
Claude Fable 5.2 Released: Specs, Benchmarks, and When to Upgrade
On November 9, 2026, Anthropic released Claude Fable 5.2, its latest Mythos-class reasoning model designed for long-running agentic tasks and technical workflows. Launched alongside the restricted-access Claude Mythos 5.2, Fable 5.2 serves as Anthropic's flagship generally available model for software engineering, financial modeling, and scientific analysis.
The release brings a 20% price cut compared to Fable 5.1, an expanded 1-million-token context window, and improvements across multi-step autonomous benchmarks. However, with a pricing structure that sits at twice the rate of Opus 5.5, developers must evaluate whether their workloads genuinely require autonomous deliberation or if standard daily-driver models remain the economical choice.
Core Specifications and Context Limits
Fable 5.2 is designed specifically for autonomous execution across complex codebases and messy monorepos. Anthropic structured the release to support persistent tool use and long agent sessions without context fragmentation.
- Model ID:
claude-fable-5-2 - Context Window: 1,000,000 tokens
- Maximum Output: 128,000 tokens
- Knowledge Cutoff: June 2026
- Thinking Architecture: Adaptive extended thinking is always on, supporting reasoning effort levels from low to maximum (defaulting to high) with thinking states preserved across conversational turns.
- Speed: Anthropic reports output generation speeds up to 30% faster than Fable 5.1, alongside a Fast tier that runs up to 2.5 times faster.
For enterprise deployments requiring administrative oversight before spinning up subagent clusters, teams can refer to our guide on how Anthropic adds IP allowlists and admin API member sign-out controls to protect organizational access.
Benchmark Performance: Fable 5.2 vs. Fable 5.1 and Opus 5.5
Anthropic positions Fable 5.2 as an architectural leap over Fable 5.1, outperforming Opus 5.5 in multi-step agentic benchmarks.
In official vendor evaluations (run at maximum available reasoning effort unless otherwise noted):
| Benchmark | Fable 5.2 | Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% (xhigh) | 32.4% (xhigh) | 45.5% |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% |
| AutomationBench | 51.4% | 31.4% | 42.5% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% |
| Remote Labor Index | 41.3% (high) | 17.9% | 21.3% |
| OSWorld 2.1 (offline) | 89.3% | 83.5% | 87.2% |
| Humanity's Last Exam (tools) | 74.2% | 65.6% | 67.7% |
The largest gains occur in agentic execution. On the Remote Labor Index, which tracks multi-step digital work, Fable 5.2 scored 41.3%, more than double Fable 5.1's 17.9% and nearly double Opus 5.5's 21.3%. On independent evaluations, Artificial Analysis placed Fable 5.2 at the top of its Intelligence Index at maximum effort with a score of 65, compared to 59 for Opus 5.5.
However, frontier reasoning has limits. In Google's competitive domain, Gemini 4 Carbon still edged past Fable 5.2 on AutomationBench (53.9% vs. 51.4%). Furthermore, maximum-effort agent evaluations are computationally intensive; Artificial Analysis measured tasks averaging 47.8 minutes and $26.54 per run at top reasoning tiers.
Fable 5.2 vs. Claude Mythos 5.2
Claude Fable 5.2 shares its foundational weights with Claude Mythos 5.2, but the two differ in safety filtering and availability:
- Safety Classifiers and Fallbacks: Fable 5.2 is generally available on AWS Bedrock, Google Vertex AI, Microsoft Foundry, and the Claude Platform. When requests trigger Anthropic's sensitive-domain boundaries (such as offensive cybersecurity or chemical/biological synthesis), Fable 5.2 falls back automatically to less autonomous legacy tiers (Opus 4.8 for security queries, Opus 5 for biology). Anthropic noted that Fable 5.2 experiences roughly 55% fewer false-positive blocks than Fable 5.1.
- Mythos Restricted Access: Mythos 5.2 removes dual-use task guardrails for authorized teams and powers Claude Security for Claude Enterprise accounts. Direct access remains restricted to organizations verified under Anthropic's Cyber and Life Sciences verification programs.
- Data Retention Rules: Both versions fall under Anthropic's Covered Model policy, enforcing a default 30-day data retention standard for safety monitoring, with zero data retention available only as a transition program for qualifying enterprise contracts.
API Pricing and Cost Efficiency
Anthropic lowered token rates for Fable 5.2 by 20% compared to Fable 5.1:
- Input Tokens: $8.00 per million tokens
- Output Tokens: $40.00 per million tokens
- Prompt Caching Read: $0.20 per million tokens
For comparison, Fable 5.1 launched at $10.00 input and $50.00 output per million tokens. While cheaper than its predecessor, Fable 5.2 remains priced at roughly double the per-token rate of Claude Opus 5.5. If you are deciding whether your team needs this level of compute, see our detailed breakdown of GPT-6 Astra vs Claude Fable 5.1 for context on frontier model operational economics.
Because adaptive thinking outputs count as generated tokens, prompt structure directly dictates API consumption. Reviewing how to write better AI prompts can keep tasks scoped properly to avoid runaway recursive loops during unattended agent runs.
When Should You Use Fable 5.2?
Anthropic still recommends Opus 5.5 as the default workhorse for standard coding, content processing, and knowledge retrieval. Fable 5.2 should be reserved for specific operational environments:
- Autonomous Codebase Refactoring: Tasks involving dependency resolution across multi-repo architectures where an agent must run tests, observe failures, and rewrite modules iteratively.
- Long-Horizon Research Loops: Workflows requiring hill-climbing metrics, parameter tuning, or validating reproducibility over hours without human intervention.
- Complex Data Extraction: Scenarios handling hundreds of disparate source files where the 1M context window and high reasoning effort are needed to reconcile conflicting data points.
For everyday tasks, drafting, and interactive chat, Opus 5.5 remains far more cost-effective. Fable 5.2 is an infrastructure upgrade aimed squarely at autonomous task completion rather than conventional conversational assistance.


