Start With the One That Breaks Existing Code
Forced tool calls are gone. Setting tool_choice to any or to a specific tool name now returns a 400 error. Anthropic's stated reasoning is that thinking is always on for this model, and a forced call would skip it, pushing reasoning into the tool arguments and lowering their quality. Automatic tool choice and none remain supported. The recommended replacement is auto combined with strict tool use or structured outputs. If your application genuinely needs a specific tool on the current turn of a long conversation, Anthropic's suggested approach is to append a message with role system after the latest user turn, naming the tool and stating the call is required — which keeps earlier turns byte-identical so the prompt cache stays warm. One gateway-level trap is worth recording separately: LiteLLM maps OpenAI's tool_choice of "required" onto Anthropic's any, so code that ran fine on Fable 5 fails on 5.1 without a single line changing. Two other breaking changes: earlier models cannot read its thinking blocks, and editing earlier conversation turns invalidates thinking blocks. The latter is enforced today for API accounts created on or after August 31, 2026, and is expected to apply to everyone at a future release.
Unpack the "45% Cheaper" Figure
What is certain first: $10 per million input tokens and $50 output, unchanged. What changed is cache reads, from $1.00 to $0.25 per million — a 75% cut, which as a multiplier is 0.025× base input against 0.1× on every other Claude model. That is a genuine structural change, and it matters most for agentic workloads leaning heavily on long cached context. Now the discount on the discount: the 25% and 45% figures come from Anthropic's own four weeks of August 2026 usage across Claude Enterprise, Claude Code and the API, measured at default effort. Vendor self-measurement, not third-party benchmarking. And the savings apply where usage is billed per token, so subscription users see no such discount. Two third-party measurements point in opposite directions. Artificial Analysis found that at maximum effort, Fable 5.1 costs more per completed task than Fable 5, because it produces roughly 1.7× as many output tokens. Cognition measured the reverse on its coding benchmark: $2.68 per task, down from $5.84. These do not contradict each other — they show the conclusion depends heavily on effort level and task shape. To know your number, run your own workload.
Effort Levels and Behavioral Changes
Levels go from three to five: low, medium, high, xhigh, max, with no way to turn thinking off entirely. Defaults vary by surface — High in Claude Code, Medium in Claude Cowork and on Claude.ai. A beta header allows switching effort per message without invalidating the prompt cache. One easily missed note sits in the official docs: effort level names do not represent the same amount of thinking across models, so an effort sweep run on Fable 5 does not transfer. Several behaviors need re-adaptation too. Parallel tool calling is more variable, so an agent loop where Fable 5 batched several calls may now issue one per turn. The model narrates less and answers from memory more often at low effort. When editing files it prefers whole-file rewrites over targeted edits. Also, neither Fable 5.1 nor Mythos 5.1 is supported on Priority Tier, though Fable 5 is.
Placed in This Product Line's History
Fable 5 launched June 9, was switched off worldwide, and returned 19 days later on June 30 — this site covered that "alive for four days" shutdown at the time. 5.1 arrives just three months after 5. Anthropic also cites a safeguard adjustment this time: 60% fewer false positives on cybersecurity, allowing vulnerability discovery but not exploit development. The practical implications split in two. For subscription users, what you will notice is response style — less narration, more whole-file rewrites — with no direct cost change. For API users, work in this order: grep your code for tool_choice with any or a named tool (including indirect mappings through gateways like LiteLLM), then rerun an effort sweep on real workloads because the old level conclusions do not carry over, and only then evaluate how much the cache price cut actually saves you.
via: Anthropic's announcement, Claude docs: what's new in Fable 5.1, Claude docs: migration guide, MarkTechPost