Claude Fable 5 and Mythos 5: what the first Mythos model means for your organization

Claude Fable 5 is the first Mythos-class model Anthropic has made available to the general public, launched on June 9, 2026. This is not an incremental update: it represents a new tier of capability, with an 11-point lead on SWE-Bench Pro over the best competing model, a 50-million-line codebase migrated in a single day in one reported case, and a one-million-token context window.

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

The Mythos class is built 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), the Claude apps, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Its context window exceeds one million tokens, with a generation capacity of 128,000 tokens per request.

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

In the vast majority of sessions, Fable 5 behaves like Mythos 5. The fallback to Claude Opus 4.8 only occurs on requests detected as high-risk in specific domains (cybersecurity, biology, chemistry, model distillation).

Suspension and restoration of access: an episode worth noting

Fable 5 and Mythos 5 have not been continuously available since their launch. On June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls. These controls were lifted on June 30, 2026, and Anthropic restored access on July 1, 2026. Anthropic has published an official statement regarding this episode.

For an organization evaluating Fable 5 today, this point matters: for nearly three weeks, the model's availability depended on a regulatory context external to Anthropic, which is worth incorporating into any risk analysis before production deployment.

What benchmarks really say

On SWE-Bench Pro, which evaluates real-world problem solving from public GitHub repositories without assistance, Fable 5 achieves 80.3%, compared to 69.2% for Claude Opus 4.8 and 58.6% for GPT-5.5. The gap with the top competitor is over 20 points on this specific benchmark.

On aggregated leaderboards like BenchLM, which compiles over 120 models, Fable 5 ranks among the very top. These rankings evolve continuously as new models are added and scores are revised; therefore, Fable 5's exact position at the time you read this article may differ from its launch position. It is recommended to check the live leaderboard rather than relying on a static figure.

The most revealing pattern is not an isolated score, but the nature of the tasks where Fable 5 pulls furthest ahead: long, multi-tool, multimodal, and ambiguous tasks that resemble real work rather than simple prompt-response exchanges. The gap with previous models widens as tasks become more complex.

Areas where the gap is most significant

Advanced coding. Large codebase migrations, complex implementations, and autonomous multi-day sessions. The model writes its own tests to verify its work and uses vision to compare its output against the initial goal.

Long-form knowledge work. Multi-source research, analysis of large document portfolios, and structured syntheses, including planning and delegation toward a review-ready deliverable.

Applied vision. Understanding diagrams, charts, and tables embedded in PDFs and files, which is particularly useful in finance, legal, and analytics.

Autonomous agents. Deployed in a multi-agent system, the model plans across multiple steps, delegates to sub-agents, and verifies its own work.

A report from the launch mentions a migration in a 50-million-line codebase completed in one day, compared to an estimate of over two months of manual work. This type of feedback remains an individual testimonial, to be validated against your own use cases rather than generalized as-is.

Guardrails and their practical implications

Fable 5 comes with the strictest constraints Anthropic has ever imposed on a general-purpose model.

Domains with automatic fallback. For queries involving cybersecurity, biology, chemistry, or model distillation, Fable 5 automatically switches to Claude Opus 4.8. This switch is silent in most interfaces. Via the API, blocked requests return a stop_reason: "refusal" with an HTTP 200 status, without being billed at the Fable 5 rate. The classifiers have been calibrated conservatively, meaning they sometimes flag harmless requests; careful integration via the Fallback API is necessary for applications active in these domains.

Data retention requirements. Using Fable 5 requires accepting a 30-day data retention period to power security classifiers. For organizations subject to GDPR or strict industry regulations (healthcare, finance, legal), this is a key point to analyze before any deployment. Anthropic offers an option for inference localized exclusively in the United States, at 1.1x the standard rate, for organizations with data residency requirements.

Pricing: how to evaluate the cost

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, with a 90% discount for prompt caching. With batch pricing, these rates drop to $5 and $25, respectively. This is approximately double the rate of Claude Opus 4.8.

Evaluating based solely on cost per token is misleading for a model designed for long-running tasks. At a cost per task of approximately $8 to $10, Fable 5 in "high effort" mode would resolve a significantly higher proportion of difficult problems compared to Opus 4.8, according to analyses published at launch. If the task is equivalent to a day's work for a senior-level professional, the equation remains favorable even at this price point.

The recommendation emerging from early feedback: do not route all requests to Fable 5 by default. Reserve this model for tasks that truly justify it, and keep Opus 4.8 or Sonnet for routine use, using a gateway that intelligently routes requests based on their complexity.

Strategic implications for your organization

Some projects that were previously impossible to delegate now become feasible. For projects spanning several days, the consistency of previous models would degrade. Fable 5 changes this limit: large-scale codebase migrations, comprehensive analysis of contract portfolios, multi-step research with synthesis, and automation of end-to-end reporting workflows. These are no longer pilot projects; they are operational workflows.

The gap between teams that integrate and those that wait will widen. Teams that integrate this type of model into their workflows now are accelerating on tasks that their competitors are still performing manually.

AI governance is becoming an immediate operational issue. Who validates deliverables produced autonomously over several days? How do you manage the 30-day data retention period in light of your regulatory constraints? How do you distinguish tasks that justify the Fable 5 cost from those that do not? These questions arise as soon as you go into production, and organizations that have thought them through in advance progress faster and with less risk.

Availability and access

Fable 5 is available for Claude Pro, Max, Team, and Enterprise subscriptions. It was free until June 22, 2026, and is now billed via usage credits (2x the standard credits on these plans). Via API, it is accessible on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, as well as in GitHub Copilot, where data retention must be manually enabled by the Copilot administrator.

As noted above, access to both models was suspended from June 12 to June 30, 2026, for compliance with U.S. export controls, before being restored on July 1, 2026. For access to Claude Mythos 5 outside of Project Glasswing, Anthropic has opened an interest form for cybersecurity and biology research organizations.

‍