AI Update: Four top models in 72 hours, and Europe adds €3 billion

1. Four Top-Tier Models in 72 Hours

Between 1 and 3 September, Anthropic, Google, Meta, and OpenAI each released a new flagship model.

• Claude Fable 5.1 (Anthropic, 1 September): the biggest performance leap of the week, and around a quarter cheaper. Details below.

• Gemini 3.8 Flash (Google, 2 September): Google's workhorse, the third Flash model in six weeks. Price unchanged at 0.75 dollars per million input tokens and 3.75 dollars output – but only until 31 December 2026. From 1 January, Google doubles this to 1.50 and 7.50.

• Meta Muse Spark 1.3 (2 September): coding model. Meta claims a fifth fewer tool calls than its predecessor. Relevant for development teams, not otherwise.

• GPT-6 Astra (OpenAI, 3 September): the model that operates the computer itself.

Tokens are the unit providers use for billing text – roughly word fragments.

Context: You will see Google's price footnote often in the coming months. The entry price attracts, the standard price comes later. Anthropic did the opposite in August, making the introductory price of Sonnet 5 permanent. If a model price appears in a business case, the expiry date belongs next to it.

Also: do not hard-wire any application to a single model. At this pace, how fast you can switch matters – not who is currently in the lead.

2. GPT-6 Astra: Operates the Computer, but Fails to Beat Claude

OpenAI released Astra on 3 September, calling it "the world's smartest and best-aligned model".

The innovation: Astra works on the computer instead of in a chat window. It fills out forms, maintains records in the CRM, cleans up calendars, and writes results directly into a document or email. In OpenAI's own benchmarks, it completes these tasks in about half the time of its predecessor. It adheres to existing templates – corporate layout, standard document structure, spreadsheets with working formulas. And it asks clarifying questions where the answer changes the outcome, rather than guessing.

Price: 10 dollars per million input tokens, 50 dollars output. Above 272,000 input tokens, OpenAI doubles the input price.

Context: The "smartest model in the world" claim does not hold up under independent measurement. For the provider Artificial Analysis, Claude Fable 5.1 led the rankings. In the next index update, both are tied on 53 points, followed by Claude Opus 5 on 51. Yet Astra is significantly cheaper: 3.26 dollars per test task compared to 7.63 for Fable 5.1.

More important than the order is how it came about. Artificial Analysis revised the index twice in one week following criticism that the initial rating of Astra was too low. A ranking that flips twice in seven days is no basis for decision-making. Test on your own use case.

One sentence from OpenAI's announcement is telling: Astra's reasoning is harder to monitor than its predecessor's because it solves tasks with fewer visible intermediate steps. Honest phrasing, and it says something about the direction.

What to do with this: Take a task where agents failed you last year. No prestige project, but something boring with lots of clicks – retrieving invoices from a supplier portal, reconciling master data between two systems, pulling numbers from a dashboard. That is exactly where the ground has shifted.

3. Claude Fable 5.1: The Biggest Leap of the Week

On 1 September, Anthropic released Claude Fable 5.1 and Mythos 5.1. It is the same model with two levels of safety: Fable for everyone, Mythos only for vetted professionals in cybersecurity and life sciences. The name sounds like an incremental step. The numbers say otherwise.

For business workflows, performance has almost doubled. In Anthropic's own evaluations of automated workflows, the score rises from 17.1 to 31.4 per cent. For scientific research in the terminal, it goes from 24.7 to 52.6 per cent. Both are tasks where the model must work autonomously over many steps – exactly where automation projects have previously failed.

More convincing than proprietary benchmarks are reports from companies that tested in advance:

• Hebbia (analytical software for financial institutions): the best presentations of all tested models, and the best accuracy rate when citing financial documents.

• Crosby (contract review): when redacting contracts, up from 47.9 to 57.0 points. The gain is primarily in the first pass – less rework.

• Glean (enterprise search): its own testers preferred Fable 5.1 over Fable 5 roughly two to one across everyday queries, research, and drafting.

• Rogo (financial analysis): same accuracy with 20 per cent fewer tokens, and clearly better at presentations.

• Millennium (investment firm): a crash that occurred roughly once every million runs, which no one could explain for four to five years. Fable 5.1 found the root cause in a third-party library.

