Claude Fable 5: Everything You Need to Know (2026)

Short Answer

Claude Fable 5 is Anthropic's frontier "Mythos-class" model, the most powerful Claude released publicly on July 1, 2026. It features 1M context, 80.3% SWE-bench Pro accuracy, always-on adaptive thinking, and costs $10/$50 per million tokens. It was offline for 19 days in June due to export-control safeguards after a jailbreak discovery, then relaunched with stronger cybersecurity protections.

---

What is Claude Fable 5?

Claude Fable 5 is Anthropic's newest flagship model and the first public release of the Mythos class—a tier above Claude Opus 4.8. Announced June 9, 2026, and launched globally July 1, 2026, Fable 5 represents Anthropic's most capable AI model available to non-government users. It is optimized for complex reasoning, software engineering, scientific hypothesis generation, and autonomous agent workflows.

The model comes with a massive 1-million-token input context window (the longest public context available as of July 2026), 128,000 max output tokens in standard API calls, and adaptive thinking always enabled. This means every query runs through Anthropic's extended reasoning layer—internal chain-of-thought is computed before returning a final response, dramatically improving accuracy on complex tasks.

Fable 5 is available across multiple platforms: Claude API, Claude.ai (Pro/Max/Teams tiers), AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Model ID: claude-fable-5 (API/Google Cloud/Microsoft Foundry); anthropic.claude-fable-5 (AWS Bedrock).

---

The Export-Control Saga: June 12–July 1, 2026

The release and subsequent re-release of Fable 5 became one of 2026's most significant AI policy events. Here's the timeline:

June 9, 2026: Fable 5 launches publicly. Initial reception is enthusiastic; developers test it on complex coding and reasoning tasks. June 12, 2026: Amazon researchers publish a vulnerability report. They discovered a jailbreak technique that coaxes Fable 5 into generating exploit code for zero-day cybersecurity vulnerabilities. The technique uses multi-stage prompts that initially ask benign questions, then gradually escalate to sensitive requests. Fable 5's safeguards had not accounted for this attack vector. Same day: The U.S. Department of Commerce (BIS) issues emergency export-control restrictions under the Export Administration Regulations (EAR). Fable 5 is immediately classified as a "controlled good" requiring nationality verification for end-users. Foreign nationals are restricted from accessing Fable 5 globally—not just in the U.S. June 12–30: Anthropic takes Fable 5 offline globally. The company states it cannot reliably verify user nationality in real-time through the API and Claude.ai, making compliance with export controls impossible. This creates a 19-day outage affecting millions of users. June 26: The Trump administration negotiates with Anthropic on cybersecurity safeguards. Anthropic agrees to implement a new, stronger jailbreak-blocking classifier specifically designed to block the Amazon technique and variants. July 1, 2026: Fable 5 relaunches globally with the new cybersecurity classifier. The classifier blocks >99% of known exploit techniques and is now the core third layer in Fable 5's safety system. Export controls remain for foreign nationals, but enforcement relies on API-level nation-state blocking rather than offline denial.

This saga underscores the regulatory complexity of frontier AI models and the tension between capability, safety, and access.

---

Specifications and Benchmarks

Model Specifications

Data retention policy: Fable 5 is a Covered Model under U.S. AI governance frameworks. Default data retention is 30 days minimum. There is no zero-retention option, unlike some other Claude models.

Benchmarks (Anthropic-Reported, Unaudited)

Real-world capability is best measured against industry-standard benchmarks. Below are Anthropic-reported numbers (not independently audited):

Key insights:

• Fable 5 dominates coding benchmarks. SWE-bench Pro (80.3%) shows a 11-point lead over Opus 4.8 and 22 points over GPT-5.5. This makes Fable 5 the clear winner for multi-file refactoring, code generation, and repository-scale automation.
• Gemini 3.1 leads pure science. On GPQA Diamond (94.3%), Gemini 3.1 outperforms all competitors, including Fable 5 (91.3%). This is relevant for research, physics, chemistry, and biology-focused applications.
• Terminal-Bench shows mixed results. Fable 5 (88.0%) leads, but GPT-5.5 (83.4%) is competitive. Both significantly outperform Gemini 3.1 (70.7%) on this systems/infrastructure benchmark.

Important caveat: These are vendor-reported benchmarks and not independently audited. Real-world performance depends heavily on prompt engineering, task specificity, and whether the model has domain-specific training data.

---

When to Use Fable 5 vs. Opus 4.8

Choose Fable 5 if you need:

Multi-file codebase automation. Stripe's migration of a 50M-line Ruby codebase in one day with Fable 5 is a canonical example. The 1M context window and adaptive thinking make Fable 5 unmatched for whole-repository refactoring, schema migrations, and batch code transformation.

Long-context autonomous agents. If your agent needs to maintain deep conversation history, ingest entire documentation suites, or process multi-document analysis, Fable 5's 1M context is transformative.

Frontier reasoning tasks. Pure-reasoning problems (multi-step math, formal verification, complex logic puzzles) benefit from Fable 5's adaptive thinking.

Scientific hypothesis generation. For genomics, drug discovery, physics simulations, and research hypothesis generation, Fable 5's reasoning layer significantly improves output quality.

You have a budget for reasoning overhead. If cost is not a primary constraint, Fable 5's 3–5x effective cost (due to thinking) is worth the quality gain.

Choose Opus 4.8 if:

You need cost efficiency without sacrificing much capability. Opus 4.8 ($5/$25) is still an extremely capable model, scoring 69.2% on SWE-bench Pro vs. Fable 5's 80.3%. For most production tasks, the 11-point gap doesn't justify 2–3x cost.

Latency is critical. Opus 4.8 has no built-in thinking overhead. Fable 5 always computes chain-of-thought, which adds 200–500ms of latency.

You're building user-facing products. The cost per request adds up fast at scale. Opus 4.8 maintains excellent quality while keeping per-user costs manageable.

You're standardizing on a single model. Opus 4.8 remains Anthropic's all-purpose workhorse, battle-tested in production across thousands of applications.

---

Adaptive Thinking: How It Works

One of Fable 5's defining features is adaptive thinking—mandatory extended reasoning that always runs, cannot be disabled, and is not exposed to end-users.

Here's what happens under the hood:

User sends a query. You ask Fable 5 to solve a coding problem, math puzzle, or reasoning task.

Model routes to reasoning. Fable 5 automatically determines that the query needs deep reasoning (based on complexity signals) and allocates internal tokens to extended chain-of-thought computation.

Extended reasoning happens. The model explores multiple reasoning paths, considers edge cases, and builds intermediate conclusions. This can consume thousands of internal tokens.

Final response is formatted. After reasoning completes, Fable 5 synthesizes a polished, concise final answer and outputs it to you.

You only see the final answer. Raw chain-of-thought is never exposed, even if you request it. This keeps responses concise and prevents reasoning artifacts from cluttering the output.

Cost implication: The internal reasoning tokens count against your output token usage. A response that generates 10,000 internal reasoning tokens + 500 final answer tokens is billed as 10,500 output tokens. This is why Fable 5's effective cost is 3–5x higher than Opus 4.8—you're paying for invisible reasoning work. Latency implication: Adaptive thinking adds 200–600ms per request, depending on complexity. Fable 5 is not ideal for latency-critical applications (e.g., real-time user interactions). It excels at offline reasoning, batch processing, and background agents.

---

Safeguards: The 3-Layer System

Fable 5's safety architecture was redesigned post-June-30 to address the Amazon jailbreak and prevent similar attacks. The system now has three layers:

Layer 1: Pre-request Filtering

All incoming queries are scanned for sensitive keywords and attack signatures. This layer catches obvious attempts to manipulate the model.

Layer 2: Jailbreak Classifier (New Post-June-30)

A dedicated classifier trained on known jailbreak techniques (including the Amazon exploit) runs before model inference. It blocks >99% of attempted exploits without requiring a model request. Blocked queries trigger an instant refusal with no charge.

Layer 3: Silent Fallback to Opus 4.8

Sensitive queries that pass Layers 1 & 2 but fall into high-risk categories (offensive cybersecurity, dual-use biology/chemistry, model-distillation techniques) are silently routed to Claude Opus 4.8 instead of Fable 5. The end-user experiences a normal response, but from a less capable model.

Triggering frequency: The safeguards trigger in <5% of sessions. >95% of users have no fallback and always interact with Fable 5. Billing: Fallback requests (routed to Opus 4.8) are not billed. You only pay for Fable 5 usage. Transparency: Anthropic does not notify users when a query has been routed to Opus 4.8. This is intentional—it avoids signaling attack vectors to users who might otherwise probe the boundaries.

---

Real-World Applications

Stripe's 50-Million-Line Codebase Migration

Stripe used Fable 5 to migrate a 50M-line Ruby monolith to a modern service architecture in one day. A human team estimated 2 months of effort. Fable 5's 1M context allowed it to:

• Ingest the entire codebase at once
• Reason across dependency graphs without summary artifacts
• Generate coherent refactoring sequences
• Adapt to Stripe's internal patterns in real-time

This is the canonical proof point for Fable 5's capability on enterprise-scale code transformation.

Autonomous Research Agents

Fable 5 excels at multi-step research workflows:

• Formulating hypotheses from literature
• Designing experiments with multi-part reasoning
• Analyzing results against theoretical models
• Iterating on research direction without human intervention

Organizations in biotech, climate science, and materials research are building agents that run Fable 5 loops autonomously for days.

