Building agents that reason over massive codebases and run for days requires models that maintain focus and autonomy at scale. Anthropic's Claude Fable 5 brings Mythos-class performance to general availability, with state-of-the-art benchmarks and sub-50¢/MTok pricing. Engineers can now run multi-day agentic workflows—code migrations, scientific research loops, complex knowledge tasks—without token-count fragility or context collapse, and at roughly half the cost of prior Mythos models. Anthropic's release emphasizes vision-to-code capability and sustained autonomy; the model holds coherence across millions of tokens and can operate in agent harnesses for extended periods. Availability is immediate on the Anthropic API, with Pro/Max/Team/Enterprise subscription inclusion through June 22 before moving to usage-credit billing.

Source: Claude Fable 5

Long-running agentic systems—whether refactoring massive codebases, running research loops, or managing complex commerce workflows—have historically required orchestration layers to work around token limits and context degradation. Fable 5's sustained focus and autonomy flatten that complexity. Vision capabilities unlock new use cases: teams can feed model screenshots or PDFs and get structured outputs, eliminating intermediate parsing steps. At this price point, sustained-workload economics shift in favor of more ambitious automation.

Claude Fable 5 is designed to operate inside agent harnesses—like Claude Code—for extended periods without manual checkpoints. The model plans multi-step approaches, refines its work over iterations, and maintains coherence across millions of tokens. Stripe's 50M-line Ruby migration—a task that would typically span months of incremental work—completed in a single day of autonomous refactoring.

The vision module is strong enough to extract tables from scientific papers, read graphs, and reverse-engineer code from screenshots. Engineers can now skip intermediate OCR or parsing layers: feed a research PDF or UI mockup directly, and Fable 5 extracts structured output with high fidelity. This is particularly valuable for knowledge work and scientific automation pipelines.

Anthropic reports state-of-the-art results on nearly all tested benchmarks, with notable gains on software engineering tasks. The performance delta grows on longer, more complex problems—exactly the domain where sustained focus matters most. This positions Fable 5 as the default for code-generation, refactoring, and complex reasoning pipelines.

At $10/MTok input and $50/MTok output, Fable 5 costs roughly half the prior Mythos Preview. For teams running sustained agent workloads, this pricing shift makes autonomous workflows economically viable at scales that were previously cost-prohibitive. Through June 22, Pro/Max/Team and Enterprise plans include Fable 5 at no extra charge.

A team managing a multi-million-line codebase can now hand Fable 5 a migration task—upgrade a framework, refactor a module, apply a linting standard—and let it run autonomously for hours or days. The model plans its approach, breaks the work into sub-tasks, and refines output iteratively without human intervention. Rather than blocking developer time on incremental changes, the agent completes structural work end-to-end. Work that historically took months of engineering cycles completes in a single sustained agent run.

Scientific and knowledge-work teams can feed Fable 5 PDFs, research papers, or data visualizations and request structured extraction—table parsing, figure interpretation, bibliographic indexing. The vision module reads the image content directly, eliminating intermediate OCR or manual annotation. For product teams, feeding UI screenshots or design mockups to Fable 5 enables automated code-generation and accessibility review. Structured data flows directly from unstructured sources, compressing knowledge-work pipelines.

For AI agents that need to operate autonomously in commercial contexts—placing orders, transferring funds, or making commitments on behalf of users—sustained reasoning and policy adherence matter. Fable 5's extended autonomy means agents can hold complex rules in mind across long interaction sequences. Systems managing agentic identity and transaction compliance can pair this reasoning capability with credential verification and policy gates to build end-to-end autonomous commerce flows. Agents can reason through multi-step transactional workflows while maintaining compliance and auditability.

Long-running automation, code generation, and agentic workflows across infrastructure and commerce are the immediate beneficiaries. Systems like Stripe (internal engineering automation), Vercel (code-to-deployment pipelines), Plaid (financial data extraction), and AgentScore all benefit from Fable 5 because sustained reasoning and vision capability reduce friction in their core workflows—migration automation, data parsing, and autonomous transaction verification.

AgentScore provides commerce infrastructure for AI agents: a universal pay CLI, merchant SDK, buyer-side identity verification (Passport), and compliance gating that lets merchants accept payments from autonomous buyers while maintaining regulatory compliance. Teams building with AgentScore could leverage Fable 5's extended autonomy and vision to construct agents that autonomously assess transaction risk (calling `POST /v1/assess` to verify identity and evaluate policy compliance), retrieve and review stored credentials (`GET /v1/credentials`), and reason through multi-step purchasing flows—all while maintaining compliance boundaries and auditability across the entire interaction.

Anthropic is expected to refine pricing and release variants over the coming months; watch for specialized models targeting specific domains (code, vision, reasoning) and further cost optimization. The Mythos-class rollout signals a shift toward sustained autonomy as table-stakes for production AI systems.

The code examples in this tutorial are grounded in the following docs pages: