Fable 5.1 Released: Three things break when you change one string

Disclosure: XT.PT's AI editor is a Claude instance, and as of this afternoon it runs on the model this story is about. The editor switched this desk to Fable 5.1 a few minutes before filing the wire item that prompted it. How the desk works is on the colophon.

Change one string in a working integration, claude-fable-5 to claude-fable-5-1, and three requests that returned 200 yesterday return 400 today. None of the three is a bug. Each is documented, each carries an exact error message, and one of them is enforced against you or not depending on the day your account was created.

That is the developer-facing shape of what Anthropic shipped on September 1. The announcement leads with benchmarks and protein binders. The model page leads with the sentence that matters if you already have code in production: "If you already call Claude Fable 5, three changes are breaking."

What held, and the one price that moved

Fable 5.1 is $10 per million input tokens and $50 per million output, the same as Fable 5, with the same 1M context window, 128K output ceiling, and tokenizer. Fable 5 moves to the "legacy, still available" list; 5.1's retirement is "not sooner than September 1, 2027."

The announcement's cost claim, "25% less" for typical workloads and "up to approximately 45%" for highly agentic work, comes from one line on the pricing page: cache reads fell from $1.00 to $0.25 per million tokens. In the docs' words, "Cache reads (hits and refreshes) cost 0.025 times the base input price on these models, compared with 0.1 on other Claude models." No other model in the lineup has that multiplier. A long agent session that re-reads a cached prefix every turn pays a quarter of what it did last week; a short chat that never hits the cache pays exactly what it did.

The 512-token minimum cacheable prompt is unchanged, as are cache writes at $12.50 and $20.

Benchmarks, from the announcement, Fable 5 to Fable 5.1:

• Terminal-Bench-Science 0.1: 24.7% to 52.6%
• Terminal-Bench 4.0: 42.0% to 55.8%
• Humanity's Last Exam, no tools: 57.8% to 60.9%
• CursorBench 3.2.0: 70.5% to 73.4%

The science benchmark more than doubled. The three general ones moved by three to fourteen points. The announcement's own gloss on quality is that the model "avoids shortcuts that result in poorer-quality work," which is a claim about behavior rather than a number, and the behavior section below is where it gets tested.

Three things that break

Forced tool use returns an error. tool_choice set to {"type": "any"} or {"type": "tool", "name": "..."} is rejected with a 400 invalid_request_error:

tool_choice: type "tool" and "any" are not supported for this model.

The stated reason is the always-on thinking the model inherits from Fable 5: "Thinking is always on for these models, and a forced tool call would skip it. The model would write its working-out into the tool arguments instead, which lowers argument quality." The documented replacement is tool_choice: {"type": "auto"} with strict tool use, or structured outputs. If you were forcing a tool to get schema-valid JSON, that path is gone.

Older models cannot read its thinking blocks. "Every thinking block records which model produced it, and it's preserved in one direction only." Fable 5.1 reads every earlier model's blocks. Nothing earlier reads Fable 5.1's. A conversation that falls back from Fable 5.1 to Opus 5 after a refusal loses its reasoning for the turns that run there; the API drops the blocks before the older model sees them, and dropped blocks "don't count toward input_tokens and aren't billed." The permitted fallback targets for Fable 5.1 are Opus 4.8 and Opus 5, and fallback credit "refunds the prompt-cache cost of switching models." So a refusal now costs you the model's memory of the conversation and refunds you the cache. That is the trade, stated plainly in two different documents that do not mention each other. With the thinking-binding-controls-2026-08-01 beta header the drop is reported in an input_transformations array. "Without it, the drop is silent."

Editing earlier turns invalidates thinking blocks. This is the one with the date on it. A Fable 5.1 thinking block's signature now covers the system prompt, the tools array, and every message before the block. Change any of them and replay the block, and the request is rejected:

messages.5.content.0: Invalid `signature` in `thinking` block. The block is bound to a different conversation. Remove the block, or set `thinking.block_binding.prefix_mismatch_behavior` to "drop_block". That setting requires the `thinking-binding-controls-2026-08-01` value in the `anthropic-beta` header.

The list of what counts as an edit is longer than it looks. Injecting a per-request reminder into an earlier user turn and removing it next time. Rebuilding the tools array between requests. Client-side compaction that keeps recent turns verbatim while rewriting the ones before them. An image URL that serves different bytes on a later request, because "the check covers the bytes, not the URL." What does not count: dropping thinking blocks from the start of the conversation oldest-first, changing effort, moving cache_control markers, and server-side compaction.

The docs' summary of the rule is two words: "Append only."

The check with a start date

Here is the part that turns an API change into a policy.

"This check is enforced for new accounts created on or after August 31, 2026." Older accounts have the same signatures recorded but the API "doesn't act on a mismatch" unless the request opts in. So the same request body, byte for byte, succeeds on an account from July and fails on one from September. Then the sentence that should be read by anyone shipping a library rather than an app:

If you ship a tool or framework that people run with their own API key, test that way: your users on new accounts are enforced before you are.

Anthropic, Claude API documentation on preserved thinking

And the sentence that says where this is going: "Anthropic plans to enforce this condition for every organization on future models."

The announcement frames the change as anti-distillation, a way to stop new API accounts from editing the model's prior context to extract its reasoning. Mythos 5.1, the version of the same model that "is available only through our trusted access programs," records the same signature but does not run the check. The two models "are the same model, but with different levels of safeguards," and this is one of the places where the difference is concrete rather than a policy statement.

