Claude Fable 5.1: 75% Off Cache Reads, Change Reasoning Budget Mid-Conversation

Same price, cache cost cut to a quarter. In real operational terms, switching to Fable 5.1 means a noticeably smaller monthly bill for cache-dependent apps.

Anthropic released Claude Fable 5.1 on September 1, 2026. As the version number suggests, this is an incremental improvement on Fable 5. But "incremental" does not fit this release. The 75% cut to cache read costs and the addition of per-message effort control are changes that create real economic impact — especially for developers and businesses calling Fable via API.

1. Core Change 1: 75% Off Cache Reads

Prompt caching reduces token costs when reusing the same system prompt or document across multiple calls. You pay the expensive "cache write" cost on the first call, then only the much cheaper "cache read" cost for subsequent calls.

In Fable 5, cache reads cost $1.00 per million tokens. In Fable 5.1, that is **$ 0.25** — exactly one quarter of the original price.

Model	Input (base)	Output	Cache Write	Cache Read
Fable 5	$10/M	$50/M	$12.5/M	$1.00/M
Fable 5.1	$10/M	$50/M	$12.5/M	$0.25/M

Here is a concrete example of why this matters. Suppose you build a service that maintains a long legal document as system context while answering multiple questions. If you reference a 100,000-token contract 1,000 times per day:

Fable 5 cache read cost: $0.10 × 1,000 = **$ 100/day**
Fable 5.1 cache read cost: $0.025 × 1,000 = **$ 25/day**

That is a $2,250 difference per month. This gap scales linearly as operation size grows.

2. Core Change 2: Per-Message Effort Control

This is the more technically interesting change. Starting with Fable 5.1, per-message effort control is available in beta.

Until now, Claude''s reasoning budget was set at conversation start and fixed for the entire conversation. Setting "use up to 10,000 tokens for reasoning in this conversation" applied the same budget whether the question was simple or complex.

There was a larger problem: changing the reasoning budget invalidated the prompt cache. If a more complex question arose mid-conversation and you raised the budget, the cache built up to that point was wiped.

Per-message effort control solves both problems simultaneously.

# First message: quick answer is enough
message_1 = {"role": "user", "content": "...", "effort": "low"}

# Third message: deep analysis needed
message_3 = {"role": "user", "content": "...", "effort": "high"}

# Prompt cache remains intact

Reasoning depth adjusts per message, while system prompt and earlier context cache remain unchanged.

Real-World Scenarios

Customer Support Agent: Route simple FAQ questions to "low" effort, contract interpretation or edge cases to "high" effort automatically. Optimize both cost and quality simultaneously.

Document Tool: "Keep writing the next paragraph" gets low effort; "Critically review the logic structure of this section" gets high.

Education Platform: Dynamically provide explanations at different depths for the same question depending on the student''s understanding level.

3. Base Specs: 1M Context and Always-On Adaptive Thinking

Fable 5.1''s base specs are the same as Fable 5.

Context window: 1,000,000 tokens (~750,000 words)
Max output: 128,000 tokens
Training cutoff: June 2026
Thinking mode: Always-on adaptive thinking

Always-on adaptive thinking is the Fable line''s defining characteristic. Reasoning happens internally even without explicitly enabling a thinking mode. Depth of thought adjusts automatically based on problem complexity.

Performance Improvements

Three specific improvements are reported in Fable 5.1:

Front-end code generation: Smoother results for React, Vue, and other UI component generation reported compared to the predecessor.
Precondition retention in long reasoning chains: Earlier preconditions are maintained more stably to the end in multi-step reasoning.
Tool call reliability in autonomous workflows: Tool calls behave more consistently in agent pipelines.

4. Important Note: Thinking Block Backward Compatibility

One thing to watch when adopting Fable 5.1:

Fable 5.1 can read thinking blocks generated by earlier Claude models. However, earlier models cannot read thinking blocks generated by Fable 5.1.

Systems with multi-model pipelines or fallback structures need to account for this. If a conversation running on Fable 5.1 falls back to an earlier model, the reasoning performed by Fable 5.1 is not passed to the next model.

5. New Beta Features

In addition to per-message effort control, two more beta features were added.

Turn-scoped system messages: Set system messages that apply only to a specific message turn. The system prompt automatically reverts to the original on the next turn. Useful when specific pipeline steps need different personas or instructions.

Updates display mode: The model streams readable progress updates between tool calls. For long agent tasks, users can see "what is happening right now" in real time.

Tips

Migrate cache-dependent services immediately: Same price, same performance, cache reads at a quarter of the cost. If you are using Fable 5, there is no reason not to switch to Fable 5.1.
Classify effort into three levels: "low" (simple fact-checking, short answers), "medium" (general analysis), "high" (deep reasoning, creative work) — design a message router using these tiers.
Cache long documents in the system prompt: Do not pass reference documents on every call. Put them in the system prompt once and cache them. Maximize the benefit of Fable 5.1''s lower cache read cost.
Improve UX with updates mode: Adding updates display mode to long-running agents reduces user drop-off. It addresses the common complaint "I have no idea what the AI is doing."

Sources: