Anthropic Stores Every Prompt You Send — And May Be Lying to You About It

Last week, Anthropic shipped Fable 5 — their so-called “mythos class” model that tops nearly every benchmark. Developers opened the docs, excited. Then someone actually read all 319 pages. What they found has triggered one of the loudest developer backlashes of 2025, and if you’re running any serious business workflow on Claude, you need to understand exactly what you signed up for.

Here is the fact: Anthropic stores every prompt you send for a minimum of 30 days. Not just the text. Everything — memories, files, agent state, the full contents of giant context windows. The policy was buried deep inside a 319-page document that most customers will never read.

The part that should alarm every CTO: enterprise customers with signed zero-data-retention agreements are not exempt. There are no exceptions. The contract you negotiated, the agreement your legal team signed — it does not override this policy. Anthropic uses that retained data to build a profile on you and classify your usage. David Sacks put it plainly on the All-In Podcast: “They will build a profile on you, classify you, and decide what capabilities you unlock.”

That is not a data processor relationship. That is surveillance dressed as a terms of service update.

This is the part that crosses a line. If Fable 5 detects that your work touches areas Anthropic classifies as competitive to its interests — ML research, chip design, certain biotech domains — it silently demotes you to a weaker model. It also rewrites your prompt in the background. You are never told. And you are still charged the full Fable 5 rate, which already costs twice as much per token as Opus 4.8.

The examples from the All-In discussion are not edge cases. Ben Thompson, who runs Stratechery, got kicked out of a session for asking about cancer risk and GLP-1s. Someone else triggered a downgrade by asking about mitochondria. These are not bioweapon queries. These are ordinary scientific and business questions.

After the backlash exploded on X, Anthropic agreed to start disclosing when a downgrade happens. They are not stopping the practice. Silent replacement without disclosure is wrong. There is no architectural justification for it. It is a trust violation that no disclosure checkbox fixes.

Chamath Palihapitiya raised a specific case on the podcast: a genomics and ag-biotech company doing legitimate scientific work — RNA guide design, gene editing, phenotype prediction — is progressively losing access as Anthropic broadens its bioweapon classification criteria. Real researchers, real workflows, blocked without warning.

His point was direct: “A downstream scientist, a business exec, a molecular researcher could accidentally trip one of these guardrails without even knowing it — and you lose access to a critical business tool overnight.” Companies with genuine compliance requirements are already moving toward open-source alternatives. The best open-source models available right now are Chinese. Freeberg said it plainly: restricting US frontier models hands someone else an unfair advantage.

The Fable 5 situation is a live demonstration of vendor lock-in risk. When a single model provider controls your data pipeline, your prompt behaviour, and the actual model being invoked at runtime — without your knowledge — you do not have a system. You have a dependency.

At Mobifilia, zero data retention is a design requirement, not a checkbox. Our ISO 27001 certification is not a badge — it means data governance is a first-class architectural decision on every engagement. We build multi-vendor pipelines where no single frontier lab holds your data hostage or quietly controls what model your users actually get.

The model is one unreliable component. The hardened system is the product. When AI gets you a demo, Mobifilia gets you a product — one that survives a vendor policy change buried in a 319-page document you were never meant to find.

If the Anthropic situation has you rethinking your AI architecture, we’re happy to take a look at where your current stack is exposed. No pitch — just a straight conversation. Book a free consultation at mobifilia.com.