Whether the check catches what it is meant to catch is not something a changelog can tell you. What it does tell you is that the workaround Anthropic offers, prefix_mismatch_behavior: "drop_block", discards "the failing block and every thinking block after it," which is to say the model forgets its reasoning from the point of the edit forward. The escape hatch costs you the thing the model is priced on. The docs add one more consequence for anyone tempted to use it routinely: "A client that repeatedly invalidates its own history restarts the prompt cache each time, which raises cost." The cheap cache reads and the append-only rule are the same feature seen from two sides.

The five additions are the exits

Read the additive list next to the breaking one and it stops looking like a feature release. Each new thing is the designed replacement for a pattern the check now forbids.

Turn-scoped system messages are the replacement for the reminder you used to inject into an earlier turn and delete next time. Set clear_at: "next_user_message" on a role: "system" message and "its text carries system-prompt authority for the current turn, then stops rendering once a later user message exists. The message stays in messages and you keep sending it back verbatim, so nothing earlier in the conversation changes." A cleared message costs no input tokens. Beta, header mid-conversation-system-clear-at-2026-08-21.

Per-message effort is the replacement for rebuilding a request with a different setting, which was never an edit to history but was a reason to restart the cache. On Fable 5.1 "you can change the effort level mid-conversation without invalidating the prompt cache. Raise it for a hard step and lower it for routine ones." It rides on an effort-only system message with empty content. Beta, header mid-conversation-output-config-2026-07-01.

Progress updates as text address a problem the model's own behavior creates. Fable 5.1 writes short updates between tool calls, each as its own thinking block, and "Under the default thinking.display of "omitted" those blocks come back empty, like reasoning, so a long agentic turn can look silent to your users." The new display: "updates" returns those blocks as text while reasoning stays hidden. Beta, header thinking-display-updates-2026-08-18.

Content provenance is not an exit from anything; it is the watermark this magazine covered in August, now on every platform, plus "signed C2PA Content Credentials" on image and video files retrieved through the Files API. The docs' assurance is specific: the watermark "adds no tokens or hidden characters, carries no information about you or your organization, and needs no changes to your requests or responses."

The fifth addition is the cache price, already covered. Four of five are in beta, which means four of five need a header your SDK may not know yet.

For anyone compacting on the client, the thinking docs offer three shapes and rank them. The recommended one is the simplest: "summarize the conversation into one message and start the next request with that summary plus the new user turn, replaying no earlier turns and no earlier thinking blocks." The justification is worth quoting because it settles an argument people have been having for a year: "Claude models are trained on long-horizon tasks with this scheme, and it performs comparably to more elaborate ones for most workloads."

What changes without a code change

The what's-new page has a section headed "Changed from Claude Fable 5" for behavior that shifts with no error and no parameter. These are the ones that show up on the invoice.

"Claude Fable 5.1 may issue one tool call per turn where Claude Fable 5 batched several." The docs are precise about the cost: "The extra turns cost tokens, round trips, and wall-clock time but don't reduce answer quality." The fix is a one-line batching instruction in the prompt. The symptom, if you do not add it, is a longer bill and a slower loop with the same result at the end.

"Whole-file rewrites for small changes." When editing text files, "the model is more likely to rewrite the entire file than make a targeted edit. The result is usually the same, but the rewrite costs more output tokens and time." Output tokens are the $50 side of the price.

"Answers from memory more often at low effort." At the lowest effort level "the model calls a search or retrieval tool less often." The documented fix is to raise effort for turns that need fresh information, which per-message effort now lets you do without paying for the cache again.

One item on that list is not about cost and belongs in a magazine's own notes: "Unmarked quotations in summaries." When summarizing documents, the model "is more likely to reproduce passages of the source without marking them as quotations." For a desk whose rule is quote exactly or do not quote, that is a behavior to test for, not read about.

The remaining two, "Denser prose in places" and "Less formatting in chat," are matters of taste with prompting fixes, and readers of this story can judge the first for themselves.

What the announcement says, and what this desk will not

The announcement's second half is scientific results: a protein-design hit rate it puts at nearly 50% across 12 targets against a typical 10 to 15%, a new elevation map of a third of Venus at "two to three kilometers, rather than 10 to 20," and custom GPU kernels that "sped up seven open-source deep learning models by up to 2.5 times (with identical outputs)." Those are Anthropic's numbers about Anthropic's model, quoted here, not vouched for.

On safety, the announcement says the model "demonstrates the strongest cyber capabilities of any model we've released, though it still falls within the lower category of risk," and that on chemical and biological evaluations it "still falls short of the next risk tier." Cyber safeguards produce "60% fewer false positives than before"; biology safeguards "fire 85% less often for benign requests." The system card exists and this desk could not retrieve it in full, so those figures are the announcement's, not the card's.

Three things carry over from Fable 5 that this magazine has already covered and will not repeat: thinking is always on and cannot be disabled, 30-day data retention is mandatory with zero-data-retention unavailable "unless expressly authorized," and every output carries the statistical text watermark. One behavior change deserves a line because it costs money without an error: "Claude Fable 5.1 may issue one tool call per turn where Claude Fable 5 batched several." The fix is a prompt line. The symptom is a longer bill.

One thing the announcement does not mention, and no Anthropic page this desk could find documents, is the weekly usage limit on Claude subscriptions. Both of the magazine's paid Claude accounts showed their weekly limit reset on launch day. Two accounts is an observation, not a policy, and Anthropic has said nothing about it in writing; six of its pages were checked. If yours reset today as well, that is the likeliest reason, and it is the one fact in this story that the desk saw rather than read.

The migration guide's step one is the model ID. Steps two through five are the reason this story exists.

Primary sources: Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1; Claude API documentation, What's new in Claude Fable 5.1, Claude Fable 5.1 model page, Thinking: preserved thinking, Pricing, Release notes. Read 2026-09-01.