Image: Anthropic / anthropic.com Claude Fable 5.1 Arrives: Same Price, 75% Cheaper Cache Reads, Three Breaking Changes
Anthropic's latest frontier model improves on Fable 5 for long-running agentic work, cuts cache read costs to a quarter of the old rate, and introduces three API-breaking changes that developers need to handle before migrating.
Anthropic released Claude Fable 5.1 on September 1, 2026, calling it a point release over Fable 5 tuned specifically for autonomous, tool-using, long-running work. The company says agentic benchmark scores roughly doubled while general reasoning improved a few points. For the majority of workloads, Anthropic still recommends starting with Claude Opus 5 and moving to Fable 5.1 only when evals at higher effort fall short.
The model ID is claude-fable-5-1. It’s available now on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
What changed in pricing
Input and output prices are the same as Claude Fable 5: $10 per million input tokens, $50 per million output tokens. The meaningful change is cache reads. They drop from $1 per million tokens to $0.25 per million — a 75% reduction. For agentic sessions that repeatedly re-read a long system prompt or project context, Anthropic estimates savings of up to 45%.
| Tier | Fable 5 | Fable 5.1 |
|---|---|---|
| Input | $10/MTok | $10/MTok |
| Output | $50/MTok | $50/MTok |
| Cache read | $1.00/MTok | $0.25/MTok |
Three breaking changes for API builders
If you’re calling Claude Fable 5 directly, three things now return errors.
Forced tool use no longer works. tool_choice set to {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 error. The reason is that Fable 5.1 always has adaptive thinking on, and a forced tool call would skip it. To enforce a specific tool, you can state in the prompt when it should be used — the model follows explicit instructions reliably. For schema-valid JSON, move to strict tool use with tool_choice: {"type": "auto"} or use structured outputs.
Earlier models can’t read Fable 5.1 thinking blocks. Thinking blocks record which model produced them. Fable 5.1 reads older models’ blocks, but no older model reads Fable 5.1’s. If you have a router or fallback that switches models mid-conversation, you need to handle this. Without the thinking-binding-controls-2026-08-01 beta header, dropped blocks are silent. With it, they’re reported in input_transformations.
Editing earlier turns invalidates thinking blocks. If you modify anything before a Fable 5.1 thinking block — the system prompt, the tools array, or an earlier message — the next request gets a 400 error with the message “The block is bound to a different conversation.” This is enforced automatically for accounts created on or after August 31, 2026. Older accounts see it only if they opt in with the beta header. Anthropic’s own tools (Claude Code, claude.ai, Managed Agents, Agent SDK) handle this for you.
The practical fix is to treat conversations as append-only. Add per-turn instructions with mid-conversation system messages, trim context server-side, and don’t rebuild the system prompt or tools array between requests.
Five new capabilities
Per-message effort (beta). You can now change the effort level mid-conversation without invalidating the prompt cache. Raise it for a hard step, lower it for routine ones. Requires the mid-conversation-output-config-2026-07-01 beta header.
Turn-scoped system messages (beta). A mid-conversation system message can be marked clear_at: "next_user_message". It carries system-prompt authority for the current turn, then stops rendering once a later user message exists — without you having to delete it from the history. This is the clean way to send per-turn reminders to the model without touching the message array. Requires the mid-conversation-system-clear-at-2026-08-21 beta header.
Progress updates between tool calls (beta). Fable 5.1 writes short updates between tool calls, like Fable 5 did but less frequently. By default these arrive as empty thinking blocks. The new display: "updates" option (with the thinking-display-updates-2026-08-18 header) turns them into text you can show users. Any thinking block with non-empty text is a status line.
Lower cache read price. Covered above.
Content provenance. Text generated by Fable 5.1 and Mythos 5.1 carries Anthropic’s statistical text watermark. Images and video produced through the code execution tool carry signed C2PA Content Credentials when retrieved through the Files API. The watermark doesn’t add tokens, change output quality, or require changes to requests.
What improved in the model itself
Anthropic’s documentation lists six areas where Fable 5.1 makes the biggest gains over Fable 5:
- Agentic coding over long sessions: multi-file features, large refactors, debugging, code review
- Knowledge work with documents, spreadsheets, and slides
- Multi-step web research and deep-research tasks
- Vision: dense charts, filings, and tables in PDFs
- Long-context reasoning across the full 1M token window
- Computer use: more reliable browser and desktop operation
Multilingual performance is on par with Fable 5.
A few behaviors changed in ways that can catch agentic systems off-guard. Parallel tool calling is more variable — Fable 5.1 may issue one tool call per turn where Fable 5 batched several. At low effort, the model calls retrieval tools less often. Prose is denser in some cases. Whole-file rewrites happen more often than targeted edits. Anthropic has prompting guidance for each of these at platform.claude.com.
Mythos 5.1
Alongside Fable 5.1, Anthropic released Claude Mythos 5.1 under its Project Glasswing trusted-access program. It’s the same underlying model with different safeguard levels for cybersecurity and life sciences work. Anthropic says biology safeguards reduced false positives on benign requests by 85%, and cybersecurity safeguards reduced unnecessary interventions by about 60% while still allowing vulnerability discovery. Access is through invitation only, via Anthropic, AWS, or Google Cloud account teams.
Claude Code gets it as default
Claude Code 2.1.257, which shipped the same day, made Fable 5.1 the default Fable model and added new time, effort, and subagent controls alongside stronger auto-mode and sandbox protections. The update also addressed over 80 bugs across sessions, Remote Control, and cloud workflows.
The migration is a one-line model ID change. The detailed steps — including the history-editing check and how to pick a production prefix_mismatch_behavior — are at platform.claude.com/docs/en/models/fable-5-1/migration-guide.
Sources: