as your default on Pro and Max. Switch to Sonnet or Haiku for basic tasks, and only use Fable 5 for exceptional cases.

2. Claude Goes Mobile – And Gains a Voice

Two new updates make Claude much more practical for daily use.

Cowork launches on Web and Mobile on 3 August. Cowork is Anthropic’s autonomous agent mode where Claude completes workflows on its own – researching, drafting documents, and organising files. Previously restricted to the desktop app, it will run in browsers and on mobile from 3 August because sessions run in Anthropic's cloud. You can switch devices mid-run, and scheduled tasks will continue in the background even if your device is offline. No admin setup is required; the feature activates automatically on 3 August. Double usage limits apply until 5 August.

A stat from the announcement blog is telling: over 90% of Cowork usage is not for software development. The largest sectors are business operations and content creation, accounting for roughly half of all workloads. This dispels the myth that AI agents are only useful for programmers.

Advanced Voice comes of age. Claude's voice feature previously only ran on its smallest model, Haiku. It now runs on Opus and Sonnet, allowing you to switch models mid-conversation. The update is available on iOS, Android, Desktop, and Web in eleven languages, including German. Crucially: in voice mode, Claude can access connected tools – Gmail, Google Calendar, Slack – to take action. You can dictate an email, and Claude will save it directly as a draft.

Our take: Anthropic does not lead on pure conversational feel. The mode is turn-based – Claude waits until you finish speaking – and sounds less natural than OpenAI's or Google's live voice modes. The value lies elsewhere: only with Claude does a spoken conversation result directly in a completed task, like a draft email in your inbox. Voice as an interface, not just a chat.

3. Claude Now Takes Action in Microsoft 365

Previously, Claude's Microsoft 365 integration was read-only. It can now execute actions directly inside corporate accounts.

• Outlook: Send emails, manage drafts, labels, filters, and out-of-office replies; create, edit, or delete calendar invites.

• SharePoint: Create and update files.

• Read and search capabilities remain unchanged.

Deployment note: For enterprise accounts where the integration was installed before 7 June, write permissions are disabled by default and must be enabled manually. Newer installations have them active. Microsoft global admins must approve these new permissions once.

Our take: "Writing in Microsoft 365" means Claude executes actions inside the account. It does not mean Claude will automatically write Word documents, Excel sheets, or PowerPoints – that is a different feature. The practical value is the elimination of manual steps: Claude reviews your inbox and draft replies directly, rather than just giving you text to copy and paste.

4. European Sovereignty: Apertus 1.5 and Flux 3

Two European models prove that relying on US tech giants is not the only option.

Apertus 1.5 – the Swiss model goes multimodal. ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5 on 24 July. The latest version of the fully open Swiss language model can now process images – reading and analysing documents, charts, or technical drawings (with audio capabilities still in experimental stage). Key additions include an optional reasoning mode (solving steps internally before responding), a four-fold context window (handling much longer documents), and better instruction-following and tool use. It was trained on the "Alps" supercomputer at CSCS in Lugano. "Fully open" here means more than other open-source models: not just weights, but training data, code, and values are fully auditable. This is the core of the sovereignty argument – organizations can audit the training data and run the model on-premise. The models are available on Hugging Face, with a technical paper and benchmarks arriving in the coming weeks.

Flux 3 from Germany. Black Forest Labs, a German venture, unveiled Flux 3 on 23 July. The model trains on image, video, and audio simultaneously. It can generate video with natively synced audio (up to 20 seconds), convert text or images to video, handle keyframe transitions, and stitch multiple clips into cohesive scenes. A robotics offshoot is currently being trialled by Audi.

A note on performance: Black Forest Labs' internal data shows Flux 3 beating rivals like Luma, Runway, or Kling with a preference score of up to 93%. These are self-reported tests and have not yet been independently verified.

Our take: For Swiss enterprises, Apertus is the standout news. Whilst it does not match the peak performance of Opus 5 or Fable 5, it holds its own against open models like Google’s Gemma 4. It easily covers 60% to 70% of standard office tasks like writing, summarising, and basic research. It is the logical choice where data sovereignty, transparency, and independence from US hyperscalers are paramount. This is what digital sovereignty actually looks like.

5. Security Leak: Shared Claude Chats Exposed on Google

An embarrassing security oversight at Anthropic was exposed over the weekend. Chats shared publicly via Claude's "Share Link" feature were indexed and searchable on Google, Bing, and Yandex. A Reddit user demonstrated on 25 July that a simple search query could surface pages of private user chats.

While only shared chats were exposed – not private account history – the leaked contents were highly sensitive: legal notes, medical advice, internal financial models, source code containing API keys, CVs, and even crypto keys. A GitHub archive collected 453 Claude sessions and 519 Grok chats containing over 11,000 cleartext messages – Grok suffered from the exact same issue.

The cause was a classic configuration error. Anthropic had blocked search engines via a robots.txt file, which stops bots from crawling. However, this does not prevent a URL from being indexed if it is linked elsewhere. Fixing this requires a "noindex" meta tag – which was missing. Anthropic added the tag on Monday, and Google removed the indexed results starting 26 July. Team and Enterprise accounts were unaffected.

Our take: The fault lies squarely with Anthropic. However, the lesson for enterprises is fundamental: a shared link to an AI chat is public. Treat it as such. Review your active shared sessions in privacy settings, and establish strict policies on what belongs in an AI prompt: no passwords, no commercial secrets, and no credentials.

6. In Brief

Competitors remained highly active. OpenAI launched the GPT-5.6 family (Sol, Terra, Luna) into general availability, making it the new default for ChatGPT. They also launched GPT-Live-1 as the default voice engine with real-time translation. Google released Gemini 3.6 Flash on 21 July, cutting prices. Meta launched Muse Spark 1.1, its first paid agentic coding assistant. xAI opened public access to Grok 4.5 as a low-cost coding model.

Google flags Gemini 4. During its earnings call on 22 July, Alphabet CEO Sundar Pichai confirmed Google is already training Gemini 4. This is their most ambitious training run yet, utilising a significantly larger base model. Pichai candidly admitted Google needs to catch up on agentic coding. The launch is expected in November or December; Google will release monthly incremental updates to its Flash models in the meantime. Our take: Google continues to project a strong future vision, while Anthropic and OpenAI are shipping today. Notably, Gemini 3.5 Pro missed its June release window.

Mistral enters robotics. French AI firm Mistral demonstrated a compact model that guides robots through tasks via spoken commands and a single camera feed – marking Mistral's first step into embodied AI.

OpenAI targets SMBs. OpenAI launched a dedicated SMB enablement programme featuring training webinars and on-site academies, alongside "ChatGPT Work" powered by GPT-5.6. Anthropic is matching this with Cowork and deeper Slack integrations. The land grab for the daily desktop is fully underway.

Three Things to Focus on This Week

1. Switch your default to Opus 5. On Pro and Max accounts, you now get near-Fable-5 performance at half the computational cost. Opus 5 is the natural default; save Fable 5 strictly for your most complex logic tasks.

2. Prepare for Cowork on mobile. From 3 August, Claude's agent workflows run in the browser and on mobile. Leverage the double limits until 5 August to test background agent jobs.

3. Evaluate Apertus. View this not as a direct competitor to US frontier models, but as an open, sovereign Swiss alternative for regulated industries where data residency and transparency are non-negotiable.