Claude Fable 5 and Mythos 5: Anthropic's Most Powerful AI Models Explained

Anthropic Just Released Its Most Capable Models Ever

On June 9, 2026, Anthropic released Claude Fable 5, the most capable AI model the company has ever made publicly available. Alongside it, they announced Claude Mythos 5, the same model without safety classifiers, restricted to vetted partners through a programme called Project Glasswing.

I have spent the past two weeks running both models through production workloads for our AI agency clients, and the performance jump from the previous Claude generation is substantial. This is not an incremental upgrade. Fable 5 and Mythos 5 represent a new capability tier that changes what is practical to build with AI.

Here is everything you need to know, from technical specifications to pricing to practical deployment considerations for AI agencies and engineering teams.

Claude Fable 5: Technical Specifications

The Numbers That Matter

Context window: 1 million tokens. Fable 5 accepts up to 1 million tokens of input context in a single request. To put that in perspective, that is roughly 750,000 words, or approximately 10 to 15 full-length novels. For practical AI work, it means you can load entire codebases, complete legal case files, months of email correspondence, or full research paper collections into a single prompt.

We have written extensively about how context windows shape AI architecture decisions. Fable 5’s 1M context window eliminates most of the retrieval-augmented generation complexity that AI agencies have been building around context limitations. In many cases, you can now simply load the full document set directly into the prompt instead of building vector databases and chunking pipelines.

Output tokens: 128K per request. This is the largest output capacity of any commercially available model. Previous Claude models capped at 8K or 16K output tokens per request, requiring multi-turn generation for long documents. Fable 5 can generate an entire research report, a complete codebase module, or a full legal brief in a single response.

Pricing: $10 per million input tokens, $50 per million output tokens. This is premium pricing, positioned above GPT-4o and previous Claude models. For an AI agency managing costs carefully, the pricing demands thoughtful routing. Use Fable 5 for tasks that require its full capability. Use smaller, cheaper models for routine tasks. The cost difference is significant enough that model routing is not optional, it is a financial necessity.

Core Capabilities

Complex reasoning. Fable 5 handles multi-step reasoning chains that would cause previous models to lose coherence. Mathematical proofs, legal argument construction, financial modelling logic, and scientific hypothesis evaluation all show marked improvement. The model maintains reasoning consistency across much longer chains than its predecessors.

Long-horizon agentic work. For AI agencies building autonomous agents, this is the headline capability. Fable 5 can maintain coherent goal pursuit across extended agent loops, tracking sub-task dependencies, managing state, and recovering from intermediate failures more reliably than any previous Claude model. Agent tasks that previously required frequent human intervention can now run for longer before needing a check-in.

Software engineering. Code generation, review, refactoring, and debugging are all significantly stronger. Fable 5 understands large codebases loaded into context, identifies cross-file dependencies, and generates code that accounts for project-wide conventions and patterns. For AI agencies offering development services, this capability directly translates into faster delivery.

Advanced vision. Fable 5 processes images with strong understanding of charts, diagrams, screenshots, handwritten notes, and complex visual layouts. For workflows that involve analysing dashboards, extracting data from scanned documents, or understanding architectural diagrams, the vision capability is production-ready.

Claude Mythos 5: The Unfiltered Variant

What Mythos 5 Is

Claude Mythos 5 is architecturally the same model as Fable 5. Same parameters, same training, same capabilities. The difference is that Mythos 5 does not include the safety classifiers that Fable 5 uses to filter responses on high-risk topics.

This does not mean Mythos 5 is “unsafe.” It means it does not have the additional classification layer that Fable 5 applies to detect and handle potentially harmful requests. For most commercial applications, this distinction is irrelevant because the tasks do not trigger safety classifiers in the first place. But for specific research, medical, legal, and security applications where safety classifiers can interfere with legitimate work, Mythos 5 removes that friction.

Project Glasswing: Access Restrictions

Mythos 5 is not publicly available. Access is restricted to organisations that have been vetted through Anthropic’s Project Glasswing programme. The vetting process evaluates the organisation’s use case, security posture, and commitment to responsible deployment.

