Claude Fable 5.1 and Mythos 5.1: Pricing, Access, and What Changed

*This is a research-based analysis of Anthropic’s September 1, 2026 launch materials and system-card disclosures. DiscoverAI has not independently benchmarked either model. Performance, cost-reduction, and safeguard figures are provider-reported unless otherwise stated.*

The short answer

Anthropic released Claude Fable 5.1 for ambitious coding, knowledge work, vision, and long-running agents, alongside Claude Mythos 5.1 for a limited group of vetted cybersecurity and life-sciences organizations. Anthropic says the models share the same underlying capability, while Fable applies safeguards and fallback routing in higher-risk cyber, biology, and chemistry domains.

Fable 5.1 is available through paid Claude plans, the Claude API, major cloud marketplaces, and supported enterprise surfaces. API pricing is $10 per million input tokens and $50 per million output tokens. The notable cost change is cache reads at $0.25 per million tokens; Anthropic estimates this lowers typical workload cost by about 25% and highly agentic workload cost by as much as 45%. Those savings depend on cache reuse, output volume, tool calls, retries, and orchestration—not the model name alone.

What actually changed

Anthropic positions Fable 5.1 around work that spans hours, applications, files, and recovery steps. It highlights codebase-wide implementation, document-heavy analysis, browser work, asynchronous agents, and scientific problem-solving. Provider benchmarks show large gains on its chosen coding and science tasks, but benchmark setup, harness design, safeguards, and statistical uncertainty matter. A leaderboard result is a hypothesis for your evaluation, not a purchase order.

The safeguard architecture is equally important. Anthropic says many flagged cyber requests route to Opus 4.8 and biology requests to Opus 5. API customers must configure the Fallback API. The company reports fewer benign interventions than the previous release while retaining restrictions on penetration testing, exploit generation, binary vulnerability scanning, and dual-use scientific work.

Mythos 5.1 reduces some of those restrictions for approved defensive and research programs. Access is not general: organizations must be vetted, availability is initially limited, and default safety monitoring carries a 30-day data-retention requirement. That makes Mythos a governed-access program, not a higher toggle hiding in an ordinary Claude subscription.

How teams should evaluate Fable 5.1

Choose one representative long-running task with a fixed repository or source packet, acceptance tests, a time limit, and an explicit tool budget. Run it at more than one effort setting and compare completion quality, human correction time, failed actions, cache hit rate, input/output tokens, tool costs, wall time, and recovery after a deliberately broken step.

Test safeguard behavior separately. Use approved benign security and scientific prompts that resemble real work, record fallbacks and refusals, and verify whether the routed model still meets the task. Do not infer sensitive-domain suitability from general coding scores. Teams with residency requirements should also account for Anthropic’s stated 1.1x price for US-only inference.

The practical takeaway

Fable 5.1 is most relevant when a difficult job can justify premium tokens and genuinely reuse context. The buyer question is whether it reduces total reviewed-work cost, not whether it wins a launch benchmark. Mythos 5.1 represents a different decision: whether an approved organization can accept vetting, monitoring, retention, and narrow program access in exchange for higher-risk research capability.

The safest rollout is measured and reversible. Keep human approval around external actions, cap spend and duration, preserve tool-call evidence, and compare the completed deliverable—not the most impressive intermediate reasoning trace.