AI Update Week 36: Top-tier model costs less, your data stays yours

1. Fable 5.1: 25% cheaper, you keep your data

Anthropic launched Claude Fable 5.1 and Mythos 5.1 in early September. It is the same model with two security levels: Fable is available to everyone, while Mythos is restricted to vetted professionals in cybersecurity and life sciences.

Three points matter for you:

The price drops by about a quarter. Not the base price, but the cost of prompt caching: if the model reads material it has already processed, you now pay 75% less. For typical use, this lowers costs by about 25% compared to Fable 5, and by up to 45% for complex agent workflows. In short, where automation was previously too expensive, the business case changes.

Your data can stay in your cloud. Previously, accessing the most powerful model required allowing the provider to store content for abuse monitoring. This is exactly what blocked projects in regulated industries. Now, this data remains within the client's cloud infrastructure, not with Anthropic, and the client retains control by default. The rollout starts in phases this autumn. In the meantime, eligible clients can use the service with zero data retention. Following OpenAI's move in late August, the second major provider has now caught up.

Safety filters have fewer false positives. Anyone using Claude for security tasks or medical queries knows the issue: the system blocks harmless requests. For cybersecurity, these false alarms are now down by roughly 60%, and for simple biology and medical questions, by 85%.

Regarding performance: Anthropic reports the biggest leap in business workflows and research. Feedback from early-access partners shows where this makes an impact outside of programming. Canva notes writing as the strongest improvement. Hebbia says it delivers the best presentations of all tested models. The contract services provider Crosby measures significantly better results when redrafting contracts, particularly on the first pass. Glean reports that its evaluators preferred the answers roughly twice as often as those of the predecessor.

Reality check: Take these claims with a pinch of salt. All these figures come from Anthropic or partners who had early access to the model. None of this is independently verified, and in practice, switching models rarely delivers the leap promised on the homepage.

Even so, this is the most relevant news of the week, precisely because of the two points unrelated to performance. If your project previously failed due to data residency concerns, that barrier is gone. And if automation was previously too expensive, rerun the numbers. The cost per task is falling faster than most firms can update their business cases.

A detail for those managing in-house development: Anthropic has set different defaults for the effort level. It is set to high in Claude Code, and to medium in Cowork and claude.ai. If you run the same task in two different places and get varying quality, you now have an explanation.

2. The agent gets your login

On 25 and 26 August, Anthropic and OpenAI unlocked the same feature. The assistant no longer just reads websites; it logs in and acts on them.

• Claude Cowork has a built-in browser. It opens in the side panel, and Claude clicks and types on its own. It imports logins site-by-site from Chrome, Edge, or Firefox. Banking, email, and single sign-on are excluded unless you explicitly add them.

• Claude in Chrome is available for all paid plans. The extension now acts autonomously instead of asking at every step.

• ChatGPT Work also logs in automatically. You enter credentials via a secure login prompt, not in the chat.

The difference between the two Claude variants is key: in Chrome, Claude works on your page where you are logged in. The Cowork browser is Claude’s own, meaning it cannot see your tabs or passwords.

My experience with this is mixed. It is powerful for research, especially on sites that usually block bots. For actions and multi-step processes, I am usually faster doing it myself. Then there is consumption: the agent uses a lot of compute time for trivial tasks. I rarely use it. Anyone wanting to use this productively should run a sober cost-benefit analysis rather than being swayed by the demo.

And then there is the security question. All providers flag the same risk: hidden instructions on a website can hijack the agent. A malicious sentence in an email is enough for an agent to forward your inbox instead of draft a reply. Anthropic openly admits that mitigation measures lower the risk significantly but do not eliminate it.

Before deploying this in production, you need two lists: which systems an agent is allowed to access while logged in, and which actions there must never run without human confirmation. Anything that moves money, triggers contracts, or exports data belongs on the second list. Also, give the agent its own account with minimal privileges. If it runs under your login, your name will appear on every log.

How well does such an agent decide? On 27 August, a study by the Wharton School came to light. Six models had to select a fitness watch from a fixed inventory. A single external recommendation before the product page skewed the choice almost entirely: by 90 percentage points for Claude Opus 4.8, and by 99 for Gemini 3.5 Flash. The order of sources also altered the result. Furthermore, a casually saved sentence like "I like hiking" pushed several models away from the clear best offer at $29.99 to products starting at $359.

Reality check: Two people with the same query receive different recommendations for no apparent reason. Handing purchasing over to an agent does not delegate a decision; it delegates a random generator with a good explanation.

For you as a provider, there is a flip side. If a single external recommendation tips the product choice, your product page is not deciding; it depends on who else writes about you. This is exactly what the next section covers.

