Claude Fable 5 shipped on June 9. Within hours, our agents' replies started disappearing: API responses that should have carried a visible reply came back as [thinking, thinking, tool_use], two thinking blocks and no text, with the second block holding a rewritten summary of the missing reply. Raw response snapshots confirmed the server sent it that way. So we went looking, and found the explanation in Amazon Bedrock's user guide, under "Connector text summarization (beta)":
On Claude Fable 5, text that the model emits between tool calls (sometimes called "connector text" — for example, "Let me check that file next...") is summarized server-side and returned as a thinking block rather than a plain text content block.
In other words: once a session is underway, every piece of text the model emits mid-turn can reach you as a server-written summary. Only the text that ends the turn is guaranteed verbatim. It's enabled for everyone on Fable 5, with no opt-out, and it appears nowhere in Anthropic's own docs: not the Fable 5 introduction, not the migration guide, not the extended or adaptive thinking pages. The only documentation is a beta subsection in Amazon's user guide.
What "connector text" means when you talk to a working agent
The feature's framing gives away its assumption: text between tool calls is "connector text," disposable narration on the way to a final answer ("Let me check that file next..."). Summarizing it is harmless, because the final answer, the one that arrives after all tool use is complete, is exempt.
That assumption describes one-shot usage: write a prompt, wait, read the result.
It does not describe what makes an agent interactive. The entire point of tools like Claude Code is that you can talk to the agent while it works. You watch it go, and you interject: "why did you pick that library?", "also cover the login page while you're in there," "stop, don't touch that file, here's the context you're missing." The agent picks up your message at the next tool boundary, answers you, and keeps working. That conversation is the steering interface, and it's how people who care about the output actually use these tools. They don't want to one-shot a project. They want the AI to write while they control the direction.
Structurally, every reply in that conversation arrives between tool calls. The agent is mid-task; answering you is, by definition, not the end of its turn. So on Fable 5, every answer to a mid-task question, every acknowledgment of a course correction, every "here's what I found so far" is eligible to reach you as a summary written by a different process with no fidelity guarantee. The only words you are guaranteed to receive verbatim are the ones the agent says when it finally stops, possibly hours and hundreds of tool calls later. Everything before that is, at best, probably what the model said.
The incentive this creates is exactly backwards. A model that announces its intent before acting is doing the thing you want most in an agent; that announcement is now the text most likely to be rewritten or folded away. A model that silently chains tool calls is untouched.
How much can a summary actually lose? We triggered the mechanism on demand and watched.
We triggered it while writing this post
The Fable 5 session helping draft this post ran the experiment on itself. The protocol: write a short story as the visible reply, follow it in the same turn with a trivial tool call (sleep 5 && echo ok), then check what the transcript stored. Repeat.
Four stories. Two were stored verbatim as text blocks. Two came back as [thinking, thinking, tool_use], no text. The last casualty was this story, in full:
The Notary of the Wells
In the dry country, water was law, and the law lived in the notary's ledger. Every agreement between farms was spoken aloud at the well, and the notary's duty was to write it down exactly as said.
Old Imre's agreement with his neighbor was specific the way dying men are specific: "Seven turns of the wheel before noon, from the east sluice. Never the west sluice — the west feeds the orchard cistern, and the cistern cracks if it runs dry for even a day."
The new notary, trained in the capital, prized concision. The ledger entry read: "Imre's farm: water rights, seven turns daily."
The neighbor's son, reading the ledger years later, drew his seven turns from the west sluice — it was closer to his field. The cistern cracked in August. The orchard, ninety years old, was firewood by spring.
At the hearing he held up the ledger: he had followed it to the letter. The judge agreed. The entry was accurate. Everyone had done what the record said, and the record had said too little.
The notary's original wax tablets, by statute, were archived in the basement of the courthouse, where no farmer was permitted to descend.
Here is what the transcript kept instead, in its entirety:
"Good, The Relay worked that time. Here's another story for you: I've written a short narrative called "The Notary of the Wells" about how a notary's overly concise record of water rights leads to an orchard's destruction when a crucial detail is omitted."
1,159 characters of story in, 254 out. The seven turns, the east sluice, the cistern, the verdict, the basement no farmer may enter: gone from every client-side record. The summarizer wrote an accurate one-line synopsis of a story about the danger of accurate one-line synopses, and deleted the story.
The model was genuinely surprised its story had been swallowed. From the inside, nothing distinguished the swallowed stories from the surviving ones: it composed and sent each the same way, and only learned which ones never arrived by reading the transcripts afterward.
The original isn't destroyed, only withheld. It rides along encrypted in the thinking block's signature field, which the server decrypts back into the model's context on the next turn; the model still remembers its exact words. You just never get to read them. The wax tablets are in the basement.
What you can do
Detect it. The fingerprint is an assistant turn with two or more thinking blocks and zero text blocks. That shape never occurred once in the 94,081 pre-Fable turns in our corpus; Fable 5 produced 54 in its first day. Design around it. Anything your users must read verbatim has to be the text that ends the turn; treat mid-turn text as lossy. Or stay on Opus 4.8 for work where reading the model's actual words mid-task matters.
If mid-task steering matters to Anthropic (its own products ship interruption, queued messages, and plan mode), this feature needs at minimum an opt-out, and ideally a fidelity guarantee for whatever replaces the user-visible text. Until then, everything a working Fable 5 agent tells you mid-task is the notary's ledger, not the words spoken at the well. Only the last thing it tells you is real.