99.9% SLA
Claude Fable 5: Model Introduction & Practical Guide
Claude Fable 5 is Anthropic's public frontier: a Mythos-class model that shares its underlying system with the restricted Claude Mythos 5 and adds conservative safety guardrails. It posts SOTA results across nearly all public benchmarks - SWE-Bench Pro at 80.3%, FrontierCode 29.3% at medium effort - and demonstrates capabilities that read like research milestones: migrating a 50-million-line Ruby codebase in a day, playing through Pokémon FireRed from raw screenshots, and sustaining million-token tasks with persistent file-level memory.
Here's the short version: Fable 5 is what frontier capability looks like when it's packaged for general availability. The "Fable" tier sits above Opus in Anthropic's hierarchy, with an independent classifier that quietly routes sensitive requests (cybersecurity, biosecurity, distillation attempts) to Claude Opus 4.8 rather than refusing them - about 5% of traffic triggers downgrade. The model's signature strengths are long-horizon autonomy, persistent memory across extended work, and visual reasoning that works from raw pixels. Pricing is premium: $10/$50 per million tokens. The trade-offs: restricted-domain requests get downgraded by design, the absolute model (Mythos 5, without guardrails) is limited to vetted institutions, and 30-day monitoring/data retention policies apply.
This guide covers model overview, core features, technical specifications, capability comparison, core advantages, recommended use cases, example prompts, and selection recommendations.
Quick Facts
| Attribute | Value |
|---|---|
| Model name | Claude Fable 5 |
| Developer | Anthropic |
| Category | Mythos-class frontier model (public, guardrailed) |
| Base | Shares underlying system with Claude Mythos 5 |
| Safety model | Independent classifier; sensitive requests routed to Opus 4.8 (~5% of traffic) |
| Key benchmarks | SWE-Bench Pro 80.3% · FrontierCode 29.3% (medium effort) |
| Autonomy | Multi-day tasks; 50M-line Ruby codebase migration in a day |
| Memory | Million-token tasks with file-level persistent notes |
| Visual reasoning | Beats Pokémon FireRed from raw screenshots; chart value extraction; webpage reconstruction |
| Pricing | $10 / 1M input · $50 / 1M output |
| Access | Claude app/API (claude-fable-5), Claude Code |
Table of Contents
- Model Overview
- Core Features
- Technical Specifications
- Capability Comparison
- Core Advantages
- Recommended Use Cases
- Example Prompts
- Selection Recommendations
- FAQ
- Sources & Further Reading
1. Model Overview
Anthropic introduced Fable as a new capability tier: Mythos-class, positioned above Opus. Fable 5 and Mythos 5 share the same underlying model - the difference is packaging. Mythos 5, without meaningful guardrails, is limited to vetted institutions (security and life-science organizations); Fable 5 is the version the public can use, with a conservative safety layer and standard data policies.
The safety mechanism is worth understanding because it shapes what the model does. Rather than refusing sensitive requests outright, Fable 5 runs an independent classifier that detects cybersecurity, biosecurity, or distillation intent and routes those requests to Claude Opus 4.8 - roughly 5% of traffic. For frontier LLM development requests specifically (pretraining pipelines, distributed training infrastructure), invisible interventions - prompt modification, steering vectors, parameter-efficient fine-tuning - limit effectiveness, affecting about 0.03% of traffic. Practically: the model rarely says "no," but a small slice of requests silently lands on a different model.
The capability demonstrations are the strongest published record of frontier-agent behavior anywhere. In software engineering: a full migration of a 50-million-line Ruby codebase completed in one day; SWE-Bench Pro 80.3%; FrontierCode 29.3% at medium effort (the highest frontier-model score, at lower cost than peers). In visual reasoning: completing Pokémon FireRed from raw screenshots without external tools, extracting precise values from scientific charts, and reconstructing webpage source from images. In memory and autonomy: sustained million-token tasks using file-level notes it writes and consults itself, with Slay-the-Spire performance reportedly 3x Opus 4.8, and tasks that run for hours or days without human intervention.
2. Core Features
Ultra-long autonomous coding. Day-scale migrations and development sessions; SWE-Bench Pro 80.3%.
