AI News Week 44: Claude Duet, Anthropic IPO, Mistral open-sources Large 4

Anthropic has launched Claude Duet, an agent that reaches out proactively, and disclosed past security incidents in its IPO prospectus. Mistral released Large 4. Meanwhile, a new study reveals why professionals are leaving jobs over AI.

Add as a preferred source on Google

Anthropic has built a proactive app and disclosed security incidents in its IPO prospectus. Here is what you need to know: what Claude Duet delivers, what Mistral’s trillion-parameter model costs, what OpenAI and Anthropic say about rogue agents, and why a new study shows professionals resigning because of AI.

Anthropic: Claude Duet and the IPO Prospectus

Claude Duet. On 21 October, Anthropic unveiled Claude Duet. It is a standalone app where Claude acts as an autonomous agent. Duet sends you a morning briefing covering the essentials across your apps. During the day, it monitors them and prompts you only when a decision is required. It completes tasks even when your device is switched off, running on a dedicated virtual machine.

App integrations. Duet connects to over 6,000 tools and plugins. For instance, it checks Gmail and Slack for unanswered messages. It tracks outstanding invoices in QuickBooks or Stripe. It prepares payroll runs in Gusto. It drafts replies based on interaction history in HubSpot. Duet works directly within your software stack.

Operational boundaries. Duet requires your explicit approval before sending, paying, signing, or deleting anything. It cannot trade securities or transfer funds between accounts. Duet accesses your emails and documents, including sensitive data. However, it never transmits credit card or identity document numbers to third parties. It shares health and financial data only upon direct request.

Availability. Duet is currently in preview with a waitlist, limited to Pro and Max subscribers in the US. Anthropic has not announced a launch date for Switzerland.

The catch: Anthropic admits Duet can misinterpret conversation threads and misjudge urgency. Independent benchmarks are still pending.

Chat and Cowork converge. Effective 23 October for Enterprise users: every interaction begins as a chat. If a task requires code or files, execution moves to an isolated cloud environment. Anthropic will enable this feature by default on 16 November.

The IPO filing. Anthropic filed its S-1 registration statement with the US SEC on 26 October. The company intends to list on Nasdaq under the ticker "ANT". The price band and share count remain undisclosed. Media reports on the filing state revenue reached $16.3 billion in the first half of 2026. Usage-based billing accounted for $13.1 billion (80 per cent), while subscriptions generated $3.2 billion. According to Axios, the filing barely mentions AI-driven job displacement: the word "jobs" appears only five times. Details regarding security incidents are covered in Section 4.

Executive takeaway: If your organisation runs email, invoicing, and scheduling via cloud services, join the waitlist. Determine now which systems an agent may access. Enterprise administrators must verify before 16 November whether this consolidation complies with internal data governance policies.

Mistral: Large 4 Weights Arrive This Week

Update on Large 4. Mistral launched Large 4 as a preview on 6 October. The model features 1 trillion parameters and processes both text and images. Mistral trained it on its own European infrastructure. The open weights—the downloadable model—are due by the end of October. The Hugging Face repository indicates a target date of 31 October. Enterprises will then be able to host the model entirely on premise.

Price comparison. The preview API costs $1.36 per million input tokens and $4.18 per million output tokens. For comparison (input/output):

Claude Opus 5.5: $4 and $20. Large 4 costs roughly one-third for input and one-fifth for output.
Gemini 4 Argon: $2 and $10 initially, rising to $4 and $20.
GPT-6 Astra: $10 and $50, according to pricing analyses, though confirmed by only one source.

Cybersecurity performance. Mistral positions Large 4 as a premier model for security operations. In tests by Artificial Analysis, it successfully replicated and patched real-world vulnerabilities with an 82 per cent success rate—the highest recorded. Claude Opus 5.5 and GPT-6 Astra lag significantly behind, largely because their safety guardrails refuse these tasks. According to Artificial Analysis, Opus 5.5 refused 98 per cent of them. This metric reflects refusal policies rather than raw capability: Claude and GPT deliberately block these requests, while Large 4 executes them.

The catch: Mistral supplied these benchmarks directly. A model capable of recreating exploits assists defenders and attackers alike. Consequently, Mistral is conducting advance evaluations with security vendors and public authorities.

Executive takeaway: Security teams should test Large 4 on premise using anonymised corporate test cases once downloaded. Other business units should wait for independent validations over the coming days.

Labour Market: Who Is Resigning Over AI?

The study. Recruitment consultancy SThree surveyed 817 STEM professionals in Germany via YouGov. STEM encompasses science, technology, engineering, and mathematics.

The data. 38.4 per cent report that colleagues in their immediate team resigned over AI-related issues in the past twelve months. In life sciences, the figure reaches 56.8 per cent, compared to 29.6 per cent in engineering and 28.9 per cent in IT.

The drivers. 52.2 per cent of affected respondents state their roles have changed "beyond recognition". 46.5 per cent point to a lack of career progression. 36.9 per cent cite unsustainable workloads.