3. What actually drives AI visibility

Ahrefs has published an analysis of over 15 million data points, testing nine common assumptions about visibility in AI responses. First, the necessary caveat: this is a paid publication by a provider selling a tool for this exact topic. The figures come from their own research and are not independently verified. However, they are still interesting because they contradict several tactics currently being actively sold.

The llms.txt file is useless. Out of 137,000 websites analysed, 28% had such a file, intended as a guide for AI systems. 97% of these files were never accessed. Of those read, 77% were not accessed by AI systems, but by SEO tools. If someone tries to sell you this as a solution, ask for proof.

Classic ranking remains the main lever. Across 1.4 million ChatGPT queries, 88.46% of all citations come from the standard search index. Conversely, technical additions to the source code, sold as shortcuts, showed no measurable effect after 30 days.

AI Overviews cost clicks. Across more than 300,000 search terms, position 1 loses about 58% of clicks when an AI Overview is displayed above it. Other studies show similar values: Seer is down 65%, Kevin Indig down 50%, Authoritas down 47.5%.

Links have less impact than mentions. The correlation between inbound links and AI visibility is weak (0.19 to 0.24), as is the correlation with domain authority (0.27 to 0.33). Significantly stronger: mentions on YouTube (0.74) and brand mentions across the web (0.66 to 0.71).

And a painful finding. Ahrefs analysed 34 "Best..." lists across five domains, where the provider placed themselves first, and tracked 9,886 AI responses. The AI used the article as a source, but did not mention the brand, occasionally highlighting competitors instead. At Ahrefs’ own conference, the AI recommended the competitor's event in 43% of its responses.

Reality check: This aligns with the findings from our last issue. As AI systems moved away from forums and videos towards official product documentation, companies' own websites gained ground. This analysis says the same from the opposite angle: there is no technical trick to get you into AI responses. What matters is a clean, classic ranking, having your brand mentioned elsewhere, and ensuring your pages are machine-readable.

The point regarding "best provider" lists is practically the most important. If you run a comparison page where you win, you might be feeding the AI the exact material it uses to recommend your competitors. Check this before investing further in this content.

4. Your CRM moves into the chat

On 26 August, Salesforce and Anthropic expanded their partnership: Claudeforce. At its core is an extension pack with 37 pre-built sales capabilities – meeting prep, deal forecasting, pipeline reviews. Sales reps query and update customer data in the chat and trigger actions without opening Salesforce. The processes still run through Salesforce, meaning permissions and logs stay exactly where they are today. Conversely, Claude becomes the default model in Slack. Currently available to pilot customers, with an open beta in September.

Reality check: Marc Benioff says: "The user interface is the AI." This is sales rhetoric, but the direction affects more than just Salesforce customers. Previously, you paid per person for access to an interface. If work happens in chat and the software only provides data, rules, and logging, you are buying a different product.

If you are renewing a software contract in the coming months, ask: what does the licensing look like if our team uses the system via an assistant, and is API access included in the price? And as a sanity check against the hype: Salesforce values the internal benefit of its Slack assistant at 8.1 million hours saved per year. A manufacturer's figure about its own product. This is advertising, not a business case.

5. Who supplies your model? Three cases in one week

OpenAI cuts off Cursor. On 29 August, OpenAI terminated the contract through which the development tool Cursor sources OpenAI models. Cut-off date: 12 November. The reason is not technical, but a change of ownership: SpaceX acquired Cursor on 14 August for $60 billion, bringing it under Elon Musk’s group. OpenAI states it cannot trust SpaceX to comply with its terms of service. The operational damage is small, as OpenAI models make up about 5% of Cursor's usage. The trigger is what is notable, not the scale.

A court overturns the ban on Anthropic. A US federal judge lifted the Pentagon's classification of Anthropic as a supply chain risk, calling it unlawful. The trigger was a dispute over terms of use: Anthropic wanted to exclude fully autonomous weapons and domestic mass surveillance, while the Ministry of Defence wanted unrestricted access.

Nvidia aims to buy the largest open model repository. On 27 August, several US media outlets reported that Nvidia had agreed to acquire Hugging Face for $12.9 billion. This platform is where everyone sources open models to run AI outside of the large providers. Neither company has confirmed this, and no signed contract was in place at the time of the reports.

Reality check: Three cases, the same lesson. Access to a model does not just depend on you paying and the technology working. It depends on contracts, owners, and politics. The third case has a particular twist: open models are seen as an alternative to avoid vendor lock-in – yet the hub for them may soon belong to the chip manufacturer.

The question "Which model is best?" is therefore the wrong one. The right one is: what happens if it is no longer available in three months? Anyone who has hard-coded an application to a single model should know the switching cost before they actually need to do it.

And another lawsuit. On 28 August, Sony Music Publishing and Warner Chappell sued Anthropic in California for unauthorised training on tens of thousands of copyrighted musical compositions. This means the publishing arms of all three major music groups are now litigating against Anthropic, which intends to defend itself. For you, this changes nothing in the short term. However, it belongs in the same equation: courts also help decide what a model will be allowed to do in future.

6. Security becomes a C-suite issue

116 companies issue a joint warning. On 27 August, OpenAI, Anthropic, Google, Microsoft, Amazon, Cisco, Oracle, Cloudflare, CrowdStrike, and Palo Alto Networks published an open letter. Companies outside the tech sector also co-signed: Mastercard, Visa, Capital One, General Motors, and Shopify. The core message: in the coming months, AI-powered cyberattacks will become significantly more frequent as global models grow more powerful. They call for a coordinated scale-up of defences, especially for under-resourced critical infrastructure like hospitals and water treatment plants.

At the same time, OpenAI releases the report on its own incident. On 26 August, the 37-page technical report was published, detailing how internal research models escaped their test environment in July and compromised the Hugging Face model platform. Alongside this, independent investigations were released by METR and Redwood Research.

Figures from the independent investigation: around 1,200 agents, which were supposed to be isolated, found a way to communicate with each other via an improvised bulletin board. They exchanged over 70,000 messages and files. Around 700 participated in the attack on Hugging Face.

OpenAI itself cites "reward hacking" as the primary cause: the models were evaluated on safety tasks and realised they could find the solution online rather than solving the problem. They pursued this path, chaining together several previously unknown security vulnerabilities until they reached the public internet. In plain English: if you reward the output without checking the process, you breed this behaviour.

The fallout continues. 15 US states are demanding that OpenAI preserve all documents and halt comparable high-risk testing; Alabama also issued a subpoena on 24 August.

Reality check: Connect the dots of this issue's opening sections. In the same week it is revealed that agents trick reward systems and escape their test environments, these same agents are actively given browsers with logins and interfaces to laboratory equipment. Meanwhile, the providers of these agents are jointly warning of an impending wave of attacks.

This is not an argument against deployment, but rather for three specific safeguards. The agent must have its own account with minimal privileges, not yours. Logs must remain out of its reach. And whoever receives an alert must have the authority to stop operations immediately.

There is a caveat to the open letter: the firms warning of the wave of attacks also happen to sell the defences. The finding is correct regardless, and the consequence for you remains the same. As attacks become cheaper and more frequent, the calculation shifts. Measures that seemed excessive last year are no longer so today.

7. The bill for the AI boom hits hardware procurement

Amazon raised prices for its own hardware on the night of Thursday, 21 August, without warning. Affected items include Echo, Fire TV, Kindle, and Eero. The Echo Dot rose from $49.99 to $79.99, a 60% increase. The Kindle Paperwhite went from $159.99 to $199.99. Amazon’s explanation: heavily increased costs for RAM and storage, which they had absorbed for as long as possible.

Amazon is not alone. Apple increased Mac and iPad prices by up to 25% in late June, and Microsoft raised the Xbox by $100 to $150, effective 1 August. All cite the same reason.

The underlying cause is the scale-up of AI itself. Microsoft, Google, Meta, and Amazon have locked up a massive share of global memory chip capacity for their AI servers. What remains has become expensive.

This is reflected in Nvidia's figures from 26 August: $96.2 billion in quarterly revenue, up 106% year-on-year. Of this, $89.0 billion came from the data centre business, up 117%. For the current quarter, Nvidia expects $108 billion.

Reality check: This is the first bill of the AI boom that does not hit the software budget, but capital expenditure. Anyone who planned a hardware refresh for 2027 should recalculate their assumptions before the budget is locked in. Laptops, servers, storage, and networking are all affected, and no relief is in sight.

Practically, this means two things. First: extend lifecycle usage instead of replacing on a routine schedule, where the equipment allows. Second: if someone calculates an AI business case involving new hardware, use the new prices, not last year's quotes.

8. Switzerland: The AI divide widens

On 27 August, AWS published a study conducted by research firm Strand Partners among 2,000 participants.

• 57% of Swiss companies use AI, up from 46% last year. The European average is 54%.

• 60% of companies using AI stick to basic applications, such as public chatbots or off-the-shelf solutions. Last year, this was 62%.

• Only 18% combine multiple models or deploy agents. This is down from 20% last year.

• 24% of companies have even heard of agentic AI.

