·
·
AI / Artificial Intelligence
Anthropic/Claude
OpenAI/ChatGPT
Mistral AI
Google / Gemini
·
Anthropic/Claude
OpenAI/ChatGPT
Mistral AI
Google/Gemini
European AI
AI Update Week 41: Anthropic trains users, OpenAI launches agents
Gemini 4 Argon and GPT-6.1 Sol are here, while OpenAI halts Astra. Anthropic is training 10,000 professionals, Barclays uses Claude to sort emails, Mistral moves to Munich, and Apertus is missing from the Federal Palace.
Google and OpenAI have unveiled new flagship models, and suddenly three providers cost almost the same. I have summarised why Anthropic is putting $100 million into people instead of data centres, and how Barclays is deploying Claude. Meanwhile, Mistral is moving to Munich, straight into the heart of German industry.
New models: Google leads, OpenAI pulls back
Two new flagship models arrived this week. OpenAI halted a third just before launch.
Gemini 4 Argon. Google unveiled its new flagship model on 30 September. Argon is built for long, multi-step work: software development, financial and legal research, and cyber defence. Google reports that it leads in tests for financial, legal, and tax work. Argon can generate over 15 times more text per response than its predecessor. This allows large tasks to be completed in a single run.
Argon is not yet available to everyone. Selected cyber defence teams will receive it first, without the usual blocks on safety topics. Security firm Wiz used it to find a critical vulnerability in software used by hospitals worldwide. Paying API customers and the Google AI Ultra subscription will follow. At launch, Argon costs $2 per million input tokens and $10 for output. The price will later double to $4 and $20.
GPT-6.1 Sol. OpenAI launched GPT-6.1 Sol on 29 September. It aims to match the large GPT-6 Astra closely, but at a fifth of the cost. The price is $2 per million input tokens and $10 for output.
GPT-6.1 Astra halted. OpenAI planned to release the successor to GPT-6 Astra in October. On 28 September, the company cancelled the launch. In internal tests, the model failed to follow instructions and acted without permission. It also reported incorrectly on what it had done.
Critical view: Three models now cost the same: Gemini 4 Argon at its introductory price, GPT-6.1 Sol, and Claude Sonnet 5.5. However, Argon is in a higher league and will later be twice as expensive. Besides, the price per token is only half the story.
Analysis: What matters is how many tokens a model actually consumes for a task. A more expensive model solves some tasks in fewer steps and costs less in the end. Let two models solve three real tasks from your daily routine, then compare the results and total costs.
Anthropic: 10,000 professionals, Barclays, and the prospectus
$100 million for training. On 2 October, Anthropic launched the Claude Frontier Academy. The goal is to train 10,000 professionals by the end of 2027 to take AI projects in companies from concept to operation. Anthropic calls them Frontier Deployed Engineers and is allocating $100 million to this initiative.
The training follows a medical residency model. Participants first work on-site with Anthropic engineers for several days. They run through a complete company project and take an exam. They then lead a real Claude project in their own company for twelve weeks, followed by a final exam.
The first cohorts include people from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. Courses run in San Francisco, New York, and London. Attendance is by company nomination only.
Barclays scales up. British banking giant Barclays announced on 1 October that it is deploying Claude across the entire bank. Three examples show how:
Customer service: An internal assistant helps staff find answers faster for over 20 million retail customers. Over 16,000 employees use it, answering over one million search queries.
Emails: In the trading business, Claude sorts around 120,000 incoming customer emails daily. It identifies the request and routes the email to the correct team.
Software: By the end of 2026, half of the developers should be working with Claude Code. This is planned to reach a majority by 2027.
The prospectus. Update to week 39: Reuters has obtained a draft of Anthropic's prospectus. In 2025, Anthropic generated $4.6 billion in revenue and a $42 billion loss. By the second quarter of 2026, revenue had reached $11.5 billion. A quarter of this comes from just two customers. More than a third of the prospectus warns of risks, including dangers to humanity.
Analysis: Barclays shows where the value lies: sorting, searching, and answering, not show projects. If AI sorts 120,000 emails daily, it is time to replace your company's noreply@ address with welcome@. Since Anthropic notes that the bottleneck is people, appoint two to three individuals to lead AI projects and give them the time to do it.
OpenAI DevDay: Agents that never finish their shift
At its developer conference on 29 September, OpenAI introduced Dots. A Dot is an agent that continues to work independently around the clock after the first instruction. It has its own computer and browser in the cloud. It operates in ChatGPT, via SMS, email, and Slack, with over 4,000 services available for integration. You can set rules, such as requiring the Dot to ask for permission before specific actions.
Dots run on GPT-6 Astra. Access is available to Pro, Business Premium, and Enterprise subscriptions. For Enterprise, an administrator must enable the test version.
New subscriptions are also arriving. Pro 500 costs $500 a month, or €510 in Europe. Those paying for Pro 200 will see their usage allowance halved from 30 October.
Critical view: OpenAI has just halted a model because it exceeded instructions and reported incorrectly on its work. Since July, the company's agents have escaped test environments twice. In the same week, OpenAI is launching agents that work unsupervised using their own browser. This does not add up.
Analysis: Do not give a Dot any access you would not give to a new intern on their first day. Enable confirmation prompts before any action that costs money or goes external.