From what I have observed, Project Glasswing participants include major research institutions, defence and national security contractors, select healthcare organisations, and a small number of AI agencies with demonstrated enterprise deployment track records. If your AI agency serves regulated industries and needs unfiltered model access for legitimate professional applications, applying to Project Glasswing is worth the effort.

Practical Differences Between Fable 5 and Mythos 5

For 95% of AI agency workloads, Fable 5 and Mythos 5 produce identical results. The safety classifiers in Fable 5 only activate on specific categories of content that most business applications never encounter. I have run both models through our standard evaluation suite covering customer service automation, code generation, document analysis, research synthesis, and multi-agent orchestration. The outputs are functionally indistinguishable.

Where the difference becomes apparent is in edge cases. Medical AI applications that discuss drug interactions, dosages, or treatment protocols. Legal applications that analyse cases involving violence or abuse. Security research that involves malware analysis or vulnerability assessment. In these domains, Fable 5’s classifiers occasionally trigger fallback behaviour, while Mythos 5 handles the content directly.

The Safety Architecture: How Fable 5’s Classifiers Work

Classifier-Based Filtering

Fable 5 uses a set of built-in classifiers that evaluate both the input prompt and the generated output for content that falls into predefined high-risk categories. When a classifier triggers, the model has two possible responses depending on the severity and category.

For moderate triggers, the model may add caveats, disclaimers, or framing to its response while still providing the requested information. For severe triggers, the model falls back to Claude Opus 4.8, an older and more conservative model that handles the request with tighter safety constraints.

The Opus 4.8 Fallback

The fallback to Claude Opus 4.8 is worth understanding for production deployments. When Fable 5 encounters a request that triggers its high-severity classifiers, it does not simply refuse. Instead, it routes the request to Opus 4.8, which processes it with a different set of safety parameters. The response comes back from Opus 4.8 but is returned through the Fable 5 API endpoint, so from your application’s perspective, the API contract is maintained.

The practical implication is that Fable 5 responses on sensitive topics may have noticeably different quality, style, and capability compared to non-sensitive topics, because they are actually being generated by a different, older model. For AI agencies building applications that might occasionally touch sensitive domains, you need to test for this fallback behaviour and design your user experience accordingly.

For deeper context on managing model behaviour unpredictability, see our guide on LLM hallucination management for AI agencies.

Pricing Analysis for AI Agencies

Cost Structure

At $10 per million input tokens and $50 per million output tokens, Fable 5 is positioned as a premium model. Let me put those numbers in practical terms for AI agency budgeting.

Scenario: Document analysis agent. Processing a 100-page legal document (approximately 50,000 tokens) and generating a 5-page summary (approximately 3,000 tokens). Input cost: $0.50. Output cost: $0.15. Total: $0.65 per document. At 100 documents per month, that is $65 in model costs.

Scenario: Coding assistant. Loading a 20,000-line codebase (approximately 80,000 tokens) and generating a code review with suggestions (approximately 10,000 tokens). Input cost: $0.80. Output cost: $0.50. Total: $1.30 per review session. For a development team running 20 reviews per day, that is $26 per day or roughly $780 per month.

Scenario: Long-running agent loop. An autonomous agent processing a complex research task with 15 reasoning steps, each involving context refresh with approximately 200,000 input tokens and 5,000 output tokens per step. Total input: 3M tokens ($30). Total output: 75K tokens ($3.75). Total: $33.75 per complex agent task.

These costs are manageable for high-value professional applications but prohibitive for high-volume, low-value tasks. Every AI agency using Fable 5 in production needs a model routing layer that sends routine tasks to cheaper models and reserves Fable 5 for work that justifies its pricing.

Comparison with Alternatives

For context, comparing across the current LLM landscape: GPT-4o is roughly $5 per million input and $15 per million output. Gemini 2.0 Flash is significantly cheaper for many tasks. Open-source models running on Ollama have zero per-token cost after hardware investment.

Fable 5 is not the cost leader. It is the capability leader. The pricing makes sense when the task genuinely requires its capability tier, and using a cheaper model would produce inferior results that cost more in human review and correction time.

