Claude Fable 5.2

• Nov 9, 2026

  Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2, and cuts Fable's price 20%

Same model as the restricted Claude Mythos 5.2 (tracked separately), which shares its specs and pricing and differs only in safeguards. Specs, ids and pricing are from Anthropic's model page (platform.claude.com/docs/en/models/fable-5-2/overview): 1M context, 128K synchronous output, reliable knowledge cutoff June 2026, default effort high, cache reads $0.20/MTok (2.5% of input), 5-minute cache writes $10, 1-hour $16, batch 50% off; fast mode is $16/$80 for up to 2.5x faster output. Bedrock id anthropic.claude-fable-5-2; Google Cloud and Microsoft Foundry use claude-fable-5-2. Retirement is 'not sooner than November 9, 2027', a floor rather than a scheduled date. Benchmarks from the Fable 5.2 / Mythos 5.2 system card (Table 8.1.A, max effort, safeguards on, five-trial average): SWE-bench Pro 93.0%, HLE 73.2% no tools / 74.2% with tools, plus untracked FrontierCode 2.0 49.1% (xhigh), FrontierSWE v2 69.9%, SWE-bench Multilingual 96.6%, OSWorld 2.1 offline partial 89.3%, Remote Labor Index 41.3%, AutomationBench 51.4%, Terminal-Bench 4.0 67.6% and Terminal-Bench-Science 0.1 74.6%. GDPval-AA is reported only as v2.1 (1831), a different scale from the v2 this key records, so gdpvalAA stays null. ARC-AGI-2 93.8% is ARC Prize-verified at xhigh effort (arcprize.org/results/anthropic-claude-fable-5-2, 2026-11-09; max 92.5%, high 90.4%, medium 89.6%, low 86.7%). Cost per task and speed are Artificial Analysis' max-effort variant, the only one it lists: Intelligence Index 65, $26.54 per task, 103 tok/s, 231.55 s to first answer token, 275M output tokens across the index. predecessorId is null: Anthropic's launch post and migration guide describe moving from Fable 5.1 but never say 5.2 replaces it, and Fable 5.1 stays available as a legacy model. Anthropic's launch post says that in prerelease testing the model helped complete otherwise-infeasible software migrations, designed three working electronic devices, solved two open problems in black-hole physics and formalised Fermat's Last Theorem in 30 hours (a 1,000-agent team on an earlier snapshot, per the system card). The launch table also gives RoboDojo Sim 64.9% (Fable 5.1 27.6%), WANDR relaxed 82.1%, Chartography no tools 73.4%, AA-Briefcase v1.1 66.1% rubric / 1785 Elo (Opus 5.5 leads at 1800), and AutomationBench 51.4%, where Gemini 4 Carbon leads at 53.9%.