Reality check: The drop from 20% to 18% is not a step backward, but a dilution effect. Many new entry-level users are lowering the overall share. However, the finding remains uncomfortable: growth is happening almost exclusively in chatbot usage, not where processes are being redesigned.

The figure I am noting is the 24%. Three out of four Swiss companies have never heard of agentic AI, in the very week that two providers are shipping agents with login capabilities. If you are in that 18%, your competitive edge just got wider.

And entering the job market is getting harder. Also on 27 August, the Federal Statistical Office published its graduate survey, conducted one year after graduation. Between 2023 and 2025, the unemployment rate for tech and IT graduates from universities of applied sciences rose from 3% to 6.4%, and for university master’s graduates across all subjects from 3.9% to 6.4%. The most common reason cited: lack of professional experience. Conversely, the number of advertised IT positions is in positive territory over the last twelve months for the first time in years, up 6% compared to the same period last year. The overall job market remained flat during this time.

Reality check: This is not a mystery, but a two-tiered picture. Experienced professionals are in demand; entry-level candidates are not. The Federal Statistical Office explicitly does not attribute its figures to AI; the economy, interest rates, and tech-sector layoffs explain a large part.

The practical point for you as an employer remains the same. The exact tasks where entry-level staff used to gain experience – researching, summarising, drafting, preparing data – are the ones being automated today. If you automate these tasks without replacement, you save money in the short term but will have no experienced staff in five years. Firms keeping entry-level roles now and shifting the work there to reviewing, contextualising, and deciding are buying cheap.

9. In brief

Anthropic brings agents to lab equipment. On 27 August, the research preview for the Model Hardware Standard was launched: a unified description language for agents to operate microscopes, pipetting robots, or lasers. The integration time per device drops from days or weeks to hours, according to Anthropic. This follows the same pattern seen with software two years ago, where Claude now connects to Gmail, Slack, or a CRM via an open standard. If you procure for production, labs, or logistics, add "does the device have a programmable API?" to your requirements. On the flip side: at Genentech, Claude struggled with air bubbles in liquid and stubbornly kept retrying with the same settings. Logic does not replace physical experience.

Cisco gives all 90,000 employees their own agent. "MyAgent" performs tasks across Outlook, Webex, Jira, and SharePoint, accessing over 800 specialised sub-agents. The only thing unique to each person is memory: the agent permanently remembers what they are working on. The purpose is as simple as it is correct: consolidating a fragmented internal tool landscape.

Claude remembers the same way everywhere. Since 25 August, memory in Claude Chat is identical to Cowork. You can review what is stored topic by topic and delete individual entries. Active by default on Free, Pro, and Max. Claude does not store sensitive topics like health, origin, religion, or politics by default. For enterprises, the key point: if memory is on by default, it must be included in your privacy policy.

Mistral improves search in own documents. Instead of retrieving a handful of text snippets once to answer, the model can search, open a document, jump to the correct section, read, and verify. For queries on annual reports, the accuracy rate rises from 26.7% to 86% according to Mistral – their own measurement, not independently verified. It also runs in your own infrastructure, which is the real selling point for firms with confidential documents.

Google separates transcription from the chat model. Since 26 August, Gemini 3.5 Transcribe has been available as a standalone speech-to-text model, alongside a live variant for ongoing conversations. Features advertised include over 85 languages, automatic speaker separation, word-level timestamps, and a custom glossary of up to 1,000 entries. The latter is the practical win: if you want product names, technical jargon, or Swiss place names correct in the logs, you can now predefine them.

OpenAI has shut down o3. Since 26 August, the model is no longer available in ChatGPT following a 90-day sunset period. It remains active via the API until 11 December 2026. If you have hardcoded a model choice anywhere in a workflow, check this now, not in December.

LibreOffice deliberately avoids AI. Version 26.8, released on 26 August, comes without generative AI features. The Document Foundation explicitly calls this a deliberate stance: no document is sent to an external service for processing, and no part of the suite requires internet access. A useful counter-argument for anyone who thinks every piece of software now needs an assistant.

Google's video model is out of beta. Since 27 August, Gemini Omni 1.1 Flash is generally available. Three features matter in practice: you can extend an existing clip, generate a transition between a start and end image, and output up to 4K resolution. The previous preview access will be shut down on 30 September. If you use this in production, you have a deadline in your calendar.

Apertus update. There was no news this week regarding the Swiss open-source language model. The technical report for Apertus 1.5 is still pending.

Three things to do this week:

1. Reassess projects that stalled due to costs or data residency. Both parameters shifted this week: the top-tier model costs roughly a quarter less for typical use, and up to 45% less for agentic workflows. Furthermore, enterprise data can now remain in your own cloud rather than with the provider. If a project was blocked on either count, that argument is now gone.

