Anthropic Details Claude Fable 5 Autonomous Coding at Base44

Anthropic recently published a case study detailing how engineering firm Base44 uses Claude Fable 5 for high-stakes autonomous coding. The firm deployed the Mythos-class model to manage multi-day autonomous sessions that previously required constant human supervision.

Autonomous Engineering Workflows

Base44 integrated Fable 5 as an assistant engineering manager to draft specifications, design architecture, and review pull requests. The deployment relies heavily on Claude Code version 2.1.170+ and Claude Cowork.

The firm attributes the success to the model’s 1,000,000-token context window and its continuous verification loops. This allows developers to scale Claude Code across enterprise monorepos without losing architectural context during extended reasoning sessions.

If you build long-running agents, the Base44 implementation shows how to leverage large context windows for self-correction. The system automates ambiguous, highly multi-step problems by validating its own outputs before proceeding. Developers looking to implement similar systems can build Claude Code routines that handle continuous verification.

Technical Specifications and Pricing

Claude Fable 5 operates as a Mythos-class model, positioned directly above the Opus tier. It features a maximum output of 128,000 tokens per request.

Pricing is set at $10 per million input tokens and $50 per million output tokens. Anthropic has extended a promotional period offering free Fable 5 access for paid Pro, Max, Team, and Enterprise subscribers through July 19, 2026. During this window, users can allocate up to 50% of their weekly usage limits to Fable 5 at no additional cost.

Model	SWE-Bench Pro
Claude Fable 5	80.3%
Opus 4.8	69.2%
GPT-5.5	58.6%

The system includes a strict safety routing mechanism. Requests flagged as high-risk in cybersecurity, biology, or chemistry automatically fall back to the weaker Opus 4.8 model. While this satisfies compliance requirements, developers note the fallback downgrades performance without user confirmation.

Regulatory and Competitive Pressures

The publication of the Base44 architecture arrives during a turbulent month for Anthropic. Access to Fable 5 and its unrestricted counterpart, Mythos 5, was globally suspended on June 12, 2026. The U.S. government mandated the suspension following the discovery of a jailbreak technique by Amazon researchers.

Anthropic restored access on July 1, 2026, after implementing new safety classifiers. The delayed rollout impacts how enterprise teams approach desktop agent workflows involving sensitive codebases.

The case study also follows OpenAI’s July 9 release of GPT-5.6 and its Sol model. OpenAI benchmarks indicate Sol outperforms Fable 5 by 13.1 points on the Agents’ Last Exam evaluation.

For engineering teams building autonomous coding systems, the Base44 deployment validates the utility of million-token context windows for complex multi-day tasks. You should evaluate the 128,000-token output limit against your codebase size and monitor your workflows for unexpected Opus 4.8 safety downgrades.