Cloud Availability and Deployment

Where You Can Access Fable 5

Fable 5 is available through Anthropic’s direct API, as well as through three major cloud platforms:

AWS Bedrock. Full integration with AWS Bedrock, supporting all Fable 5 features including the 1M context window and 128K output tokens. For AI agencies with existing AWS infrastructure, this is the path of least resistance.

Google Cloud Vertex AI. Available through Vertex AI’s Model Garden, with integration into Google Cloud’s broader AI and data infrastructure. Useful for teams that use BigQuery, Google Cloud Storage, or other GCP services as part of their data pipeline.

Microsoft Foundry. A newer deployment option through Microsoft’s AI infrastructure platform. Particularly relevant for enterprises that use Azure for their core infrastructure and want to keep all AI workloads within the Microsoft ecosystem.

Export Control Considerations

Anthropic has implemented export control restrictions for Fable 5 and Mythos 5 in certain regions. The specifics are not publicly documented in full detail, but in practice this means that API access may be limited or unavailable from certain jurisdictions. If your AI agency serves international clients, you need to verify that your deployment regions and client locations are not affected by these restrictions.

For multi-region AI deployments, this is a practical constraint to factor into architecture decisions. You may need to use alternative models for clients or operations in restricted regions.

What Fable 5 and Mythos 5 Mean for AI Agencies

The Capability Threshold

Fable 5 crosses several capability thresholds that make previously impractical applications viable:

Full codebase understanding. Loading an entire application codebase into context and asking the model to reason about architecture, identify bugs, or plan refactoring was impractical with smaller context windows. With 1M tokens, most production codebases fit entirely within context.

Complete case analysis. Legal, medical, and financial analysis that requires reviewing complete case files, entire patient histories, or full portfolios of documents can now happen in a single model call instead of requiring complex document chunking and retrieval.

Extended agent autonomy. The combination of the large context window, strong reasoning, and 128K output capacity means agent loops can run longer, handle more complex tasks, and produce more comprehensive outputs before requiring human review.

For AI agencies evaluating their model strategy, Fable 5 is not a replacement for all models. It is the model you reach for when the task exceeds the capabilities of more cost-effective options.

Building for Fable 5’s Strengths

The practical advice for AI agencies is straightforward. Design your model routing to use Fable 5 for:

• Tasks that require reasoning over large document sets
• Agent loops that need extended autonomy and coherence
• Code generation and review for large, complex codebases
• Analysis tasks where accuracy is critical and the cost of errors exceeds the cost of premium model pricing
• Client-facing outputs where quality directly impacts perceived value

Use cheaper models for everything else. The 80/20 rule applies. Most AI agency workloads involve many routine tasks and a smaller number of complex, high-value tasks. Route accordingly.

The Safety Trade-Off

Fable 5’s safety architecture is both a strength and a complication. For client-facing applications, the built-in safety classifiers provide a valuable layer of protection that reduces the risk of harmful or embarrassing outputs. For internal tools and specialised professional applications, the classifiers can create unexpected behaviour that disrupts workflows.

Understanding when and how the Opus 4.8 fallback triggers is essential for production deployments. Test your specific use cases thoroughly. If your application consistently triggers safety classifiers, explore whether your use case qualifies for Mythos 5 access through Project Glasswing.

Getting Started with Fable 5

1. Sign up for API access at Anthropic’s developer portal
2. Start with a small evaluation: send a complex reasoning task with a large context payload
3. Compare the output quality and latency against your current model for the same task
4. Implement model routing to direct high-value tasks to Fable 5 and routine tasks to cheaper models
5. Monitor costs carefully during the first month and adjust routing thresholds based on actual spend

For a broader perspective on choosing the right model for different use cases, see our complete LLM comparison for AI agencies.

Ready to Integrate Fable 5 Into Your AI Systems?

If you are building production AI systems, evaluating frontier models for your business, or designing agent architectures that need the full capability of models like Fable 5 and Mythos 5, contact our team. We help businesses and AI agencies architect model strategies that balance capability, cost, and safety for real-world deployment.