Claude Fable 5.1: What's New, Pricing, and Migration Guide

Claude Fable 5.1 at a Glance

Anthropic released Claude Fable 5.1 on September 1, 2026. It is available in Claude for Pro, Max, Team, and Enterprise users, and through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The Claude API model ID is claude-fable-5-1.

The important positioning detail is easy to miss: Anthropic's own model guide says to start with Claude Opus 5 for most workloads. Fable 5.1 is the specialist choice for work that remains difficult at higher Opus effort, especially long-horizon agents, codebase-wide engineering, multistep research, and complex document work.

What Actually Improved in Fable 5.1

Anthropic says the largest gains over Fable 5 are in six areas: long agentic coding sessions, finished knowledge-work deliverables, multistep research, vision over dense documents, reasoning across the full context window, and computer use. The practical theme is not a better one-shot answer. It is staying coherent while a task expands across tools, files, and many steps.

• Coding: multi-file features, migrations, debugging, code review, and long autonomous sessions.
• Research: following evidence across several searches instead of stopping after the first plausible answer.
• Documents: moving from analysis to a finished document, spreadsheet, or slide deck.
• Vision: reading charts, tables, filings, and diagrams inside PDFs.
• Computer use: operating browsers and desktop apps while recovering from failed steps.

These are Anthropic's release claims, not a promise that every task will improve. Teams should test Fable 5.1 against their own hard cases and compare task success, wall-clock time, and total cost with Opus 5 before moving production traffic.

The Pricing Change That Matters

Base input and output prices have not changed from Fable 5. The material change is prompt-cache reads: $0.25 per million tokens, down from $1 per million tokens on Fable 5. That is a 75% reduction for cache hits and refreshes.

Anthropic estimates the lower cache-read price can reduce the cost of typical Fable workloads by about 25% and highly agentic workloads by as much as roughly 45%. Those are vendor estimates, and the real saving depends on how much stable context your application caches and reuses. A one-shot prompt will see little benefit. A long agent that repeatedly reads the same repository instructions, tools, or document set can benefit much more.

Batch API pricing remains a 50% discount on base input and output tokens. Cache writes cost $12.50 per million tokens for five minutes and $20 per million tokens for one hour.

New Controls for Long-Running Agents

Fable 5.1 adds three beta controls that are especially useful in agent products. They are API features and require the relevant beta headers, so they should not be treated as universally available defaults.

1. Per-message effort: raise effort for a difficult planning or debugging step, then lower it for routine work without invalidating the prompt cache.
2. Turn-scoped system messages: add a high-authority instruction for one turn without rewriting earlier conversation history.
3. Readable progress updates: request short status text between tool calls with thinking.display: "updates" while private reasoning remains hidden.

The model also adds content provenance. Anthropic says generated text carries a statistical watermark, while supported image and video files retrieved through the Files API can carry signed C2PA Content Credentials. This does not add visible characters or tokens to the response.

Three Breaking Changes Developers Must Handle

Changing claude-fable-5 to claude-fable-5-1 is not the whole migration. Anthropic documents three breaking behaviors.

1. Forced tool use is unsupported. A tool_choice of any or a named tool returns a 400 error. Use automatic tool choice with strict schemas or structured outputs instead.
2. Older models cannot read Fable 5.1 thinking blocks. Moving a conversation from Fable 5.1 to an earlier model causes those blocks to be dropped before the older model sees them.
3. Editing earlier turns can invalidate later thinking blocks. Changing the system prompt, tools, or an earlier message while replaying later Fable 5.1 thinking can produce a 400 error for affected accounts.

Treat Fable 5.1 conversations as append-only. Pass thinking blocks back unchanged, add new instructions as later messages, and use server-side context editing or compaction when the history must shrink. If your application rebuilds prior messages on every request, test that behavior before migrating users.

Fable 5.1 vs Opus 5: Which Should You Use?

A sensible production route is Sonnet 5 for routine volume, Opus 5 for difficult general work, and Fable 5.1 only for tasks whose value justifies the extra cost. Do not route everything to Fable because it is the newest model. Route the failure cases that need it, then verify the improvement with your own evals.

A Safe Fable 5.1 Migration Checklist

1. Run your existing eval set on Fable 5.1 before changing production routing.
2. Remove forced tool_choice settings and test strict tool schemas or structured outputs.
3. Confirm that your message history is append-only and that thinking blocks are returned unchanged.
4. Test every fallback route, especially any route from Fable 5.1 to an older Claude model.
5. Re-tune effort instead of assuming the default is right for every turn.
6. Watch whether independent tool calls still run in parallel; Fable 5.1 can be more variable here than Fable 5.
7. Measure cache-hit rate, cost per completed task, latency, and success rate together.
8. Review the 30-day data-retention requirement before sending sensitive workloads.

For Claude.ai users, there is no API migration to manage. Select Fable 5.1 when it is available on your paid plan, reserve it for demanding work, and keep checkpoints when a conversation becomes valuable or long. ClaudeKit's usage tracking, conversation forks, and exports can help protect that work around session limits and experiments.