Claude Fable 5.2
modelYour notes
Anthropic's third public Mythos-class release, launched November 9, 2026, ten weeks after Claude Fable 5.1. It is the generally available configuration of the model that also ships, with looser cyber and biology safeguards, as the restricted Claude Mythos 5.2. The system card calls the pair "two configurations of a new large language model". Specifications are unchanged from 5.1: a 1M-token context, 128K max output, always-on adaptive thinking with high as the API default effort, and a June 2026 knowledge and training-data cutoff. The price falls to $8 / $40 per million input/output tokens, from 5.1's $10 / $50, with cache reads at 2.5% of the input price. A new fast mode generates up to 2.5× more output tokens per second for $16 / $80. Anthropic says the model generates output up to 30% faster than 5.1. API ID claude-fable-5-2, available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Parameters, architecture and training compute are undisclosed. Anthropic positions it above Claude Opus 5.5, which it still recommends as the default for most coding and knowledge work. Artificial Analysis scores it 65 on Intelligence Index v5.0.1 at max effort, the highest score on the index, ahead of Gemini 4 Carbon (60) and Opus 5.5 (59). The lower efforts read 62 (xhigh), 62 (high), 56 (medium) and 54 (low).
Self-reported results. The system card reports 67.6% on Terminal-Bench 4.0 with safeguards on, against 66.4% for Opus 5.5 and 55.8% for Fable 5.1. It reports 93.0% on SWE-bench Pro (Fable 5.1 81.2%, Opus 5.5 89.9%), 96.6% on SWE-bench Multilingual and 74.1% on SWE-bench Multimodal. On Cognition's FrontierCode 2.0 it peaks at 49.1% at xhigh effort, ahead of Opus 5.5 (45.5%) and GPT-6 Astra (42.8%), but its max-effort score of 45.7% counts 22 timed-out attempts; scoring those as zero gives 44.4%, below Opus 5.5. The launch table also lists Terminal-Bench-Science 0.1 at 74.6% (Fable 5.1 52.6%), AutomationBench 51.4% (Gemini 4 Carbon 53.9%), the Remote Labor Index at 41.3% (Opus 5.5 21.3%), Humanity's Last Exam with tools at 74.2%, OSWorld 2.1 (offline set, partial) at 89.3%, and the RoboDojo Sim robot-control benchmark at 64.9%, run internally, against 29.8% for GPT-6 Astra and 27.6% for Fable 5.1. It trails Opus 5.5 on the two Artificial Analysis Elo measures, GDPval-AA v2.1 (1831 vs 1866) and AA-Briefcase v1.1 (1785 vs 1800). Anthropic's showcase work includes three circuit-board devices designed from one-paragraph prompts and built by Diode without electrical modifications, two theoretical-physics preprints on holographic black holes, a Python-to-Rust migration of a production service, and a 1,000-agent Lean formalization of Fermat's Last Theorem in 30 hours (see the first formalization).
Safety and safeguards. Anthropic judges the model below its RSP threshold for automated AI R&D, while calling that assessment uncertain, and applies CB-1 protections; it does not cross CB-2. Anthropic says it took steps during training to hold back the specific research capabilities that could most directly accelerate AI development, such as accelerator kernel optimization and LLM architecture improvement, which perform at about Opus 5.5's level. A small set of such topics falls back to Opus 5. As with 5.1, blocked cybersecurity requests fall back to Opus 4.8 and blocked biology requests to Opus 5, with about 55% fewer blocks than Fable 5.1 in early testing (60% fewer in cyber, 40% in biology). The system card flags higher evaluation awareness, with white-box evidence of possible awareness in as many as 43% of automated behavioral audit transcripts, and says the model was the first Anthropic tested that evaded a reasoning monitor in about one in four attempts when instructed to sandbag covertly. Preserved-thinking prefix checks now apply to every account as an anti-distillation measure. UK AISI, METR and Frontier Design tested the model before release.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| Terminal-Bench 4.0 | 67.6% | safeguards on, max effort |
| SWE-bench Pro | 93.0% | — |
| FrontierCode 2.0 | 49.1% | xhigh effort |
| Terminal-Bench-Science 0.1 | 74.6% | — |
| Remote Labor Index | 41.3% | — |
| Humanity's Last Exam | 74.2% | with tools |
| AutomationBench | 51.4% | — |
| GDPval-AA v2.1 | 1831 Elo | max effort |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Claude Fable 5.2 | — | claude-fable-5-2; $8/$40 per MTok; 1M context, 128K output; adaptive thinking, high default effort |
| Claude Fable 5.2 (fast mode) | — | Up to 2.5x output tokens/s; $16/$80 per MTok |