• Every: "Fable intelligence, Opus price, Sonnet speed" – in their tests, roughly twice as fast as Opus 5 at half the token count.

The price drops without the base price dropping. This remains at 10 dollars per million input tokens and 50 dollars output – identical to Astra. Reused context becomes cheaper: if the model reads material it has already processed, it costs 75 per cent less, now 0.25 dollars per million tokens. For typical usage, this yields around 25 per cent lower costs, and up to 45 per cent for complex agent workflows. Astra charges 1 dollar for the same, which is four times more. Exactly where automation was previously too expensive, the calculation changes.

Fewer false alarms. Anyone who has used Claude for security or medical topics knows the groundless refusals. The safety guardrails now trigger about 60 per cent less often in Claude Code, and 85 per cent less often for harmless biology and medical queries. Fable 5.1 is now also allowed to find security vulnerabilities – though not build attack tools.

And for European companies: Anthropic has co-signed the EU AI Act's Code of Practice for labelling AI content. Texts from models released after 2 August 2026 carry an invisible watermark. A detection interface is running in closed preview for authorities, media, fact-checkers, and research. It contains no information about the user or organisation.

Context: Two caveats. The benchmarks are Anthropic's own, and the quotes come from partners with early access. Moreover, the 5.1 numbering undersells the leap – anyone using Claude daily probably did not realise that more has changed under the hood than in some major version releases.

In practice: If you discarded an automation use case in the last six months because the model was too inaccurate or too expensive, recalculate it now. Both obstacles shifted in the same week.

4. Mistral Secures €3 Billion – Europe's Largest Tech Round

On 8 September, Mistral closed a 3 billion euro funding round at a valuation of over 21 billion. According to Mistral, this is the largest equity round ever raised by a private European technology company.

• Led by Samsung Electronics with around 1 billion euros. Also participating: the EU-backed Scaleup Europe Fund, PSG Equity, new investor Advent, funds managed by BlackRock, and the Grand Duchy of Luxembourg.

• The valuation has almost doubled: the September 2025 round valued the firm at 11.7 billion euros.

• According to CEO Arthur Mensch, the capital will go into proprietary data centres – building and owning, rather than just renting.

• On the customer side: British retailer Tesco is entering a three-year partnership with Mistral and building a joint AI lab. Following Airbus and BMW in June, this is the next major European enterprise customer.

Context: This provides the sovereignty argument previously missing from the debate – not as a technical issue, but as a capital issue. A European provider can now finance infrastructure itself rather than buying it from the giants.

The contradiction must be noted: the largest single amount comes from South Korea, and Mistral is using it to build a data centre business with the same capital intensity as US providers. Anyone using "European independence" as a selling point should know it is funded here via international capital.

For you as a decision-maker, the reference list is the real argument. Airbus, BMW, Tesco – these are corporations that could have bought American. If you previously left Mistral off your list because the company seemed too small, that argument is gone.

5. Anthropic Releases Free Blueprint for Commerce Agents

On 2 and 3 September, Anthropic published its template for commerce agents and placed the code under an open license on GitHub. It includes two ready-to-use agents with examples for retail, travel, telecom, and entertainment.

• The shopping agent runs in the store: it browses the inventory, compares options, builds the shopping cart, and answers questions about orders and returns in the same conversation.

• The merchant agent works for the staff behind the scenes: sales figures, inventory alerts, price and promotion suggestions, campaign drafts.

Early adopters include Shopify, Visa, Mastercard, Accenture, and Priceline. Anthropic cites up to 35 per cent larger shopping carts and a 60 per cent higher conversion rate for active installations – manufacturer data from selected installations, without a control group. Treat these as a directional indicator, not a project plan target.

Context: More interesting is the core rule Anthropic reveals: money, write permissions, and identification numbers are secured in the program code, not in the prompt. The model proposes, the software executes. Anyone evaluating an agent – from any vendor – is left with one key check: where is the boundary the model cannot cross, and is it hard-coded or just an instruction?

For shop operators, this means the blueprint is free and the patterns are documented. The effort lies in connecting the catalogue, shopping cart, orders, and policies. That is integration work, no longer research.

6. ChatGPT Ads: Now Open for Booking in Switzerland

Last issue, we noted that ads were being displayed in ChatGPT in Switzerland, but could not be booked locally. This gap is closed: since 31 August, the Ads Manager has been running on ads.openai.com for 31 European markets – the 27 EU states plus Norway, Iceland, Liechtenstein, and Switzerland.

• Ads are only seen by users on Free and Go plans. Plus, Pro, and Enterprise remain ad-free.

• In Switzerland and the EEA, there is no personalised advertising at launch. Targeting is based on conversation topic, approximate location, and device type.

• OpenAI reported in early September that the ad business achieved an annualised run rate of one billion dollars in less than 200 days. An extrapolation from a current quarter, not an earned annual figure.

Context: The lack of personalisation is not bad news. Without personal data, the conversation context decides – which is closer to search engine marketing than social media. If you sell a complex service, you will find people mid-comparison here. Set a test budget with a conversion goal and an end date, not a permanent channel shift.

7. Nvidia Acquires Hugging Face – Now Confirmed

On 4 September, the contract was signed: Nvidia is acquiring Hugging Face for nearly 13 billion dollars, including around 11.9 billion for the existing owners and up to one billion as retention bonuses for employees. Antitrust regulators must still approve; both parties expect to close by 2027.

Developers and researchers worldwide use Hugging Face to source open AI models and datasets. It is the central hub for anyone wanting to run AI without the big providers. According to co-founder Clément Delangue, the initiative came from Hugging Face itself – the platform needed compute and capital. The price represents roughly eighty times the estimated annual revenue of about 150 million dollars.

Context: Open models are seen as the way out of vendor lock-in. Soon, the hub for these models will belong to the manufacturer of the chips they run on. Nvidia CEO Jensen Huang has promised the platform will remain open – for all developers and also for rival hardware. A statement of intent, not a contractual commitment.

If your application uses open models: note where you source them from and whether an alternative exists.

8. Switzerland: Capital Arrives, Data Does Not

On 1 September, Dun & Bradstreet published the latest findings of its AI Index for Switzerland, part of a quarterly survey of 10,000 businesses across 32 countries.

• Around seven in ten Swiss companies achieve a measurable return from AI: 48 per cent selectively in individual projects, 21 per cent broadly across multiple projects.

• Only 7 per cent consider their data fully ready to support AI at scale. 55 per cent say "partially", and more than one in six companies say "scarcely to not at all".

• Nearly 30 per cent expect AI spending to rise sharply over the next twelve months – almost double the global average.

• Less than half feel well-positioned to identify and manage risks such as bias, data protection, or compliance.

Context: This too is a vendor study – Dun & Bradstreet sells the exact data foundation whose absence it laments. Nonetheless, the pattern fits everything coming out of Switzerland in recent weeks: adoption is outpacing the foundation.

The combination of a 30 per cent spending increase and 7 per cent data readiness describes a year in which a lot of money is flowing into projects built on shaky ground. If you manage one of these budgets: the best franc spent right now is not on the next tool, but on ensuring your customer, supplier, and product data align clearly across all systems.

Add to this a labor market figure. On 3 September, Raiffeisen published an analysis of around 600 occupational groups. The result: about two-thirds of office workers are in roles heavily affected by AI – most heavily in finance, IT, and public administration, and regionally in Zug, Zurich, and Geneva. In a realistic scenario, the bank expects 1.6 to 7.9 per cent less demand for office space in five years.

Raiffeisen makes the caveat itself: no broad reduction in office jobs is visible in employment data yet. The forecast assumes AI will partially replace work rather than just supplement it. Anyone renewing an office lease in the coming years should choose the term deliberately rather than automatically.

And two dates. The second Swiss {ai} Weeks are running from 1 September to 4 October: over 180 events in 37 cities, around 220 partners, supported by ETH, EPFL, and Swisscom. At the AWS Summit in Zurich, private bank Pictet demonstrated how research in investment research drops from several days to minutes – bound by strict rules on who can access what.

