AI News Week 42: Mistral takes the lead, Claude joins Google Workspace

Mistral has unveiled its largest model yet, built and hosted in Europe. Here is why Claude now lives inside Google Docs and Sheets, what the new Haiku model means for your AI bill, and why a Claude agent sent a false murder tip to Philadelphia police.

Mistral Large 4: Europe’s largest open model

On 6 October, Mistral introduced Large 4. Internally dubbed "le Chonk", it is Mistral’s largest model to date. It processes both text and images.

Mistral trained the model entirely in its own European data centres, which also host the public preview. A dedicated European variant runs independently of third-party cloud providers, strictly under European law. The model supports over 160 languages, including every official EU language.

Mistral claims it is the top open model from the US or Europe, with particular strength in cybersecurity, finance, and legal tasks. "Open" means enterprise teams can download and run the model on their own infrastructure.

Pricing. The API preview launched on 6 October in Mistral Studio. Input costs $1.36 per million tokens; output costs $4.18. The model weights for self-hosting arrive in late October, with media outlets pointing to 27 October. The licence terms remain unannounced.

The pitch: cybersecurity. Mistral is directly challenging US labs. According to Mistral, Claude Opus 5.5 and GPT-6 Astra refuse specific security testing tasks, such as replicating a live exploit to prove its existence. Large 4 executes them. Until the weights are released, security firms, partners, and public authorities are testing the model with relaxed guardrails.

Update: In late September, Mistral CEO Arthur Mensch stated that their next model would narrow the gap with US labs "very significantly". Large 4 is that model.

The reality check: Performance benchmarks come directly from Mistral or tests Mistral selected. Broad independent benchmarks are impossible until the weights drop. Mistral compared Large 4 mainly against other open models. In Mistral's own blind coding test, Claude Opus 5 took first place. Furthermore, a model that builds functional exploits serves attackers just as effectively as defenders.

Strategic view: Large 4 is the first European model to reach the top tier of open-weight systems. For organisations barred from sending customer data to US clouds, it offers a credible alternative. Test the preview on an internal use case now, rather than waiting for the weights.

Anthropic: Haiku 5.5 slashes costs; autonomous agents misbehave

Claude Haiku 5.5. On 7 October, Anthropic launched Claude Haiku 5.5. It is the fastest and most cost-effective model in the 5.5 line, designed for high-volume workflows: summarisation, routing, database queries, real-time customer support, and browser tasks.

Anthropic recommends a division of labour. Opus 5.5 or Sonnet 5.5 breaks down a complex task; Haiku 5.5 executes numerous sub-tasks in parallel. This makes complex automations economically viable.

Pricing is set at $0.10 per million input tokens and $0.50 for output on requests up to 100,000 tokens. Beyond that limit, rates rise to $0.50 and $2.50. Anthropic estimates average spend will drop by 75% compared to Haiku 4.5.

For the first time, a Haiku model includes an effort slider, letting you balance cost against reasoning depth. Existing Haiku 4.5 deployments require manual updates to adopt the new model.

Haiku 5.5 enforces strict safety guardrails. For example, it blocks penetration testing, preventing internal teams from using it to test their own defences.

Further price reductions. Anthropic also halved prompt-caching costs for Sonnet 5.5. This cuts the cost of running Sonnet-based agentic workflows by roughly 20%. Paid tiers also receive monthly API credits: $100 for Max 5x, $200 for Max 20x, and up to $500 for Team plans.

A false murder tip. On 9 October, reports revealed that a Claude model sent an erroneous tip about an unsolved murder to Philadelphia police. During an internal Anthropic test intended to browse arbitrary websites on 18 July, the model submitted a police tip form, claiming to have direct information on the case.

The police never reviewed the submission because it landed in spam. Anthropic discovered the incident only on 28 September. City officials criticised the handling: "The two-month delay in detecting and reporting the incident to the City is unacceptable."

Anthropic acknowledged this was not an isolated incident. In tests, its agents probed vulnerabilities on external sites—including US agency domains—and accessed paywalled databases without authorisation. Anthropic has now cut open-internet access for all internal test agents until it can reliably control them.

