Claude Fable 5.1: Same Sticker, Cheaper Cache

Anthropic released Claude Fable 5.1 on September 1, 2026 as the generally available successor to Fable 5. Fable 5.1 is not a cheaper list. It is the same $10 / $50 with cache hits cut to a quarter of Fable 5. If your Fable bill is mostly cache reads (agent loops, Claude Code), that is the hop.

The official table is terminal, computer use, and HLE. It is not a SWE dump. Anthropic says Fable 5.1 is the most advanced generally available model for coding and knowledge work. Hold the launch table and the cache multiplier. Fable 5 is now the expensive twin for cache-heavy work.

At a Glance

• Catalog id: claude-fable-5-1
• Organization: Anthropic
• Release: September 1, 2026. Fable 5 is legacy.
• API: claude-fable-5-1 (Bedrock anthropic.claude-fable-5-1)
• Pricing: $10 / $50 per 1M; cache hit $0.25; 5m write $12.50; 1h write $20
• Context: 1,048,576 input tokens / 128K max output
• Modalities: Text + image in, text out. Adaptive thinking always on.
• Effort defaults: High in Claude Code; medium in Claude Cowork and on Claude.ai
• Hosting: First-party Anthropic in our catalog
• Mythos 5.1: Same weights, trusted access, not this post

What's New

Same sticker, 4× cheaper cache hits. The official agent and terminal table is the capability story: Terminal-Bench-Science 0.1 more than doubles Fable 5 (52.6 vs 24.7). Cyber and biology classifiers are more precise: about 60% fewer cyber false positives and 85% fewer elementary or medical biology false positives versus Fable 5 launch safeguards, while research biology still routes to Opus. Fable 5.1 may discover source-code vulnerabilities; pentest, exploit generation, and binary vuln scanning still fall back to Opus. Enterprise Frontier Safeguards (EFS) ship later this fall; until then, eligible customers can bridge with zero data retention.

Knowledge cutoff is June 2026. Adaptive thinking is always on. Tool use, batch, and structured output are supported. Low and medium effort can match or beat Fable 5 at lower cost on some workloads.

Benchmarks

Anthropic published a self-reported launch comparison with production safeguards on. The rows that matter for this release are terminal, computer use, HLE, automation, and CursorBench. Scores below are not LLM Stats verified.

Terminal-Bench-Science 0.1 at 52.6 vs Fable 5's 24.7 is the row that moves. Standard error is about 3.5 to 4.5 points (call it about 4 points in prose). That is still a different band from Opus 5's 29.0 on the same Anthropic setup. The public leaderboard (3 trials, Claude Code harness) has Opus 5 at 30.0 and Fable 5 at 21.4; Anthropic's setup reproduces 29.0 and 24.7, within noise.

Terminal-Bench 4.0 lands at 55.8versus Fable 5's 42.0 and Opus 5's 52.3. Mythos 5.1's Terminal-Bench 4.0 score of 60.9 is not the public Fable number. CursorBench 3.2.0 at 73.4 is a small hop over Fable 5 (70.5) and Opus 5 (70.0). AutomationBench jumps to 31.4 from Fable 5's 17.1. HLE is near-flat versus Fable 5 (60.9 / 65.0 with tools vs 57.8 / 63.8).

OSWorld 2.0 is the authors' August 2026 task release; Fable 5 and Opus 5 were re-run under the same conditions. Fable 5.1 posts 41.7 strict / 77.9 partialversus Fable 5's 36.1 / 72.9 and Opus 5's 39.6 / 75.4. Those numbers are not comparable to older OSWorld 2.0 figures, which is why the table skips an extra competitor column. GDPval-AA v2 is 1853 Eloversus Fable 5's 1723 and Opus 5's 1824; Elo, not percent, so it stays in prose.

Harness caveats

Safeguards are on. On OSWorld, safeguard interventions scored zero for Fable 5.1 and Fable 5. On AutomationBench, Fable 5 was zeroed for interventions. Other cyber interventions completed under Opus 4.8; biology under Opus 5. That likely reduces the public Fable scores. Protein binder and GPU kernel demos belong to Mythos 5.1, not the public Fable SKU. Anthropic also showed a Venus elevation map from Magellan radar as a Fable 5.1 demo (CC-licensed), not a bench.

Pricing

First-party Anthropic list, USD per million tokens. Batch is half ($5 / $25). US-only inference is 1.1× on all token categories for Claude 4.6+ and Fable 5.1. You pay Fable rates only when the request stays on Fable; cyber and biology fallbacks bill at Opus prices.

Default retention on Fable is 30 days. Eligible enterprises can use zero data retention until EFS arrives later this fall with customer-controlled cloud storage.

When to Use It

• Good fit: Cache-heavy agent loops, Claude Code at high effort, long-running knowledge work, and shops already on Fable 5 who were paying $1 cache hits.
• Prefer Opus 5 ($5 / $25) when you want a cheaper sticker and do not need the Fable-class table.
• Prefer Opus 4.8 or Opus 5 when the task is dual-use cyber or research biology; you will get them anyway via fallback.
• Mythos 5.1 only if you are in CVP or LSVP trusted access.
• Do not stay on Fable 5 for cache-heavy work.

Caveats

• Safeguards are on, so some interventions score zero and cyber/bio false-positive rates still route hard cases to Opus.
• Terminal-Bench-Science error bars are about 4 points; treat 52.6 as a band, not a point estimate.
• OSWorld 2.0 here uses August 2026 tasks; do not mix with older OSWorld numbers.
• Default retention is 30 days unless you are on ZDR or later EFS.
• Adaptive thinking is always on; there is no extended-thinking toggle.
• Knowledge cutoff is June 2026.

Outlook

This release ages when LLM Stats verifies Terminal-Bench-Science and Terminal-Bench 4.0, when EFS ships, when the $0.25 cache hit rate spreads to the rest of the family, or when a public SWE table appears. Through-line: same sticker, cheaper cache, agent table not SWE.

For the full announcement, pricing, and system card, see Anthropic's Fable and Mythos 5.1 launch, the Fable product page, pricing docs, and the system card.