Anthropic said Monday that its new defenses cut the reasoning data harvested by distillers by around 90 to 99%. The coverage ran that number and the two Chinese labs the post names, Zhipu and Moonshot.
It is Anthropic grading its own homework, measured “on our own platform.” The same post says distillers also route traffic through third-party clouds and other inference providers.
The bill sits near the bottom. Legitimate developers had to “update their integrations,” the classifier wrongly blocks “fewer than 0.1%” of harmless requests, and “we now charge users for those blocked requests.”
Breaks your stack. The mechanism is called preserved thinking. On Fable 5.1 and Opus 5.5, an API account created on or after August 31 cannot change anything that came before a thinking block it sends back.
Anthropic’s docs define “anything” as the system prompt, the tool list and every earlier message. Stamp a fresh date into the system prompt mid-session, send the earlier thinking back, and the request dies with a 400.
Fewer than 0.1% sounds like nothing. By my arithmetic, at a million requests it is up to a thousand you paid for and got nothing back.
Matters. OpenAI raised the default context for GPT-6.1 Sol in Codex and Work to 500k tokens, up from 272k. It stays configurable up to 1M.
My read: fewer compactions, and a long session draws more from your plan on every turn. Pick the number on purpose.
Marketing. Meta shipped Muse Wonder on Monday. In the Muse app it needs a $20 or $100 monthly plan, and the larger Muse Wonder Full is limited to Muse Code subscribers at $50 or $200.
Meta AI chief Alexandr Wang told CNBC the API version will be priced “meaningfully below” OpenAI and Anthropic. That is an adjective, and I budget against numbers. Meta priced its whole app portfolio last month, so it knows how to publish one.
Three Questions Before You Ship on Fable 5.1
I use Claude Code. Does any of this touch me?
No. Anthropic’s docs say nothing changes when Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests. My content engine runs through Claude Code, so my pipeline is untouched.
My API key predates August 31. Am I safe?
Your key is. Your customers may not be. On an older account the check runs without raising an error, so a clean test on your own key proves nothing, and the docs warn that people on newer accounts “get the 400 error before you do.” Send the thinking-binding-controls-2026-08-01 beta header and each response flags the blocks that would fail.
Can I tell the API to drop the bad blocks and move on?
Yes, with drop_block, and the docs call it a stopgap. Claude rebuilds the reasoning it can no longer read, which costs you output tokens. The fix is boring: keep the system prompt and tools fixed for the session and only append messages.
This is the second day running that a safety number came from the lab being graded. Yesterday it was a detection layer scored on cases Anthropic already knew about.
If your integration edits context between turns and you want a second pair of eyes on it, book a free intro call.