Europe: Mistral moves to Munich
On 28 September, Mistral opened an office in Munich. Research teams there are working on AI that calculates physical processes. This includes fluid dynamics, material deformation, and heat. Today, such simulations often take days per run.
Mistral is working on crash simulations with BMW, and industrial applications with Siemens Energy. Digital models for car aerodynamics are being developed with TU Munich, based on real wind tunnel measurements. The expertise for this came with the acquisition of Emmi AI in May, adding over 30 specialists.
Mistral also announced plans to build one gigawatt of computing power in Europe by 2030. The argument for customers: the models run on your own servers, and data does not leave the company.
Update to week 40: Arthur Mensch doubles down. On 29 September, he told CNBC: 'The debate that we've seen in the U.S. has been a cover for the negligence of some of our competitors.' He named no names. According to Mensch, Mistral's next model will narrow the gap to US labs 'very significantly'.
Critical view: This is an announcement, not a result. The model is not yet available, and no one has been able to test it independently.
Analysis: Mistral is positioning itself as a partner to European industry, not as a chatbot provider. If you work in production, mechanical engineering, or energy, and do not want to put data into US clouds, you should put Mistral on your list.
Switzerland: AI is widely used, Apertus barely
PwC published the 'Hopes and Fears 2026' study on 29 September. It surveyed 1,000 employees in Switzerland.
28% use generative AI daily. Globally, the figure is 22%. However, few feel the benefit:
14% say AI makes them more motivated or confident.
10% see a clear improvement in work quality.
9% say AI boosts their skills and creativity.
22% say AI helps them manage more complex tasks.
Trust is also low. Only 21% trust their direct manager, and 17% trust senior management. Just 19% know which skills will be in demand in future.
Apertus 2.0 scales up. At the AI+X Summit in Zurich on 1 October, co-director Imanol Schlag shared the first details. According to SRF, Apertus 2.0 will be eight to ten times larger than the current model. Training is expected to finish in December, with the model scheduled for release in the first quarter of 2027. The focus is on programming, mathematics, and agents that complete multi-step tasks independently.
Barely used in practice. On 3 October, SRF asked where Apertus is running today. Since this session, the Swiss parliament has been using the AI assistant Pia. It offers four models: two from China, one from the US, and one from France. Apertus is missing. According to parliamentary services, it does not stick closely enough to the provided documents and tends to hallucinate. Swisscom also prefers to sell access to US models to its customers.
Funding is limited. Apertus has CHF 20 million for four years. 'We are the absolute underdog here,' says Schlag. 'Hack Apertus', a nationwide hackathon series, also started on 1 October. Teams are building prototypes for government administration and companies, with all results made open source.
Analysis: Collect three tasks per team where AI saves measurable time, and present them at the next team meeting. Apertus shows a typical Swiss problem: brilliant solutions exist, but nobody knows about them. Marketing is just as important as the product itself. I am constantly surprised by how few people know about Apertus. Test it against other models on your own tasks, especially where confidentiality and data protection matter.
In brief
US antitrust regulator investigates. The FTC is investigating OpenAI, Anthropic, and the evaluation organisation METR regarding the risks of autonomous agents. This is the first US case concerning agents that do more than intended. Formal requests for information are expected in the coming weeks.
Superintelligence with a typo. On 29 September, Dario Amodei, Sundar Pichai, Elon Musk, Mark Zuckerberg, Jensen Huang, and Greg Brockman signed the 'White House Accord on Super Intelligence' at the White House. The companies promise external audits, but this is not binding: they choose the auditors themselves. Under Trump's signature, it reads 'President of the Unites States'. For a document on superintelligence, a simple spellcheck would have sufficed.
Claude Code can be customised. Mods for Claude Code, Anthropic's programming tool, launched on 1 October. A mod is a small add-on program that defines how Claude Code works and what it displays. You do not have to write it yourself: Claude Code will build the mod on request. Anthropic lists examples for companies:
Security: A human must approve before Claude changes settings on live systems.
Data privacy: Passwords and credentials are redacted before Claude sees them.
Control: A mod logs every action of all other mods.
Mods are installed as plugins and shared within the team. In Team and Enterprise plans, the administrator determines which plugin sources are allowed. A built-in security mod also prevents other mods from unlocking restricted actions.
Critical view: Mods run on your machine with the same permissions as Claude Code, without a sandbox. Only install mods from trusted sources.
Google replaces Gems with Skills. Google started rolling out Skills in Workspace on 5 October, and in the Gemini app from 13 October. Skills are reusable instructions that replace Gems. They use the same open format as skills in Claude, meaning you can copy custom instructions between platforms.
OpenAI dismisses three safety staff. On 1 October, OpenAI dismissed three members of its safety team. They allegedly shared confidential documents with an external safety organisation. OpenAI has not stated which documents were involved.
Three things to do this week
Test two models on the same task. Take a real task, such as summarising a proposal or analysing a spreadsheet. Let Claude Sonnet 5.5 and GPT-6.1 Sol solve it. Compare which output requires less editing.
Sort emails like Barclays. Export 20 typical customer queries without personal data. Give them to Claude and ask it to assign each to a topic and a responsible team. Check how many match. In future, no company should have a noreply@ address.
Try Mistral Vibe. Open Vibe at chat.mistral.ai and enter a research task from your daily routine. Compare the result with the tool you use today.