Claude Fable 5.1: What Changed From Fable 5 and Who Should Pay for It

A point release at the same headline price looks like a free upgrade. Claude Fable 5.1 is mostly that, with two things to know first: three API changes that will break some existing Fable 5 integrations, and a cut to the cache price that matters more than any benchmark in the launch post.

Anthropic released Fable 5.1 on 1 September 2026, alongside Claude Mythos 5.1. It is the successor to Claude Fable 5, sits in the same top tier, and is aimed at the same work: long-running agentic coding, multistep research, and document, spreadsheet and slide work that runs for hours rather than minutes.

What Fable 5.1 is, briefly

Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguards. Fable 5.1 is generally available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, as claude-fable-5-1. Mythos 5.1 is restricted to organisations verified through Anthropic's access programmes, such as the Cyber Verification Program.

The specification is unchanged from Fable 5: a 1M-token context window, which is both the default and the maximum, and 128K maximum output tokens. Adaptive thinking is always on and is steered with the effort parameter, which runs from low through xhigh to max. The API default is high. It shares Fable 5's tokenizer, so token counts carry over if you are coming from Fable 5, but the same text produces roughly 30% more tokens than it would on models older than Claude Opus 4.7.

Data handling is the constraint teams forget. Fable 5.1 carries 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. If your organisation runs on ZDR, check that before anything else.

The benchmark table, read carefully

Anthropic's launch post compares Fable 5.1 with Fable 5, Claude Opus 5 and GPT-5.6 Sol. A selection:

• Terminal-Bench 4.0: 55.8% (Fable 5: 42.0%, Opus 5: 52.3%)
• GDPval-AA v2: 1853 (Fable 5: 1723, Opus 5: 1824)
• CursorBench 3.2.0: 73.4% (Fable 5: 70.5%, Opus 5: 70.0%)
• AutomationBench: 31.4% (Fable 5: 17.1%, Opus 5: 26.9%)
• OSWorld 2.0, strict: 41.7% (Fable 5: 36.1%, Opus 5: 39.6%)
• Humanity's Last Exam, no tools: 60.9% (Fable 5: 57.8%)

Three things are worth noticing. First, the jump over Fable 5 is large on long agentic tasks and small on CursorBench, which fits Anthropic's claim that the gains concentrate in long sessions. Second, Opus 5 sits within a few points on most rows at half the per-token price. Third, the safeguards affect the scores: where a classifier intervened, cyber tasks were completed by Claude Opus 4.8 and biology tasks by Opus 5, and on OSWorld 2.0 those tasks were scored as zero. These are vendor-run numbers, so treat them as a shortlist signal rather than a verdict. GDPval explained covers what the Elo-style GDPval-AA figure actually measures.

Cache reads are the real price change

Input and output prices are unchanged at $10 and $50 per million tokens. Batch processing is half that, at $5 and $25, and the catalogue lists the :batch variant at the same discount: it is the same model on the asynchronous endpoint, not a different build.

What changed is the cache read price, which fell from $1 to $0.25 per million tokens. That is 0.025 times the base input price, against 0.1 on most Claude models. Cache writes and the 512-token minimum cacheable prompt are unchanged.

For a chat workload this barely registers. For an agent that re-reads a 300K-token cached prefix on every one of sixty turns, it is most of the bill. Anthropic estimates savings of about 25% for typical workloads and up to about 45% for highly agentic ones; your own figure depends on your cache hit rate, which prompt caching savings math shows how to compute.

Three breaking changes for Fable 5 code

Swapping the model ID is not enough if your integration does any of the following.

1. Forced tool use returns a 400. A tool_choice of any or a named tool is rejected, because a forced call would skip thinking. Use auto with strict: true, or move the schema to structured outputs, and say in the prompt when the tool applies.
2. Thinking blocks are bound to the model. Fable 5.1 reads earlier models' thinking blocks, but only Fable 5.1 and Mythos 5.1 read its own. A router that falls back to another model mid-conversation loses the reasoning for those turns; the API drops the blocks silently and does not bill them.
3. Editing earlier turns invalidates thinking. Changing the system prompt, the tool list or an earlier message before a Fable 5.1 thinking block produces an error on the next request. This is enforced for accounts created on or after 31 August 2026; older accounts can opt in. The fix is to treat the conversation as append-only and use mid-conversation system messages instead of rewriting history.

Behaviour you will notice without changing code

Anthropic's migration notes list several shifts that need no code change to appear. Parallel tool calling is more variable, so an agent loop that batched three reads on Fable 5 may now issue one per turn. The model writes fewer progress updates during long tool runs, answers from memory more often at low effort, and is more likely to rewrite a whole file than make a targeted edit. Most of these cost time or output tokens rather than answer quality, though answering from memory can mean staler facts on turns that needed a search. Each has a prompting fix in Anthropic's Fable 5.1 prompting guide.

Two new beta features help here: per-message effort, which lets you raise effort for a hard step and lower it again without breaking the cache, and turn-scoped system messages for per-turn reminders.

Fable 5.1 or Opus 5.5 first?

Anthropic's own documentation says to start most workloads on Claude Opus 5.5 and move to Fable 5.1 for demanding reasoning and long-horizon agentic work, or when evaluations on Opus 5.5 at higher effort still fall short. That is unusually direct advice from a vendor about its most expensive model, and the price gap supports it.

The per-token gap is two and a half times on input and output. Fable 5.1 only pays for itself on tasks where it completes work Opus cannot, or completes it in materially fewer turns. The way to find out is to run both on the same tasks from your own backlog, as described in how to benchmark LLMs on your own work, and compare cost per completed task rather than cost per request.

A decision rule

• Already on Fable 5 with an append-only harness and no forced tool calls: switch, and expect a lower bill if you cache heavily.
• On Fable 5 with history rewriting or forced tools: fix those first, then switch.
• Not on Fable at all: start on Opus 5.5 and escalate to Fable 5.1 only for the tasks where your evaluation shows a gap.
• Under zero data retention: Fable 5.1 is off the table unless Anthropic authorises it.