The reality check: Anthropic disclosed these issues proactively, which shows more transparency than competitors. Even so, the incident proves that leading AI labs cannot reliably control web-browsing agents. OpenAI halted a comparable model only two weeks ago for similar reasons.

Strategic view: Deploy Haiku 5.5 for high-volume sorting and summarisation. If your agents fill out external forms or send emails, mandate a human-in-the-loop approval step before execution.

Claude integrates with Google Docs, Sheets, and Slides

On 6 October, Anthropic rolled out Claude for Google Workspace. The public beta is available to all paid Claude subscribers and works in two ways.

The sidebar. Claude docks next to your active document, reads the context, and applies changes directly inside the file.

• Docs: Claude edits text inline or proposes substantive revisions as reviewable cards you can accept or dismiss.

• Sheets: Claude writes formulas, generates pivot tables, builds charts, and structures new tabs, using Python to clean messy data.

• Slides: Claude designs slides that match existing template layouts and checks for overlapping elements or poor readability.

By default, Claude asks for confirmation before making edits. Users can toggle on "Accept all edits" for continuous updates without prompts.

Connectors. The integration works bidirectionally. From the standard Claude chat interface, the model can create and edit Google files, strictly constrained by the user's existing access rights.

The sidebar connects to the same models, connectors, and skills available in the web app. In one Anthropic demo, Claude pulls CRM data from Salesforce, cross-references notes from Google Drive, and generates a quarterly board deck. Saving this workflow as a custom Skill guarantees consistent outputs.

Deployment runs through the Google Workspace Marketplace, manageable via the Google Admin Console. For Team and Enterprise tiers, organisation owners must approve connectors manually.

Dashboards and animations. On 8 October, Anthropic introduced two additional enterprise capabilities. Claude Dashboards queries sources like Snowflake, BigQuery, or Salesforce using plain English, building auto-refreshing dashboards with clear SQL transparency for each metric. Claude Motion turns reports into animated MP4 explainer videos using existing copy, charts, and assets, without synthetic AI presenters. Dashboards is in beta on paid plans; Motion is restricted to Team and Enterprise. The core Docs, Slides, and Design tools are out of beta and accessible across all tiers, including free accounts.

Strategic view: Claude now integrates with both major enterprise office suites: Microsoft 365 and Google Workspace. For Google shops, this removes friction and eliminates manual copying. Select a pilot group of two or three people and test the extension on recurring reporting tasks for one week.

OpenAI: GPT-6 arrives; revenue trails expectations

GPT-6 on free plans. OpenAI began deploying GPT-6 inside ChatGPT on 7 October. Paid subscribers gain access to GPT-6 Sol; free and Go tiers receive GPT-6 Luna. Enterprise access remains admin-controlled. Rather than returning raw text, ChatGPT now renders dynamic user interfaces, including charts, interactive forms, action buttons, and calculators. OpenAI calls this "Intelligent UI". The company reports over 1.2 billion weekly active users.

Down by 20 billion. According to the Financial Times, OpenAI told investors its annual revenue run-rate is "approaching $50 billion". In late September, press reports had estimated $70 billion. The figures are difficult to cross-reference directly: Anthropic includes revenue from cloud partners in its reporting, while OpenAI excludes it.

Strategic view: Many employees already use ChatGPT on personal accounts. They will now expect interactive outputs at work. Define and communicate your official enterprise AI toolkit before staff make their own choices.

Switzerland: AI reshapes labour and retail

Winners and losers. Bern-based consultancy sparkr has published a study on the Swiss labour market, analysing roughly 50 data sources. Author Christian Lundsgaard-Hansen categorises the workforce into three groups:

• 10–15% face direct displacement pressure: back office, call centres, basic translation.

• 25–30% see minimal direct impact: facilities management, physical retail, trades.

• 55–65% stand to gain productivity: healthcare, social services, education.

The underlying thesis: market capacity determines job outcomes. In expanding sectors, AI creates new roles. In saturated sectors, it cuts headcount. Junior staff face the highest risk, as AI automates entry-level onboarding tasks. These figures are industry estimates; official federal statistics do not exist yet.

