Anthropicプロプライエタリ

Claude Fable 5

Name: Claude Fable 5
Author: Anthropic

このモデルを比較

Anthropic開発の最高性能推論モデル。100万トークンのコンテキストウィンドウ、12.8万トークン出力対応。最も複雑な推論タスクと長時間のエージェント作業に最適。

シェア:X はてブ LINE

パラメータ

非公開

コンテキスト長

ライセンス

プロプライエタリ

リリース日

2026-06-09

日本語性能

✅高品質日本語

多言語対応モデルのうち、日本語処理に優れた性能を持つモデル。

API料金

入力料金（1Mトークンあたり）

出力料金（1Mトークンあたり）

課金モード: standard

強み

弱み

活用例

深度分析

Arena Elo

First place

Leading in Text, Code, and Agent Arena as of July 2026

SWE-Bench Pro

80.3%

vs GPT-5.5: 58.6%, Opus 4.8: 69.2%

Humanity's Last Exam (w/ tools)

64.5%

vs GPT-5.5: 52.2%, Opus 4.8: 57.9%

FrontierCode Diamond

29.3%

2x Opus 4.8 (13.4%), 5x GPT-5.5 (5.7%)

Context Window

1M tokens

With 128K output tokens

Input Price

$10/1M

Output: $50/1M (2x Opus 4.8)

強み

・State-of-the-art on complex coding benchmarks (SWE-Bench Pro: 80.3%) and long-horizon agentic tasks
・Exceptional vision capabilities: completed Pokémon FireRed with vision-only, rebuilds web apps from screenshots
・Superior memory and autonomous learning: 3x better performance improvement with persistent memory vs Opus 4.8

弱み

・Mandatory 30-day data retention policy blocks use in regulated industries requiring zero-retention
・Safety classifiers trigger on ~8% of tasks, silently rerouting to weaker Opus 4.8 model
・High cost ($10/$50 per million tokens) with 1.5-2x higher token consumption than predecessors

競合比較

Model	Arena	SWE	GPQA	Price
GPT-5.5	10th in Text Arena	58.6%	94.4% (Diamond)	$5/$30
Claude Opus 4.8	2nd in Text Arena	69.2%	85.7% (Diamond)	$5/$25
Gemini 3.1 Pro	7th in Text Arena	54.2%	92.4% (Diamond)	Not specified

概要

Claude Fable 5 represents Anthropic's first Mythos-class model available to general users, positioned as the strongest generally available AI model as of mid-2026. It sits above the Opus tier in capability, sharing its underlying architecture with the restricted Claude Mythos 5 model used by cybersecurity defenders. The model excels particularly in long-horizon, autonomous tasks where it demonstrates remarkable persistence and self-correction abilities, compressing what would be months of human engineering work into days. The release introduces a novel safety architecture where sensitive queries in cybersecurity, biology, chemistry, and AI model development are automatically rerouted to Claude Opus 4.8 through safety classifiers. This enables Anthropic to release unprecedented capabilities publicly while managing dual-use risks. However, this creates domain-specific capability limitations that users must account for, particularly in security and life sciences research where legitimate work may trigger false positives. Fable 5's positioning is uniquely specialized: it dominates benchmarks requiring sustained autonomous work across many steps (SWE-Bench Pro, FrontierCode, Hebbia Finance Benchmark), while GPT-5.5 maintains advantages in abstract reasoning puzzles (ARC-AGI-2) and extreme long-context performance beyond 512K tokens. The model's premium pricing and mandatory 30-day data retention policy make it a strategic tool for high-value, complex tasks rather than a general-purpose replacement for previous Claude models.

ベンチマーク＆性能

Claude Fable 5 demonstrates state-of-the-art performance across multiple benchmark categories, with particularly strong advantages in coding and knowledge work. Here are detailed benchmark comparisons: **Coding & Agentic Engineering:** | Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | |-----------|---------|----------|---------|----------------| | SWE-Bench Pro | 80.3% | 69.2% | 58.6% | 54.2% | | SWE-Bench Verified | 95.0% | 88.6% | - | - | | FrontierCode Diamond | 29.3% | 13.4% | 5.7% | - | | Terminal-Bench 2.1 | 88.0% | 82.7% | 83.4% | 70.7% | | CursorBench | 72.9% | 64.9% | - | - | **Knowledge Work & Reasoning:** | Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | |-----------|---------|----------|---------|----------------| | Humanity's Last Exam (no tools) | 59.0% | 49.8% | 41.4% | 44.4% | | Humanity's Last Exam (with tools) | 64.5% | 57.9% | 52.2% | 51.4% | | GDPval-AA (Elo) | 1932 | 1890 | 1769 | 1314 | | Hebbia Finance Benchmark | Highest | Below Fable | - | - | | GPQA Diamond | 85.7% | 80.2% | 94.4% | 92.4% | **Computer Use & Vision:** | Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | |-----------|---------|----------|---------|----------------| | OSWorld-Verified | 85.0% | 83.4% | 78.7% | 76.2% | | Blueprint-Bench 2 | 38.6% | 14.5% | 36.2% | 26.5% | | GDP.pdf | 29.8% | 22.5% | 24.9% | 16.7% | **Domain-Specific:** | Benchmark | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | |-----------|---------|----------|---------|----------------| | Legal Agent Benchmark | 13.3% | 10.4% | 2.1% | 0.0% | | HealthBench Professional | 66.0% | 56.9% | 51.8% | - | | ExploitBench | 78.0% | 40.0% | 34.0% | - | **Long-Context Performance:** GPT-5.5 leads on MRCR v2 at 512K-1M tokens (74.0% vs Fable 5's unpublished scores), while Fable 5 demonstrates superior practical memory usage in extended autonomous sessions. **Note:** Performance varies significantly by domain due to safety classifiers. On cybersecurity benchmarks with classifiers active, Fable 5 scores 0.0% on exploit development tasks by design.