Token-efficient frontier reasoning. FrontierCode 29.3% at medium effort - top scores at lower cost than fast-mode peers.
Native visual understanding. Raw-screenshot gameplay, chart value extraction, and image-to-source reconstruction.
Persistent memory. File-level notes written and reused across million-token tasks to prevent context amnesia.
Intelligent safety downgrade. Classifier-based routing of sensitive requests to Opus 4.8 rather than hard refusals.
Long-context continuity. Million-scale token tasks with reference-quality coherence across hours of work.
3. Technical Specifications
| Specification | Detail |
|---|---|
| Tier | Mythos-class (above Opus) |
| Underlying system | Shared with Claude Mythos 5 |
| Safety routing | Classifier detects cyber/bio/distillation → routes to Opus 4.8 (~5% of traffic) |
| Restricted-domain interventions | Prompt modification, steering vectors, PEFT on ~0.03% of traffic |
| SWE-Bench Pro | 80.3% |
| FrontierCode (medium effort) | 29.3% |
| Memory | Million-token tasks; self-authored file-level notes |
| Pricing | $10 / 1M input · $50 / 1M output |
| Data policy | 30-day monitoring/retention |
4. Capability Comparison
| Dimension | Claude Fable 5 | Claude Mythos 5 | Claude Opus 4.8 |
|---|---|---|---|
| Underlying model | Same | Same | Different |
| Guardrails | Conservative (~5% downgrade) | None/minimal | Standard |
| ExploitBench | Near Opus | 78.0% | ~40.0% |
| Protein design validation | Restricted | 9/14 in pipeline | None |
| Access | Public | Vetted institutions | Public |
| Data retention | 30-day monitoring | 30-day monitoring | Standard policy |
Positioning read. Fable 5 is the most capable publicly available Claude, with deliberate carve-outs for the domains where unrestricted frontier capability is risky. Its practical competitors are GPT-6 Astra (similar "frontier packaged for access" positioning with a Daybreak program) and Claude Fable 5.1, its own successor - which improves cost and cache economics on top of this capability base.
5. Core Advantages
- Mythos-class capability, publicly available. The strongest general-access Claude at launch.
- Day-scale autonomy. Codebase migrations and long development sessions without supervision.
- Persistent memory that works. File-level notes solve the long-task amnesia problem.
- Visual reasoning from raw pixels. Gameplay, charting, and source reconstruction without tool support.
- Graceful safety handling. Sensitive requests get routed, not rejected - fewer workflow interruptions.
- Token-efficient frontier scores. FrontierCode leadership at medium effort with relatively lower cost.
6. Recommended Use Cases
- Massive codebase migrations: multi-million-line refactors in single autonomous sessions.
- Frontier software engineering: the hardest debugging, architecture, and implementation work.
- Scientific analysis: chart and figure interpretation with precise value extraction.
- Long-horizon agent tasks: multi-hour to multi-day autonomous workflows with persistent memory.
- Visual verification workflows: screenshot-based QA, UI reconstruction, and gameplay testing.
- Research and analysis at the frontier: tasks that other models defer.
7. Example Prompts
1. Day-scale migration
2. Visual grounding task
3. Scientific figure extraction
4. Long-horizon agent
5. Visual QA loop
8. Selection Recommendations
Choose Claude Fable 5 if:
- Your tasks hit the frontier: massive migrations, hardest engineering, novel research.
- You need multi-hour/day autonomy with reliable memory.
- Visual reasoning from raw images is part of the workflow.
- You want frontier capability without applying for restricted access.
Choose Mythos 5 (if eligible) when:
- Your organization is a vetted security/life-science institution needing unrestricted frontier capability.
Choose Fable 5.1 instead when:
- You want lower cache and workload costs with the same capability class.
- Your agents hammer caches (75% cache-read cost reduction matters).
Sources & Further Reading
- Claude Fable 5 - AI Tools Directory (Chinese overview)
- Claude Fable 5 & Mythos 5 - Anthropic
- Claude models - Anthropic
Capabilities, guardrail behavior, and pricing are as published at review time; Anthropic's tier structure evolves. Restricted-domain routing means some requests execute on different models - validate outputs for compliance-critical work.
