Anthropic

Claude Fable 5

Released Jun 9, 2026 · knowledge cutoff 2026-01

Status: Frontier
Location: United States
Modality: Multimodal
Context window: 1M
Max output: 128K
Speed (max effort): 74 tok/s · 107s to first answer token
Price ($/MTok in / out): $10 / $50
Cost per task (max effort): $2.75
Open weights: No

Benchmarks5 of 8 reported

GPQA Diamond

92.6%

SWE-bench Verified

95%

Terminal-Bench 2.1

88%

Humanity's Last Exam

64.5%

LMArena Elo

1509

MMLU-Pro

—

AIME (latest)

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Holds decision quality steady across sessions that run for days, remembering why it chose something hours earlier
Plans in stages, delegates to sub-agents and checks its own work unprompted
Vision good enough to rebuild an interface from a screenshot
Acts rather than narrates — little time spent describing what it is about to do

Weaknesses

Keeps working until something stops it, so agent loops need explicit stop rules or the cost runs away
Slow, and its depth of explanation is a cost when you wanted a quick scan
Token spend and latency swing enormously between low and max effort, making budgeting hard

Shares weights with the restricted Claude Mythos 5 (vetted partners only). HLE 64.5% is with tools; GPQA/SWE figures are third-party — Anthropic retired saturated benchmarks from official reporting. Terminal-Bench 2.1 88.0% as reported in poolside's Laguna S 2.1 comparison table.

For developers

API model strings

anthropicclaude-fable-5
aws-bedrockanthropic.claude-fable-5

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor