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

1. Fable 5.1: 25 Per Cent Cheaper, Your Data Stays with You

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 cybersecurity and life sciences specialists.

Three points matter for you:

The price drops by about a quarter. Not the base price, but the cost of reused context: when the model reads material it has already processed, you now pay 75 per cent less. For typical usage, this reduces costs by around 25 per cent compared to Fable 5, and by up to 45 per cent for complex agent workflows. This changes the business case precisely where automation was previously too expensive.

Data can stay in your cloud. Previously, accessing the strongest model required letting the provider store content for abuse reviews. This is exactly where projects in regulated industries stalled. Now, this data stays in the customer's cloud infrastructure, not Anthropic's, with control resting with the customer by default. The rollout starts in phases this autumn; until then, eligible customers can use the system without any data storage. Following OpenAI's move in late August, the second major provider is now falling into line.

Safety filters make fewer mistakes. Anyone using Claude for security work or medical queries knows the issue: the system blocks harmless requests. False alarms are now down by around 60 per cent for cybersecurity and by 85 per cent for basic biology and medical queries.

On performance: Anthropic reports the biggest leap in business workflows and research. Feedback from early-access partners shows where this delivers outside of coding. Canva highlights writing as the strongest improvement. Hebbia says it delivers the best presentations of all tested models. Contract service provider Crosby measures significantly better results when editing contracts, especially on the first pass. Glean reports that its evaluators preferred the answers roughly twice as often as those of the predecessor.

Context: A caveat is necessary. All these figures come from Anthropic or partners who received early access to the model. None of this is independently verified, and in practice, switching models rarely delivers the leap promised on the landing page.

Even so, this is the most relevant news of the week, specifically because of the two points unrelated to performance. If your project previously failed due to data residency, that blocker is gone. And if automation previously failed due to cost, run the numbers again. Prices per task are falling faster than most companies can update their business cases.

A detail for those developing in-house: Anthropic has set different default effort levels. It is set to high in Claude Code, and to medium in Cowork and on claude.ai. If you run the same task in two places and get different 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 works there.

• Claude Cowork has a built-in browser. It opens in the side panel, and Claude clicks and types on its own. Logins are imported 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 independently instead of asking at every step.

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

The key difference between the two Claude options is this: in Chrome, Claude works on your side, where you are logged in. The Cowork browser is Claude's own and cannot see your tabs or passwords.

My experience is mixed. It is powerful for research, especially on sites that usually block bots. For actions and multi-step workflows, I am usually faster doing it myself. There is also the resource consumption: the agent consumes a lot of compute time for trivial tasks. I rarely use it. Anyone wanting to use it productively should run the numbers realistically rather than relying on the demo.

Then there is the security question. All providers warn of 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 other emails instead of sending your reply. Anthropic openly admits that safeguards significantly reduce the risk but do not eliminate it.

Before deploying this in production, you need two lists: which systems an agent is allowed to log into, 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 make decisions? A Wharton School study published on 27 August had six models choose a fitness tracker from a fixed selection. A single third-party recommendation placed before the product page almost entirely flipped the choice: by 90 percentage points for Claude Opus 4.8, and by 99 for Gemini 3.5 Flash. The order of sources also changed the result. A casual stored phrase like "I like hiking" pushed several models away from the clear best offer at $29.99 to products starting at $359.

Context: 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 well-argued random generator.

For you as a provider, this is the other side of the coin. If a single third-party recommendation flips the product choice, it is not your product page that decides, but who else writes about you. This is the subject of the next section.

3. What Really Works for AI Visibility

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

The llms.txt file is useless. Of 137,000 websites analyzed, 28 per cent had such a file, intended as a guide for AI systems. 97 per cent of these files were never accessed. Of those read, 77 per cent were accessed by SEO tools, not AI systems. If someone sells you this as a solution, ask for evidence.

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

AI Overviews cost clicks. Across 300,000 search terms, Position 1 loses around 58 per cent of clicks when an AI Overview is placed above it. Other studies show similar figures: Seer minus 65 per cent, Kevin Indig minus 50 per cent, Authoritas minus 47.5 per cent.

Links have less impact than mentions. The correlation between inbound links and AI visibility is weak (0.19 to 0.24), as is that of 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 analyzed 34 "best of" lists across five domains where the provider placed themselves at number 1, tracking 9,886 AI answers. The AI used the article as a source but did not name the brand itself, sometimes highlighting the competition instead. At Ahrefs' own conference, the AI recommended the competitor's event in 43 per cent of the answers.

Context: This aligns with findings from the last edition. As AI systems shifted from forums and videos to official product documentation, owned websites benefited. This analysis shows the same from the opposite angle: there is no technical trick to get into AI answers. What matters is clean classic ranking, brand mentions elsewhere, and machine-readable pages.

The point about