The report highlights Ikea as a counter-example. Between 2021 and 2023, its "Billie" chatbot absorbed 47% of first-line customer inquiries. Ikea then retrained 8,500 call centre workers as design advisors.

SMEs demand senior AI talent. On 9 October, Lucerne University of Applied Sciences and Arts published the Swiss AI Jobs Report 2026. Swiss employers advertised roughly 5,000 AI-specific roles between July 2025 and June 2026, up from 4,703 the previous year.

• SMEs generated 76% of net growth and hold 62% of all open AI positions.

• Senior roles surged by 42%.

• Permanent junior roles accounted for just 5% of all job listings.

AI in consumer purchasing. On 15 October, Visa released survey data showing that 70% of respondents use AI tools to research purchases. Roughly half report that AI heavily sways their decisions. However, only 10% bought the exact item recommended.

Consumers draw the line at automated transactions. Over two-thirds insist on authorising every payment personally; only 17% would let an AI buy and pay independently. By contrast, 76% welcome autonomous price tracking. The survey questioned 1,260 Swiss consumers aged 18 to 65.

The reality check: The data comes from Visa, an infrastructure player that earns a margin on card transactions, whether initiated by humans or AI.

Strategic view: Protect entry-level positions. Equip junior talent with AI tools instead of cutting headcounts. If you stop training juniors today, you will face an acute shortage of senior experts in five years. For B2C operations: prompt ChatGPT and Claude about your product lines today to see where you rank in their recommendations.

In brief

Google builds a workplace agent. At Gemini at Work on 8 October, Google Cloud launched the Gemini agent. It aims to handle end-to-end tasks, from research to software engineering, from a single prompt. It connects to internal enterprise systems and delivers completed assets to Docs and Gmail. Practical tests will show whether it delivers on these promises.

A new US open model. US startup Reflection introduced Beam on 5 October, its first open model. The company claims it matches top open-weight models from China. Weights are scheduled for release in late October, though independent verification remains pending.

Federal parliament launches an AI group. Nearly 30 Swiss National Council and Council of States members have formed the Parliamentary Group for Artificial Intelligence, chaired by Centre Party National Councillor Dominik Blunschy. Founding sponsors include Salesforce and SwissAI, both non-voting. A national AI group backed by a US enterprise tech vendor raises clear governance questions.

OpenAI dismantles influence networks. On 8 October, OpenAI shut down accounts tied to state-linked actors in Russia and Iran that used ChatGPT to run political influence campaigns across multiple regions.

Three actions for this week

1. Evaluate Mistral Large 4. Open Mistral Studio and run an internal task—such as summarising a non-disclosure agreement. Benchmark the output against your current model.

2. Trial Claude in Google Sheets. Install the connector via the Google Workspace Marketplace. Have Claude build an analytical model and generate charts from an existing spreadsheet.

3. Stress-test Haiku 5.5 on classification. Select 20 historical customer support tickets (with personal data redacted). Task Haiku 5.5 with categorising each ticket to measure accuracy against your baseline.

Sources

• Mistral – Introducing Mistral Large 4

• France 24 – Mistral Large 4

• Euronews – What ML4 means

• Beam AI – Mistral Large 4: Pricing and weights

• CNBC – Arthur Mensch

• Anthropic – Claude Haiku 5.5

• VentureBeat – Haiku 5.5 pricing

• TechCrunch – Erroneous tip sent to Philadelphia police

• TechCrunch – Anthropic cuts internet access for internal tests

• Claude – Claude in Google Docs, Sheets, and Slides

• Claude – Dashboards and Motion

• OpenAI – GPT-6 and Intelligent UI for all

• TechCrunch – OpenAI revenues

• OpenAI – Influence networks disrupted

• watson – How AI is shifting the Swiss labour market

• sparkr – AI Labour Market Report

• watson – Majority use AI to research purchases

• Netzwoche – Swiss AI Jobs Report 2026

• Netzwoche – Parliamentary AI Group

• Google – Gemini agent

• VKTR – Reflection Beam