詳細比較

**Claude Fable 5 vs GPT-5.5:** Fable 5 dominates in coding (80.3% vs 58.6% on SWE-Bench Pro) and agentic tasks, while GPT-5.5 leads in abstract reasoning (94.4% vs 85.7% on GPQA Diamond) and extreme long-context performance (74.0% on MRCR v2 at 512K-1M tokens). Fable 5 costs exactly twice as much per token ($10/$50 vs $5/$30) and has stricter safety classifiers that may reroute queries. GPT-5.5 offers more accessible tiered access for cybersecurity professionals through chatgpt.com/cyber, while Fable 5 requires vetted partner access for Mythos 5. For repository-level engineering and complex autonomous work, Fable 5 is superior; for security research and maximum cost efficiency, GPT-5.5 may be preferable. **Claude Fable 5 vs Claude Opus 4.8:** The same 2x pricing gap exists ($10/$50 vs $5/$25), but Fable 5 delivers significant capability leaps: +11.1 points on SWE-Bench Pro, +9.2 points on Humanity's Last Exam, and nearly 2.2x better scores on FrontierCode Diamond. However, Opus 4.8 offers zero data retention (Fable 5 requires 30-day retention), doesn't trigger safety classifiers, and is 25-30% faster on spreadsheet tasks per early user reports. Fable 5 is optimized for long-horizon autonomous work where it completes tasks in fewer turns, while Opus 4.8 remains superior for routine tasks and privacy-sensitive work. **Claude Fable 5 vs Gemini 3.1 Pro:** Limited direct comparisons are available, but Fable 5 shows large advantages in coding (80.3% vs 54.2% on SWE-Bench Pro), knowledge work (64.5% vs 51.4% on Humanity's Last Exam with tools), and vision tasks (38.6% vs 26.5% on Blueprint-Bench 2). Gemini 3.1 Pro holds a slight edge on GPQA Diamond (92.4% vs 85.7%). Pricing for Gemini 3.1 Pro isn't specified in available sources, but Google's models typically offer competitive pricing. Fable 5's superior memory and autonomous capabilities make it better for extended agent workflows, while Gemini may offer better value for less complex tasks.

コミュニティ評価

Early adopters and enterprise testers report transformative experiences with Fable 5, particularly for long-horizon tasks: **Enterprise Adoption:** - Stripe compressed months of engineering into days, completing a 50-million-line Ruby codebase migration in 24 hours that would have taken 2+ months manually. - GitHub noted Fable 5 handled complex, long-horizon coding tasks with autonomy exceeding previous benchmarks. - Cursor's CEO reported it as the state-of-the-art model on CursorBench, opening new classes of long-horizon problems. - Hebbia achieved the highest score on their Finance Benchmark for senior-level reasoning. **Developer Sentiment:** The developer community is enthusiastic but cautious about costs. Common observations include: - "Fable feels like a mature, calm, and down-to-earth programmer" - "I could give Fable 5 vague prompts and still get complete projects" - Users report completing entire applications from single high-level prompts - Anthropic's internal Parameter Golf challenge showed Fable 5 improving training pipelines 6x more than Opus 4.7 **Critical Feedback:** - Safety classifiers cause frustration: "The guardrails for Fable trip at the faintest hint of a security problem" - Cost concerns: Users on Max 20x plans report burning through 2% of allocation per minute during intensive use - Some researchers note limitations in frontier AI development work due to additional safeguards - The mandatory 30-day data retention policy is a blocker for regulated industries **Adoption Pattern:** Teams are implementing hybrid strategies—using Fable 5 for complex, high-value tasks while maintaining Opus 4.8 or Sonnet for routine work. The Advisor tool, released two months before Fable 5, enables cost optimization by letting cheaper models execute while Fable 5 provides oversight. Early data shows this reduces costs by 11.9% while improving performance by 2.7 percentage points on SWE-bench Multilingual.

ユースケース

1. **Large Codebase Migrations & Refactoring:** Fable 5 excels at understanding entire codebases and performing complex, multi-file changes. Stripe's 50-million-line Ruby migration demonstrates this capability. Choose Fable 5 over GPT-5.5 or Opus 4.8 when the task requires understanding dependencies across thousands of files and making coordinated changes that would take human engineers months. 2. **Autonomous Long-Horizon Research & Analysis:** For tasks requiring sustained work over days or weeks, such as scientific literature synthesis, financial analysis of complex datasets, or legal document review, Fable 5's memory and self-correction capabilities outperform alternatives. Hebbia's Finance Benchmark results and IMC's trading-analysis evaluations demonstrate this edge. Choose over GPT-5.5 when the work involves extended autonomous sessions rather than single-prompt reasoning. 3. **Vision-Intensive Development Workflows:** Fable 5's ability to complete Pokémon FireRed with vision alone and rebuild web apps from screenshots makes it ideal for UI/UX development, accessibility testing, or visual debugging. When tasks require understanding screenshots, diagrams, or visual interfaces as part of a larger development process, Fable 5 surpasses both GPT-5.5 and Opus 4.8. 4. **Complex Financial & Business Analysis:** For senior-level financial reasoning, trading analysis, or complex business problem-solving, Fable 5's scores on the Hebbia Finance Benchmark and ability to ace IMC's evaluations across factual lookup, conceptual reasoning, and expected-value analysis make it the top choice. Choose over GPT-5.5 when the analysis requires sustained attention to complex documents and tables rather than abstract reasoning puzzles.