9. Quick and Important

Washington looks for the emergency brake. Senator Bernie Sanders and Representative Greg Casar have announced legislation that would ban artificial superintelligence and temporarily pause advanced AI development until a federal agency establishes safety rules. It will not pass – it lacks a majority in the Republican-controlled Congress, as the sponsors themselves admit. The signal is what matters: the US debate has shifted from "do not choke innovation" to "pause development" in two months. Fittingly, Astra reaches the highest level for cybersecurity in OpenAI's own risk index – it can autonomously find and exploit unknown security gaps. The released version refuses offensive tasks.

10,000 agents, 88 hours, one proof. OpenAI set around 10,000 agents to work on one of mathematics' Millennium Problems and reported a proof on 5 September. Three caveats: the version solved does not meet the prize criteria – the million dollars remains unclaimed. A mathematician working on the same problem, who stored notes in OpenAI's development environment, asked publicly if his material was ingested; OpenAI denies this. And it is a lab setup, not a product. The methodology is what is relevant: a massive number of agents coordinating autonomously for days.

ChatGPT reads patient records. Since 1 September, healthcare organisations have been able to connect their Epic systems to ChatGPT – read-only, nothing is written back. OpenAI states that doctors rated 99.1 per cent of answers as safe across 4,363 evaluations. This is not a ready-made solution for Swiss hospitals: different systems, different data privacy laws. You should still know the direction.

IFA Berlin without Samsung. For the first time in 35 years, Samsung is absent from the main floor, with Xiaomi taking the space. The big trend: language models running directly on the device rather than in the data centre. Anyone procuring hardware in 2027 will be buying computing power designed for AI.

OpenAI DevDay on 29 September in San Francisco. Historically, this is where the announcements that impact daily business are made.

Apertus. There was no news regarding the Swiss open language model. The technical report for version 1.5 is still outstanding.

Three things you should do this week:

1. Recalculate a discarded automation use case. If you cancelled a project in the last six months because the model was too inaccurate or too expensive: Claude Fable 5.1 has nearly doubled performance on business workflows, and reused context costs 75 per cent less. Both reasons for cancellation shifted in the same week.

2. Write the expiry date next to every model price. Google doubles the price of its workhorse on 1 January 2027 – this is in a footnote. Anthropic made the introductory price of Sonnet 5 permanent. And do not hard-wire any application to a single model: with four top models in 72 hours, the ability to switch matters more than the choice itself.

3. Check the data foundation before you buy the next AI tool. Nearly 30 per cent of Swiss companies expect AI spending to rise sharply, but only 7 per cent consider their data ready. The single question: are your customers, suppliers, and products clearly the same across all systems?

Sources: https://www.anthropic.com/claude-fable-and-mythos-5-1 https://www.anthropic.com/news/enterprise-frontier-safeguards https://openai.com/index/gpt-6-astra/ https://the-decoder.com/artificial-analysis-overhauls-its-intelligence-index-after-gpt-6-astra-scoring-drew-skepticism/ https://artificialanalysis.ai/models/comparisons/gpt-6-astra-vs-claude-fable-5-1 https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/ https://techcrunch.com/2026/09/08/mistral-raises-e3b-as-sovereign-ai-becomes-big-business/ https://www.eu-startups.com/2026/09/french-ai-company-mistral-raises-e3-billion-series-d-led-by-samsung-at-over-e21-billion-valuation https://claude.com/blog/claude-for-commerce-agents https://github.com/anthropics/commerce-agents https://openai.com/index/chatgpt-ads-expands-across-europe/ https://www.moneycab.com/it/uebernahme-perfekt-nvidia-kauft-ki-plattform-hugging-face/ https://www.finanznachrichten.de/nachrichten-2026-09/69464957-ki-zahlt-sich-fuer-schweizer-unternehmen-aus-doch-die-datenbasis-bremst-den-naechsten-schritt-unternehmens-ki-auf-dem-weg-zur-reife-006.htm https://www.moneycab.com/dossiers/bueroimmobilienmarkt-steht-wegen-ki-vor-naechster-huerde/ https://ai-weeks.ch/ https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-ban-artificial-superintelligence-and-temporarily-pause-advanced-ai-development/ https://openai.com/index/navier-stokes-solution/ https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/ https://openai.com/index/chatgpt-connects-health-records-and-healthcare-sources/ https://openai.com/index/devday-2026/