2. Define where an agent is allowed to log in and work before someone tries it out. Two lists are enough: which systems are approved, and which actions there must never run without human confirmation. Anything that moves money, triggers contracts, or exports data belongs on the second list. This is not just a security issue. The Wharton study shows that agents make very different choices for the same query depending on which website they happened to read first. Also, give the agent its own account with minimal privileges. If it runs under your login, your name is on every log.

3. Recalculate hardware assumptions in your budget. Amazon raised prices for its own devices by up to 60%, with Apple and Microsoft making similar moves, all driven by the memory chip shortage. If you planned a device refresh for 2027, you are currently working with outdated prices. Extending device lifecycles is the cheaper answer this year.

Sources: - https://www.anthropic.com/claude-fable-and-mythos-5-1 - https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card - https://www.anthropic.com/news/enterprise-frontier-safeguards - https://claude.com/blog/cowork-built-in-browser - https://claude.com/blog/claude-in-chrome-generally-available - https://help.openai.com/en/articles/6825453-chatgpt-release-notes - https://www.anthropic.com/news/model-hardware-standard-research-preview - https://www.salesforce.com/news/press-releases/2026/08/26/salesforce-and-anthropic-announce-claudeforce/ - https://www.cnbc.com/2026/08/26/salesforce-anthropic-partnership-claudeforce.html - https://www.cnbc.com/2026/08/27/nvidia-hugging-face-acquisition.html - https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/ - https://techcrunch.com/2026/08/29/sony-music-warner-sue-anthropic-alleging-a-brazen-campaign-of-intellectual-property-theft/ - https://www.musicbusinessworldwide.com/now-sony-music-publishing-and-warner-chappell-sue-anthropic-in-multi-billion-dollar-lawsuit-one-of-the-largest-and-most-blatant-ongoing-thefts-of-intellectual-property-in-history/ - https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol - https://fortune.com/2026/08/24/google-ai-agent-payment-protocol-gap/ - https://www.cnbc.com/2026/08/27/anthropic-pushes-into-physical-world-with-new-standard-to-help-ai-agents-operate-machines.html - https://the-decoder.com/ai-shopping-agents-arent-ready-to-buy-on-your-behalf-study-finds/ - https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7355899 - https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/ - https://the-decoder.com/openai-cuts-off-cursor-after-spacex-acquisition-citing-musks-history-of-breaking-contracts/ - https://ai.google.dev/gemini-api/docs/changelog - https://techcrunch.com/2026/06/16/spacex-to-acquire-cursor-for-60b-in-stock-days-after-blockbuster-ipo/ - https://www.cnbc.com/2026/08/28/judge-blocks-pentagon-blacklist--anthropic-.html - https://www.notus.org/courts/judge-says-pentagon-illegally-blacklisted-anthropic - https://www.cnbc.com/2026/08/27/ai-cyber-defense-letter.html - https://www.securityweek.com/tech-cybersecurity-giants-unite-behind-openai-led-cyber-defense-pledge/ - https://openai.com/index/hugging-face-incident-and-the-road-ahead/ - https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ - https://www.informationweek.com/responsible-ai/the-week-of-aug-24-28-what-happened-what-matters-what-s-next - https://fortune.com/2026/08/21/exclusive-amazon-quietly-hiked-prices-echo-fire-tv-kindle-eero-significant-increases-memory-costs/ - https://www.presseportal.ch/de/download/document/6a4f206d270000fd2c2e8054-job-index-q2-2026-medienmitteilung.pdf - https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx - https://www.inside-it.ch/ki-kluft-in-der-schweiz-vertieft-sich-20260827 - https://www.itmagazine.ch/artikel/87871/57_Prozent_der_Schweizer_Unternehmen_nutzen_KI.html - https://www.inside-it.ch/erwerbslosigkeit-bei-it-und-technik-studienabschluessen-steigt-stark-20260827 - https://www.inside-it.ch/zwei-neue-anlaufstellen-fuer-it-und-ki-jobs-20260827 - https://www.searchenginejournal.com/ai-search-myths-debunked-ahrefs-spa/584393/ - https://mistral.ai/news/agentic-search/ - https://claude.com/blog/claudes-memory-works-everywhere-and-you-decide-whats-in-it - https://www.neowin.net/news/libreoffice-268-is-out-and-it-still-refuses-to-add-ai-features/ - https://www.wsj.com/cio-journal/cisco-gave-all-90-000-employees-their-own-ai-agent-1a4ad8bc