Claude Fable 5: The first Mythos model is powerful, expensive, and heavily filtered

What it does

Anthropic has launched Claude Fable 5, the first model in its new Mythos class of large language models. Claude Fable 5 scores at the top of nearly every benchmark, including a 95 percent on SWE-bench Verified, indicating strong performance on coding and reasoning tasks. This model is positioned as a high-end option with improved capabilities over Anthropic’s prior Opus 4.8 model.

Why it matters

Claude Fable 5 pushes performance forward but comes with a steep price tag—about twice the cost of Opus 4.8, charging between 10 and 50 dollars per million tokens depending on volume. That pricing will tighten budgets for businesses or developers weighing cutting-edge quality against operational costs. Additionally, Anthropic enforces strict safety filters that reject roughly 9 percent of input requests, aiming to reduce harmful outputs but potentially blocking valid queries. These filters increase friction and risk frustrating users in edge cases.

Another change is Anthropic’s new data retention policy. Even users on zero-data-retention contracts will have their inputs stored for 30 days. For privacy-conscious companies or those handling sensitive data, this shift raises risk and compliance issues that may force more careful contract reviews or limit use cases.

Who it is for

Claude Fable 5 targets customers who prioritize high accuracy and safety over cost. Teams building advanced applications requiring state-of-the-art reasoning, code generation, or sensitive content moderation will find the model’s capabilities valuable. Enterprises with budgets to absorb higher token pricing and who value Anthropic’s filtering may choose Mythos models as safer, more capable AI backbones.

However, startups and small developers focused on cost efficiency or those needing minimal data retention will likely hesitate or look for alternatives. Claude Fable 5’s strict safety filters and data policies may also slow deployment in creative or exploratory scenarios where content boundaries are more fluid.

The catch

The premium price and aggressive safety filtering make Claude Fable 5 less flexible for broad use cases. Nine percent of requests blocked means some requests will fail unpredictably, complicating user experience and automation flows. The 30-day input retention despite prior zero-retention terms creates new privacy and liability considerations.

Businesses should weigh the operational complexity and added cost against the potential boost in accuracy and safety. Using Claude Fable 5 will require stricter data governance, compliance checks, and tolerance for filtering suppression.

What to watch next

Monitor how Anthropic adjusts pricing and filtering thresholds over the next releases and how customers adapt to the new data retention terms. Competing models may target the price-performance gap Claude Fable 5 creates. Also watch for adoption patterns across industries to see whether demand favors very high accuracy and safety or more flexible, cost-effective models.

Anthropic’s ongoing Mythos rollout will reveal if this model class reshapes AI selection priorities for builders and enterprises balancing risk, cost, and performance.

AI Quick Briefs Editorial Desk