Claude Fable 5 and Mythos: How Anthropic's most powerful model impacts your business

On June 9, 2026, Anthropic made Claude Fable 5 available, the first Mythos-class model accessible to the general public. This is not an incremental update; it's a category change, with an 11-point lead on SWE-Bench Pro compared to the best competing model, a 50-million-line codebase migrated in a single day, and a one-million-token context window. Here's what that concretely means for an organization.

‍

Fable 5 and Mythos 5: the same foundation, two access levels

The Mythos class is based on a single underlying model. What differs is the level of guardrails applied in production.

‍

Claude Fable 5 is the public version. It is available via the Anthropic API (identifier: claude-fable-5), Claude apps, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. This is the model developers and businesses can use today. Its context window exceeds one million tokens, with a generation capacity of 128,000 tokens per request.

‍

Claude Mythos 5 is the version without cybersecurity and biology restrictions. It is only accessible to a select group of organizations chosen as part of Project Glasswing: critical infrastructure operators, and verified cybersecurity and biology researchers. For the vast majority of businesses, Fable 5 is the relevant model.

‍

An important detail: in over 95% of sessions, Fable 5 behaves exactly like Mythos 5. The switch to Opus 4.8 only occurs for requests detected as high-risk, in very specific domains. Anthropic estimates that less than 5% of sessions are affected.

‍

What the benchmarks really say

The figures are striking, but they need to be read correctly to understand where the real leap lies.

‍

On SWE-Bench Pro, which evaluates the resolution of real-world problems from public GitHub repositories without assistance, Fable 5 achieves 80.3%. Claude Opus 4.8 tops out at 69.2%, GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2%. The gap with the best competitor is 11 points.

‍

On the BenchLM platform, which aggregates over 120 models, Fable 5 ranks second on the provisional overall leaderboard with a score of 96/100, and also second on the verified leaderboard.

‍

But the most revealing pattern lies in the nature of the tasks where Fable 5 creates the biggest gap. As noted in the official benchmark guide, Fable 5 is strongest where the task is lengthy, multi-tool, multimodal, ambiguous, or closer to real work rather than a simple prompt-response exchange. The gap with previous models widens as tasks become more complex.

‍

The areas where the difference is most significant

• Advanced coding: large codebase migrations, complex implementations, multi-day autonomous sessions. Fable 5 writes its own tests to verify its work and uses vision to compare its outputs against the initial objective.
• Long-horizon knowledge work: multi-source research, analysis of large document portfolios, structured syntheses. The model plans, delegates, and delivers a review-ready output.
• Applied vision: Fable 5 understands diagrams, graphs, and tables embedded in PDFs and files. This is particularly useful in the finance, legal, and analytical sectors.
• Autonomous agents: deployed in a multi-agent system (Claude Code, Claude Managed Agents), it plans across multiple steps, delegates to sub-agents, and verifies its own work.

‍

To provide a concrete scale: a testimonial published at launch reports that a 50-million-line Ruby codebase migration, performed by Fable 5 in one day, would have required more than two months of manual work. Another user reports that it outperformed their frontier physics analysis model in 36 hours, whereas GPT-5.5 had taken four days.

‍

Safeguards and their practical implications

Fable 5 is the most powerful model Anthropic has ever made available to the public, and it comes with the strictest constraints the company has ever imposed on a general model. Understanding these constraints is essential before any deployment.

‍

Areas with automatic fallback

For queries related to cybersecurity, biology, chemistry, or model distillation, Fable 5 automatically switches to Claude Opus 4.8. This switch is silent in most clients. Via the API, blocked requests return a stop_reason: "refusal" with HTTP 200, without billing at the Fable 5 rate.

‍

Anthropic states that it has conservatively calibrated the classifiers to move quickly and safely. This means they sometimes flag innocuous requests. For teams developing applications in these areas, careful integration via the Fallback API is necessary to manage these cases.

‍

The Data Retention Constraint

Using Fable 5 requires accepting 30 days of data retention for the security classifiers to function. This is a non-negotiable point in the model's current architecture.

‍

For organizations subject to GDPR, data localization policies, or strict sectoral regulations (health, finance, legal), this is a constraint that needs to be analyzed before any deployment. Anthropic offers a localized inference option exclusively in the United States, at 1.1 times the standard rate, for organizations with data residency requirements.

‍

Pricing: how to approach the cost

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, with a 90% discount on prompt caching. For batch pricing, the rates drop to $5 and $25 respectively.

‍

This is approximately twice the rate of Claude Opus 4.8. However, focusing solely on cost per token is misleading for a model designed for long tasks.

‍

An analysis published at launch illustrates the logic: for a per-task expenditure of approximately $8 to $10, Fable 5 in "high effort" mode solves about 24% of difficult problems, compared to 13.4% for Opus 4.8. If the task is equivalent to a senior professional's day of work, paying $20 in compute for a 31% resolution rate on the most challenging cases is a very favorable equation.

‍

The practical recommendation emerging from initial feedback: do not route all requests to Fable 5 by default. Reserve this model for tasks that truly warrant it, and keep Opus 4.8 or Sonnet for common uses. Implementing a gateway that intelligently routes requests based on their complexity is a good practice from the outset.

‍

Strategic implications for your organization

1. Some projects that were previously not delegable now become so

Until now, AI models were effective for short, well-defined tasks. For projects spanning several days, consistency degraded. Fable 5 changes this limitation. Use cases that become credible include: migration of large codebases, comprehensive analysis of contractual portfolios, multi-step research with synthesis, and automation of complete reporting workflows. These are no longer pilot projects. They are operational workflows.

‍

2. The gap between teams that integrate and those that wait will widen

Fable 5 is not just another tool to evaluate. It's a threshold. Teams that integrate it into their workflows now will accelerate tasks that their competitors are still doing manually. The same phenomenon occurred with the transition from spreadsheets to ERPs, then with the cloud. The gains accumulated by early adopters don't disappear when others catch up. They compound.

‍

3. AI governance becomes an immediate operational concern

A model capable of working autonomously for several days raises concrete questions: who validates the deliverables? How do you manage the 30-day data retention period in light of your regulatory constraints? How do you distinguish between tasks that justify the cost of Fable 5 and those that don't? How do you manage the transitions to Opus 4.8 in your applications?

‍

These questions are no longer theoretical. They arise from the very first production deployment. Organizations that have considered them beforehand progress faster and with less risk.

‍

Availability and Access

Fable 5 is available immediately for Claude Pro, Max, Team, and Enterprise subscriptions (free until June 22, 2026, with usage credits thereafter). On these plans, Fable 5 consumes twice the usual credits.

‍

Via the API, it is accessible on Amazon Bedrock (US East and Europe Stockholm regions at launch), Google Cloud Vertex AI, and Microsoft Foundry, in addition to the Claude platform directly. It is also available in GitHub Copilot, with a specific constraint: data retention must be activated by the Copilot administrator, and it is disabled by default.

‍

For access to Claude Mythos 5 outside of Project Glasswing, Anthropic has opened an expression of interest form for cybersecurity and biology research organizations.

‍