This week has been jam packed with Fable news. It’s been everywhere, Substack, Medium, pop ups on the interface, and the mailbox is literally overflowing with people advocating how marvellous it is, how fast it is, that everyone should use it, basically it’s the best thing since sliced bread. The news in shorter supply is what it does with the data, because it is not the same as the normal interface. I built contAIn™, and one of my big things is data use, so I went and read the terms.
Fable 5 retains your prompts, and the outputs it gives you back, for thirty days, on every platform it runs on, and you can’t switch it off either, even on the Enterprise version. If your company negotiated a zero retention agreement, it doesn’t cover Fable. Consumer plans were already being retained anyway, so this is most important for those who paid not to be.
The questions people are actually asking:
Does Fable 5 keep my prompts?
Yes. Prompts submitted to Mythos-class models, and the outputs they generate, are retained for thirty days to support safety work, on every platform where the models are offered. [1]
Can I turn the retention off?
No. Opus 4.8, Sonnet 4.6 and Haiku 4.5 can all be run under a zero data retention agreement, Fable 5 cannot, the exception is mandatory, and an existing enterprise zero retention agreement does not extend to Fable traffic. It applies however you use it, via Claude Code on an enterprise account, or through Bedrock, Google Cloud or Microsoft Foundry. [1] The retention is not a setting. It is a condition of use.
Can a human being read my conversations?
Anthropic’s answer is that by default none of its people can. Human review only happens down a controlled path, when the automated systems flag something, only a small approved set of reviewers can do it, and every single look is written to a log they cannot tamper with. [1]
A lawyer called Jessica Eaves Mathews read the terms back in June and worked out what it means for anyone handling privileged material. Her argument is that if a human employee at the vendor can read a flagged conversation containing privileged client communications, that is disclosure to a third party, and voluntary disclosure to a third party is one of the oldest ways there is to waive privilege. [2] That is her take rather than law, but one to mull over before pasting in anything confidential or IP related.
Is Anthropic training on my conversations?
Not quite the way most people think. The policy says Anthropic may use your inputs and outputs to train and improve its models unless you opt out, so it is opt out rather than opt in, and independent reviewers have called the pre-selected consent design a dark pattern. [3]
And here is the bit worthy of note. Even where you have opted out, conversations flagged for safety review can still be used for training. The policy does not say what triggers a flag and it does not promise to tell you when one happens. Flagged material is kept for up to two years, and the trust and safety scores that go with it for up to seven. [4]
So you can turn the training off, but turning it off is no guarantee that it’s not being held. Retention is a different thing altogether, and that’s not optional. Many people appear to think that flicking the training setting has dealt with it, it hasn’t.
The pitch and the terms are two different beasts
My mailbox last week was overrun with mails advocating the use of Fable. The scarcity messages, the get it while you can tripe that turns up regularly. Not one of them mentioned the data policy. I’m not insinuating that the data policy wasn’t covered at all, it was, CyberNews picked it up, Yahoo ran it, and the terms are on the Anthropic website in plain English. [5] What I’m saying is the sales emails didn’t make it explicit, and why would they, they’re sales emails after all.
The point is
That the sales emails and the terms of service are written for two different purposes.
The email drops in your inbox selling the dream. The terms sit on a page you have to find and how many people actually go looking? When was the last time you looked at the words above the permission boxes on code or the terms and conditions tickbox on something?
As data is the new gold, something that sounds too good to be true usually has a catch. The catch this time is your data, and whoever else’s data you may have fed in, or given Fable access to when falling into the hype of too good to miss try it noise.
The clue is in the name
I have always had a love of etymology and this one really made me smile.
A fable is a short fictional story designed to teach a moral lesson. It never pretends to be true, that is the point, the fox talks, you know the fox is not real and you get the moral at the end.
A myth is the opposite. Gods, superhuman beings, the origins of the world, it is told as true and believed as such. In modern usage a myth is a widely held but false belief.
So Fable is named after the story that admits it is a story, and the tier it sits in, the one above Opus, is called Mythos. Mythos carries a second interesting meaning. It is not only a myth, it is the whole pattern of belief a culture runs on, the set of attitudes a group holds without ever really examining them. [6] The shared story everyone is inside.
So if you fell into the Fable 5 hype take the moral to be: read the terms before you use it, and be very sure about whose data you are handing over, because you have no way of knowing what happens to it once it goes into the machine.
Sam
References
[1] Anthropic, Data retention practices for Mythos-class models, Claude Help Center, effective 9 June 2026. https://support.claude.com/en/articles/15425996-data-retention-practices-for-mythos-class-models
[2] Jessica Eaves Mathews, Claude Fable 5 Just Dropped. Here Is What the Fine Print Says About Your Data, 10 June 2026.
Jessica Eaves MathewsClaude Fable 5 Just Dropped. Here Is What the Fine Print Says About Your Data.When Anthropic released Claude Fable 5 today, I did what I always do with a new model launch. I went straight to the terms of service and privacy policy.This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber…Read more2 months ago · Jessica Eaves Mathews[3] Anthropic, Updates to our Privacy Policy, published 8 June 2026, effective 8 July 2026. https://privacy.claude.com/en/articles/10301952-updates-to-our-privacy-policy
[4] Anthropic privacy policy, model training section, safety-review carve-out and flagged-content retention periods, as read by Terms.law Privacy Watchdog and Tech Coffee House, June 2026. https://terms.law/Privacy-Watchdog/ai-services/anthropic/
[5] CyberNews, Companies using Fable 5 beware: it’s collecting your data, and there are no exceptions, 11 June 2026. https://cybernews.com/ai-news/claude-fable-five-data-retention-collection/ ; Yahoo Tech, Using Claude Fable 5 means your data will be collected. It’s not optional, 10 June 2026. https://tech.yahoo.com/ai/claude/articles/using-claude-fable-5-means-152936082.html
[6] Mythos, Merriam-Webster Dictionary. https://www.merriam-webster.com/dictionary/mythos
Samantha Maeer | Founder and Creator, contAIn™ | the methodology layer that directs AI to the outcome you define | contain.digital