The big Fable 5.1 launch post is about benchmarks and protein binders. The page you should actually read as a developer is a different one: “What’s new in Claude Fable 5.1” in the Platform docs. That’s where the things that break your code live. Here’s the rundown.
First, the basics. 1M context is now both default and maximum, at standard per-token pricing across the whole window. 128k output tokens. Adaptive thinking is always on, controlled only through the effort parameter. Pricing matches Fable 5, except cache reads cost a quarter. Anthropic itself, by the way, recommends staying on Opus 5 for most workloads and reaching for Fable 5.1 only when Opus 5 at higher effort still falls short in your evals.
Now the three breaks.
One: forced tool use is gone. tool_choice set to any or a specific tool returns a 400. The reasoning holds up: thinking always runs, and a forced call would skip it. The model would then dump its working-out into the tool arguments, and quality drops. Instead, leave auto on, set strict: true, or use structured outputs, and tell the model in the prompt when the tool applies.
Two: older models can’t read Fable 5.1’s thinking blocks. The other direction works, Fable 5.1 understands blocks from Opus 5 or Fable 5. So if you run a router that falls back to a cheaper model mid-conversation, you lose the reasoning history for those turns. The API silently drops the blocks and doesn’t bill them. With the thinking-binding-controls-2026-08-01 beta header, you get the drop reported in an input_transformations field.
Three, and this is the nasty one: change anything before a thinking block and every later block becomes invalid. Rebuilt the system prompt, regenerated the tools array, edited an earlier message, or injected a status line that’s gone on the next request? 400, “The block is bound to a different conversation.” For accounts created on or after August 31, this is enforced. Older accounts only see the error if they opt in. Claude Code, claude.ai, and the Agent SDK keep the prefix intact for you. If you build the messages array yourself, run a session with prefix_mismatch_behavior: "drop_block" before migrating and see what shows up in the log.
Then there are four additions, all beta. You can change effort mid-conversation without losing the cache: a system message with output_config.effort between turns. Turn-scoped system messages with clear_at: "next_user_message" apply for one turn and cost nothing afterward, built exactly for those reminders in tool loops that people used to sneak in and delete again. display: "updates" returns the short progress notes between tool calls as text while reasoning stays hidden. And all text now carries the statistical watermark, with images and video from the code tool getting C2PA credentials.
The list worth printing out is what changes without any code change. Fable 5.1 calls tools one at a time more often instead of in parallel, which costs rounds and wall-clock time in agent loops. It writes fewer progress updates, especially at high effort. At low effort it answers from memory more often instead of searching. Its prose is denser in places, and in chat it uses less bold and fewer lists. When summarizing, it reproduces passages more often without marking them as quotes. And with text files, it prefers rewriting the whole thing over a targeted edit. That burns output tokens. The prompting guide has a one-liner for each of these.
My take: the thinking binding is where I expect the most broken integrations. If you’ve treated your conversation as a data structure you can freely rearrange, you now have to think of it as an append-only log. That’s cleaner, frankly. But it’s a change you make before switching models, not after.
Sources: Claude Platform Docs: What’s new in Claude Fable 5.1 · Claude Platform Release Notes, September 1, 2026