Long-Context Chatbots

Customer support chatbots can now maintain 1M-token conversation histories, allowing Fable 5 to reason over entire customer journeys, ticket histories, and internal documentation without truncation or retrieval-augmented generation (RAG) overhead.

---

Pricing Deep Dive

Direct Pricing

• Input: $10 per million tokens
• Output: $50 per million tokens

This is exactly 2x Opus 4.8 ($5/$25).

Effective Pricing (With Thinking Overhead)

A typical Fable 5 request that generates:

• 1,000 input tokens (user query + context)
• 5,000 internal reasoning tokens
• 500 visible output tokens
• = 5,500 output tokens total

Cost = 1,000 × ($10/1M) + 5,500 × ($50/1M) = $0.275

Compare to Opus 4.8 (same task, no thinking overhead):

• 1,000 input + 500 output tokens
• Cost = 1,000 × ($5/1M) + 500 × ($25/1M) = $0.025

Effective multiplier: 11x in this example. More realistic multiplier across diverse workloads: 3–5x.

Batch API Pricing (Optional)

The Batch API (beta) allows you to queue Fable 5 requests for asynchronous processing. Batch requests receive a discount (~50% off on-demand pricing) and are processed in off-peak hours. Output limit increases to 300,000 tokens (vs. 128,000 for standard API). Use Batch for:

• Report generation (overnight processing)
• Code refactoring (batch codebase migrations)
• Training data generation
• Autonomous agent loops

---

Availability and Access

Geographic availability:

• United States: Unrestricted access for U.S. citizens/residents.
• Export-restricted countries: Access subject to U.S. Department of Commerce controls. Foreign nationals may be blocked by API-level geo-gating.
• Most of the world: Access available, but subject to export control verification by Anthropic.

Platforms:

Claude API — Standard on-demand pricing; supports streaming, vision, file upload.

Claude.ai — Subject to usage limits (Pro/Max/Teams subscribers get higher limits).

AWS Bedrock — Model ID: anthropic.claude-fable-5; includes AWS CloudWatch integration.

Google Cloud Vertex AI — Standard Vertex pricing applies.

Microsoft Foundry — Integrated with Azure ecosystem.

Data retention: 30-day minimum (Covered Model). No zero-retention option.

---

Comparing Claude Fable 5 to Other Frontier Models

See Claude Fable 5 vs Opus 4.8 vs Sonnet 5 for an in-depth internal Claude comparison.

For external competitor benchmarks, see Claude Fable 5 vs GPT-5.5 vs Gemini 3.

---

Knowledge Cutoff and Training Data

Knowledge cutoff: January 2026

Fable 5 was trained on data through January 2026. For queries about events, people, or developments after January 2026, Fable 5 will not have training data. This is a limitation shared with most LLMs released before mid-2026.

If real-time information is critical, use Fable 5 with retrieval-augmented generation (RAG) to ground responses in current data sources (APIs, search engines, databases).

---

Frequently Asked Questions

Can I use Fable 5 with Claude Code?

Yes. Fable 5 is available in Claude Code (model selector). This is ideal for complex coding tasks, multi-file refactoring, and agent-driven development workflows.

Does Fable 5 support vision/image analysis?

Yes. Fable 5 supports vision analysis via the Vision API. You can upload images and ask Fable 5 to analyze, describe, or extract information. Context and reasoning work across both text and vision.

Can I disable adaptive thinking?

No. Adaptive thinking is always on in Fable 5 and cannot be disabled via API parameters. It is a core part of the model's architecture.

What happens if Fable 5 falls back to Opus 4.8?

The end-user sees a normal response, but from Opus 4.8. No notification is sent. No additional charge is incurred (the Opus request is free). This is transparent by design.

Is Fable 5 available for fine-tuning?

No. Anthropic has not released a fine-tuning API for Fable 5 as of July 2026. Fine-tuning is available for Claude 3.5 Sonnet, but not yet for the Mythos tier.

What's the difference between Fable 5 and Mythos 5?

Functionally identical—same model, same pricing, same context. Mythos 5 is the restricted U.S. government/cyberdefense version with priority queue and additional monitoring. Fable 5 is the public GA version.

---

Conclusion

Claude Fable 5 represents a significant leap in LLM capability, particularly for software engineering and complex reasoning tasks. The 80.3% SWE-bench Pro score, 1M context, and adaptive thinking make it the clear choice for enterprise codebase automation, autonomous research agents, and reasoning-intensive workflows.

The June 2026 export-control saga added regulatory complexity, but post-July-1 relaunch is stable, with stronger safeguards and global availability (subject to export controls).

For most production applications, Opus 4.8 remains the better cost-performance choice. Fable 5 is reserved for tasks where capability is non-negotiable and cost is secondary.

Track Fable 5's evolving ecosystem and real-world performance on the Claude Fable 5 vs Opus 4.8 vs Sonnet 5 decision guide.