Anthropic

Claude Fable 5.1

Released Sep 1, 2026 · knowledge cutoff 2026-06

Status: Frontier
Location: United States
Modality: Multimodal
Context window: 1M
Max output: 128K
Speed (medium effort): 47 tok/s · 6.77s to first answer token
Price ($/MTok in / out): $10 / $50
Cost per task (medium effort): $1.00
Open weights: No

Benchmarks7 of 10 reported

MMLU-Proretired

92.4%

SWE-bench Pro

81.2%

Terminal-Bench 2.1

91.4%

Humanity's Last Exam

60.9%

LMArena Elo

1498

GDPval-AA v2

1853

ARC-AGI-2

90%

GPQA Diamond

—

SWE-bench Verified

—

AIME (latest)retired

—

Strengths & weaknesses

Strengths

Does the task it was given and stops, instead of treating every ambiguity as more work to take on — the scope creep Fable 5 was known for is mostly gone
Code review is quieter and sharper: nitpick comments largely disappear while it still catches the same real issues
More comfortable left running unattended for hours than Fable 5, and fixes root causes rather than patching symptoms
Prose sounds less stereotypically Claude — denser, plainer, with far less bold-and-bullet scaffolding

Weaknesses

Slower in agent loops: it often issues one tool call per turn where Fable 5 batched several, and code reviews take about half again as long
At max effort it generates roughly twice the tokens of its peers and can sit for minutes before the first answer token, so per-task cost went up despite the cache-read cut
At low effort it answers from memory rather than searching, tends to rewrite whole files for small edits, and can lift passages from sources without marking them as quotes

Same model as the restricted Claude Mythos 5.1 (Project Glasswing participants only), which shares its specs and pricing and differs only in safeguards — Anthropic reports no separate Mythos score on any tracked benchmark. HLE 60.9% no tools / 65.0% with tools per Anthropic. Anthropic no longer reports SWE-bench Verified, GPQA Diamond or AIME: Terminal-Bench 2.1 91.4% is Artificial Analysis-measured at max effort (Vals AI's harness gives 85.0%), MMLU-Pro 92.4% is Vals AI's measurement at max effort, and ARC-AGI-2 90.0% is ARC Prize-verified at max effort ($4.49/task). Cache reads are $0.25/MTok (0.025x input, vs 0.1x on other Claude models); batch $5/$25. Cost per task and speed are AA's medium-effort variant — max effort runs $3.69/task at 66.2 tok/s with a 291 s first-answer latency. Anthropic's launch table also reports Terminal-Bench-Science 0.1 52.6%, Terminal-Bench 4.0 55.8% (Mythos 5.1 60.9%), GDPval-AA v2 1853, OSWorld 2.0 77.9% partial / 41.7% strict, AutomationBench 31.4% and CursorBench 3.2.0 73.4%, with Fable 5.1 scored with production safeguards on, taking zeros where they intervened. LMArena Elo 1498 is the "claude-fable-5.1-max" text-arena listing (arena.ai, 2026-09-13 update, rank 5). SWE-bench Pro 81.2% is Anthropic's system-card figure (not in the launch post), as relayed by BenchLM, CodingFleet and emergent.sh; the system-card PDF was too large to read directly this run.

For developers

API model strings

anthropicclaude-fable-5-1
aws-bedrockanthropic.claude-fable-5-1
google-vertexclaude-fable-5-1

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

Replaces Claude Fable 5

News