November 2026 Claude Fable 5.2 Scores 65 on Artificial Analysis, Highest

Anthropic released Claude Fable 5.2 in November 2026 as a proprietary reasoning model operating with default fallback configuration and full multimodal comprehension. The new release secured the top rank across seventy-four tested systems on independent evaluation leaderboards, as documented in artificialanalysis.ai's report. The model pairs advanced multi-step analytical reasoning with a large input buffer designed to ingest text and high-resolution imagery simultaneously.

This analysis examines the verified intelligence metrics, token economics, evaluation suites, and deployment pathways for engineering leads and systems architects evaluating agentic integration. It outlines the balance between raw benchmark throughput and higher operational expenditure across production workloads.

01Frontier reasoning scores across seventy-four models

Claude Fable 5.2 registered an Intelligence Index score of 65 on version 5.0.1 of the Artificial Analysis suite, ranking first among seventy-four evaluated frontier models. This result places the system well above the category median score of 40. The index aggregates performance across nine benchmark environments designed to test terminal command execution, complex document synthesis, and scientific reasoning.

The primary evaluation battery incorporates AA-Briefcase version 1.1, AutomationBench-AA, Terminal-Bench 4.0, and FrontierCode 2.0 alongside AA-Omniscience version 1.1 and GDP.pdf. It also measures frontier reasoning using HLE-Diamond, the hard subset of CritPt version 2, and Terminal-Bench-Science 0.1. Anthropic structured the release as a reasoning architecture, though standard non-reasoning execution variants may also accompany the core tier.

Across auxiliary evaluations, the system recorded top ratings in the Finance and Accounting Index, which tracks financial reasoning across five core suites. The model also completed domain-specific tests including the Harvey LAB-AA version 1.1 legal benchmark, EnterpriseOps-Gym-AA business operations, and the ITBench-AA suite for Kubernetes incident root-cause analysis.

02Benchmark metrics and token expense breakdown

The model achieved four out of four rating units for intelligence while recording significant token consumption and running costs during evaluation. Input pricing sits at $8.00 per million tokens compared to the class median of $2.00, while output pricing reaches $40.00 per million tokens against a $10.00 category median. The provider provides a 98 percent cache discount to lower costs for repeated system prompts.

Verbosity metrics placed the system at rank seventy-two out of seventy-three tracked models due to its extensive reasoning chains. The model consumed 28.3 billion input tokens and generated 275 million output tokens during the test suite, producing an output-to-input ratio of 0.97 percent. The average cost per completed task on the Intelligence Index reached $26.54, positioning the model at rank seventy-three out of seventy-three for overall operational cost.

03Multimodal agentic workflows and context capacity

The architecture processes multimodal inputs consisting of text and visual media while generating structured text output. It incorporates a context window of one million tokens, providing capacity roughly equivalent to 1,500 standard A4 printed pages in size 12 Arial font. This window accommodates entire code repositories, architectural schemata, and extensive documentation sets inside a single execution context.

The expanded context window supports specialized evaluations including MLCR-AA medical long-context reasoning, MMMU-Pro visual parsing, and AA-AnalystAgent quantitative analysis over spreadsheets. In software development tasks, the engine handles end-to-end repository modification through FrontierCode 2.0 and automated terminal manipulation through Terminal-Bench 4.0.

In creative and operational benchmarks, the model was tested on AA-3D for interactive three-dimensional spaces, Code2Video-AA for programmatic video generation, and GDPval-AA version 2.1 for real-world enterprise deliverables. Artificial Analysis additionally measures conversational traits via AA-PersonalityBench, evaluating expressive language, scannable organization, answer transparency, and reduction of generic filler text.

04CLI installation commands and environment configuration

Developers deploying Claude Fable 5.2 within existing workflows can update their terminal tooling through standard package distribution channels. The official client suite receives regular updates supporting newly released reasoning targets and configuration flags.

npm install -g @anthropic-ai/claude-code
claude update

Once installed, users can inspect installed profiles and switch active sessions to the reasoning model directly from the command line interface. The runtime supports switching targets interactively within active sessions or passing explicit model flags upon initialization.

claude --model claude-fable-5-2
/model claude-fable-5-2

For persistent project configurations, engineers can define the default model via terminal environment variables or configuration files. Setting this environment variable ensures build agents, continuous integration tasks, and terminal instances default to the designated release without per-command flags.

export ANTHROPIC_MODEL=claude-fable-5-2
claude config set model claude-fable-5-2

05Public API availability and inference economics

Artificial Analysis currently reports no publicly accessible application programming interface endpoint for Claude Fable 5.2 on its benchmark tracking dashboard. While internal benchmark testing and partner evaluations have completed, production API routing and broad developer access remain restricted pending general availability schedules from the provider.

Teams planning enterprise adoption must weigh the model score of 65 against the weighted run cost of $26.54 per task. The 98 percent prompt caching discount represents a critical requirement for production pipelines, allowing operators to amortize repeated context across high-volume pipelines.

Until broad public endpoints launch, infrastructure teams must track inference times and per-task latency across reasoning units. The model delivers top-tier performance on complex agentic workflows, but high token output necessitates strict budget controls and token monitoring across downstream client applications.