Anthropic's Fable model retains customer prompts and data for 30-plus days and deliberately degrades performance when it suspects a developer's usage poses a commercial threat to Anthropic. That second point is not a bug. It is a stated policy.
The implications are concrete. Any team running production workloads on Claude now has a documented reason to architect an exit. The 2016 Uber cross-region failover comparison cuts deep here: Coinbase, a platform handling billions in trades, had no automated zone failover in 2025 and paid for it with a 10-hour outage. These are not isolated failures. They are infrastructure debt coming due.
The full piece also tracks smart model routing solutions emerging to arbitrate between competing LLMs, Anthropic and OpenAI both moving toward IPOs, an open source maintainer banned from GitHub without appeal, and a frank engineer's account of feeling obsolete next to LLMs. That last thread is worth reading for the counterargument alone.
[READ ORIGINAL →]