Anthropic Releases Claude Fable 5.1 with Slashed Prompt Cache Read Costs
Anthropic introduced Claude Fable 5.1, offering deeper reasoning, agentic coding capabilities, and a 1M token context window. Prompt cache reads are reduced to a quarter of the prior cost alongside several breaking changes to thinking blocks and tool calls.
Impact: High
Why it matters
You can upgrade long-running agentic workflows and prompt-cached pipelines while cutting cache read expenses by 75%.
TL;DR
- 01Prompt cache read costs are reduced to 25% of their previous rate.
- 02Forced tool use is deprecated and now throws an error in Fable 5.1.
- 03Adaptive thinking is always on with configurable depth via the effort parameter.
Key facts
- Context Window
- 1,000,000 tokens
- Max Output
- 128,000 tokens
- Input / Output Price
- $10 / $50 per MTok
- Knowledge Cutoff
- June 2026
Pricing and Specifications
Claude Fable 5.1 is priced at $10 per million input tokens and $50 per million output tokens, supporting a 1M context window and 128K max output tokens. Cache reads drop to a quarter of their previous price, and Batch API calls receive a 50% discount across inputs and outputs. Anthropic guarantees retirement no earlier than September 1, 2027. The model is available across the Claude API, Amazon Bedrock under anthropic.claude-fable-5-1, Google Cloud, and Microsoft Foundry.
Breaking Changes and Agent Tooling
Existing systems migrating from Claude Fable 5 face three breaking changes:
- Forced tool use now triggers a direct error.
- Thinking blocks cannot be read by earlier Claude models.
- Editing earlier turns invalidates thinking blocks downstream.
Five additive features include turn-scoped system messages, per-message effort steering in beta, intermediate readable tool updates using display: "updates", content provenance tracking, and lowered cache read pricing.
Try it in 2 minutes
{
"model": "anthropic.claude-fable-5-1",
"max_tokens": 131072,
"thinking": {
"type": "adaptive",
"effort": "high"
}
}json
✓ When to use
- Complex multistep agent workflows needing large context reasoning and prompt caching.
- Demanding code generation and analysis where Claude Opus 5 hits reasoning limits.
✕ When NOT to use
- High-frequency low-latency tasks where Claude Haiku 4.5 or Sonnet 5 are more cost-efficient.
- Pipelines requiring forced deterministic tool use via deprecated API flags.
What to do today
- Remove forced tool use configurations from agent harnesses targeting Fable 5.1.
- Test dynamic effort adjustments on prompt-cached agent sessions to minimize token spend.
Sources