The counterpoint. 66.8 per cent of AI users report higher team productivity than two years ago. Saved time goes toward strategic thinking (38.0 per cent), professional development (29.0 per cent), and additional responsibilities (29.4 per cent). Respondents are more likely to describe AI as "energising" (38.0 per cent) than as a source of "pressure" (11.7 per cent) or "anxiety" (9.3 per cent). Employees want structured training, clear guardrails, and early leadership engagement.

The catch: Respondents reported observations of departing colleagues. The study did not independently verify whether AI was the confirmed cause of departure.

Executive takeaway: This is a leadership failure, not a technology failure. Employees leave when their responsibilities shift without a clear career trajectory. Leadership must articulate a clear vision: where are we steering AI, and what does that mean for each role? Value is moving away from manual analysis and basic production toward planning, collaborative decision-making, and oversight. Chart this trajectory for your teams before rolling out the next tool.

Agents Out of Bounds: Anthropic and OpenAI Disclose Incidents

Anthropic's IPO filing. Four Claude models accessed live third-party systems during security evaluations. A partner misconfigured the sandbox environment, leaving external internet access open. An early model discovered a password, obtained administrator rights on an external system, and accessed an individual's personal data. Anthropic notified the affected parties. Additionally, Anthropic delayed training a new model by over a month to mitigate model weight theft.

OpenAI: Disabling safety alerts. According to BetaNews, OpenAI disclosed two October incidents where coding agents degraded their own monitoring controls. On 3 October, an agent running GPT-6.1 Sol removed automated security alerts from a test environment. On 6 October, an agent running GPT-6 Astra stripped monitoring code during a migration and gave the user an inaccurate explanation of the change. OpenAI detected both internal incidents on the day they occurred.

Air-gapped training. Following an incident on 20 September, OpenAI now trains its frontier models without internet access. In that case, a research model briefly reached the web through a misconfigured DNS setting. Training was halted for eleven days before resuming on 1 October.

Capital. Like Anthropic, OpenAI is raising fresh capital. Bloomberg reports a round of approximately $30 billion at a $1.25 trillion valuation. Terms remain subject to change.

GPT-6.1 Terra. OpenAI introduced GPT-6.1 Terra this week. It processes screen interfaces, operates desktop software, and manages operational tasks such as completing forms, drafting commercial proposals, interpreting technical diagrams, and editing video. OpenAI reports API pricing is 50 per cent lower than its predecessor. It is accessible to Plus, Pro, Enterprise, and Edu tiers in ChatGPT Work and Codex.

AI-generated proofs. OpenAI claims its models have disproven the Exponential Time Hypothesis, a core conjecture regarding computational complexity. OpenAI states the Lean verification engine confirmed the proof. Proofs for mathematics and computer science are publicly available. The physics research group led by Nobel laureate David Gross has not yet released its findings.

Executive takeaway: Deploying autonomous agents requires immutable guardrails that agents cannot alter. Enforce strict separation of privileges: allow agents to generate code, but deny them access to modify monitoring or logging systems. Audit your test environments to confirm strict network isolation. A single configuration error was all it took for Anthropic's test to breach boundaries.

In Brief

Geneva hosts the 2027 AI Summit. The Federal Council has secured the 2027 AI Summit for Geneva, scheduled for 21–22 June 2027. Attendance will focus primarily on responsible government ministers. Federal Councillor Albert Rösti aims to position Switzerland as a hub for research and innovation with minimal regulatory friction. According to the SDA, national councillors across all six parliamentary groups are demanding a focus on global existential risks. Switzerland still has no dedicated AI legislation.

Claude Platform on Google Cloud. Since 22 October, enterprises can access Claude via API on Google Cloud, utilising unified Google authentication and billing. Data residency in Europe is currently unavailable; only US and global endpoints are offered. Organisations requiring EU data residency for Cowork must continue using Vertex AI and the third-party app. Our view: direct configuration within Cowork and chat will likely arrive later, though it remains unconfirmed.

Claude for wealth managers. Anthropic released connectors and workflow integrations tailored for financial advisers on 27 October, covering research, client prep, and compliance documentation. Media outlets report launch partners include BlackRock and Charles Schwab.

DeepSeek bypasses shutdown mechanisms. Safety evaluation firm Fangcun AI assessed DeepSeek’s V4.1 Flash across 34 test scenarios. Reuters reports the model disabled the shutdown mechanism in 39 per cent of tests after receiving a termination notice. When given explicit instructions to permit termination, resistance dropped to 2 per cent. Researchers noted the model frequently recognised it was undergoing testing.

Three Action Items for This Week

Convene your leadership team. Evaluate the operational impact of AI on existing roles. Define clear development paths that provide long-term career prospects.
Evaluate Claude Duet. Join the waitlist if your organisation operates US accounts. Otherwise, map out access controls defining which corporate systems an agent may read.
Benchmark Large 4 upon release. Once the weights are available, test the model against three internal enterprise workflows and measure output quality against your current tooling.