Claude Fable 5.1: 75% Cheaper Cache Reads, But Reserve It for Long-Running Hard Tasks

On September 1, Anthropic released Claude Fable 5.1, its highest-tier model now available to Pro, Max, Team, and Enterprise users via Claude Code, Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Fable 5.1 shares the same base as Mythos 5.1 but differs in safety constraints and access scope. Claude Code requires version 2.1.250 or higher to list the model.

Core Specifications

Context capacity: 1 million tokens

Max single output: 128K tokens

Default thinking: Adaptive Thinking (automatic reasoning depth)

API default effort: high

Knowledge cutoff: June 2026

The key advance is sustained agent performance on long tasks: cross-repository refactors, multi-hour debugging and code review, iterative research with self-correction, browser and desktop automation with failure recovery, dense PDF chart/table extraction, and end-to-end research-to-deliverable workflows. Anthropic's guidance: start with Opus 5 for most tasks; only escalate to Fable 5.1 when Opus 5 at high effort still falls short.

Benchmark Results

Official evaluations compare Fable 5.1, Fable 5, and Opus 5 on four developer-relevant benchmarks:

Terminal-Bench-Science 0.1: Fable 5.1 52.6%, Fable 5 24.7%, Opus 5 29.0%

Terminal-Bench 4.0: Fable 5.1 55.8%, Fable 5 42.0%, Opus 5 52.3%

AutomationBench: Fable 5.1 31.4%, Fable 5 17.1%, Opus 5 26.9%

CursorBench 3.2.0: Fable 5.1 73.4%, Fable 5 70.5%, Opus 5 70.0%

Terminal-Bench-Science more than doubles (24.7% → 52.6%), and AutomationBench nearly doubles (17.1% → 31.4%). CursorBench, a coding-centric benchmark, rises only from 70.5% to 73.4%. The author interprets this as confirmation that Fable 5.1's advantage appears on tasks with many steps, self-debugging, and extended horizons — not on typical code generation. Safety constraints caused zero scores on some OSWorld tasks for Fable 5.1 and Fable 5, and on some AutomationBench tasks for Fable 5.

A customer case illustrates the niche: Millennium chased a one-in-a-million crash for 4–5 years; Fable 5 failed, but Fable 5.1 disassembled a vendor library, matched it against a core dump, and pinpointed a supplier bug. This exemplifies the long-chain forensic reasoning Fable 5.1 targets.

Pricing and Cache Economics

Per-token prices unchanged from Fable 5:

Fable 5.1: Input $10 / 1M tokens, Output $50 / 1M tokens

Opus 5: Input $5 / 1M tokens, Output $25 / 1M tokens

Sonnet 5: Input $2 / 1M tokens, Output $10 / 1M tokens

Fable 5.1 remains 2× Opus 5 and 5× Sonnet 5. The discount is on prompt cache reads: from $1 to $0.25 per 1M tokens (75% reduction). Cache write costs stay at $12.50 (5-minute TTL) and $20 (1-hour TTL). Anthropic estimates typical task cost drops ~25%, heavy agent tasks up to ~45%, but savings scale with context reuse — long tasks with repeated system prompts, tool definitions, and history benefit most.

Subscription Tier Access

Max, Team advanced seats, Enterprise advanced seats (per-seat): Fable 5.1 usable from weekly quota, capped at 50% of weekly allowance.

Pro, Team standard seats, Enterprise standard seats (per-seat): billed from usage credits from the first token.

API and usage-based Enterprise: standard API pricing.

Free: not available.

A prior Fable 5 promotional credit period has ended and does not cover Fable 5.1. Default effort differs by surface: high in Claude Code, medium in Claude.ai and Cowork — align effort when comparing results.

Migration from Fable 5: Three Breaking Changes

1. Forced tool_choice removed

tool_choice

values "any" and "tool" now return HTTP 400. Only auto and none are allowed. To enforce structured JSON output, use strict: true or Structured Outputs; to mandate a specific tool, describe the requirement in the prompt.

2. Conversation history must be append-only

Fable 5.1 binds thinking blocks to their preceding conversation prefix. Mutating prior messages, rebuilding system prompts, or altering the tools array in subsequent requests invalidates those blocks. For API accounts created on or after August 31, 2026, such requests return bound to a different conversation. Treat history as append-only: add mid-conversation system messages for reminders, use server-side compaction for context trimming.

3. Downgrading mid-conversation drops thinking blocks

Fable 5.1 can read thinking blocks from Fable 5, Opus 5, etc., but older models cannot read Fable 5.1's blocks. A router that falls back to an older model mid-session will lose those blocks. Additionally, Fable 5.1 safety refusals return HTTP 200 with stop_reason: "refusal"; the API does not automatically fail over — implement custom fallback logic.

Author's Usage Strategy

Sonnet 5: daily tasks — small features, routine bugs, tests, documentation lookup. Fast and cheap.

Opus 5: complex but well-scoped work — module refactors, critical code reviews, cross-file features. Anthropic also recommends starting here.

Fable 5.1: reserved for (a) tasks expected to run continuously for hours, (b) sparse clues requiring cross-log/code/doc investigation, (c) situations where the cost of a mistake far exceeds the model premium. Examples: production root-cause analysis, multi-service migrations, deep audits of critical systems, agent-driven research-to-delivery pipelines.

Conversely, simple SQL fixes, CRUD endpoints, or CSS tweaks run slower and cost more on Fable 5.1.

Data Retention Consideration

Fable 5.1 defaults to 30-day data retention. Zero data retention requires explicit Anthropic authorization. Do not migrate sensitive codebases wholesale without verifying compliance.

Conclusion

The decisive factor is whether the model can complete the second half of a long task: replan when early steps go wrong, recover from tool failures, and retain the original goal through hours of execution. Yet at 2× Opus 5's price, slower latency, and API compatibility, data retention, and fallback constraints, the author sets Opus 5 as default and calls Fable 5.1 only for the hardest, longest problems.

References

Release notes: https://www.anthropic.com/claude-fable-and-mythos-5-1
Fable 5.1 docs: https://platform.claude.com/docs/en/models/fable-5-1/overview
API changes: https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1
Migration guide: https://platform.claude.com/docs/en/models/fable-5-1/migration-guide
Plan details: https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan