Claude Mythos 5 Released: 50 Million Lines of Code in One Day
Original: 刚刚，Claude Mythos 5发布！5000万行代码1天搞定
QbitAI reports Anthropic launched Claude Fable 5 and restricted-access Claude Mythos 5.
QbitAI says Anthropic introduced Claude Fable 5 for general users and Claude Mythos 5 for a small set of trusted users. The article highlights software engineering, long-context work, native vision, memory, and scientific research capabilities. It also focuses on a safety-routing design where Fable 5 downgrades high-risk requests to Claude Opus 4.8 instead of simply refusing.
Quantumbit reports on Anthropic's release of two new flagship models: Claude Fable 5 and Claude Mythos 5. According to the article, both share the same underlying capabilities but differ in product positioning: Fable 5 is a version with safety guards and risk classifiers aimed at general users, while Mythos 5 is a more complete high-capability version available only to a limited number of trusted users, particularly for sensitive capabilities involving cybersecurity and bioscience research. The article notes that Fable 5 and Mythos 5 have longer autonomous operation times compared to previous Claude models, with API pricing reduced to $10 per million input tokens and $50 per million output tokens. Technical highlights focus on software engineering, long context, native vision, memory capabilities, and research workflows. The article cites multiple benchmarks and use cases, including Fable 5 achieving 80.3% on SWE-bench Pro, and notes that in Stripe's 50 million-line Ruby codebase, it completed a full library migration that would have taken a human team over two months in just one day. On the vision front, the article mentions that Fable 5 can reason through game screens without external scaffolding to complete Pokémon FireRed; for long-horizon tasks, it emphasizes how persistent memory can enhance game and agent performance. For research, the article claims Mythos 5 can autonomously execute portions of biologist workflows, including protein binding site selection, tool orchestration and debugging, and generating molecular biology hypotheses. Notably, the report also emphasizes safety productization: Fable 5 not only refuses high-risk requests related to cybersecurity, biochemistry, and model distillation, but routes them to Claude Opus 4.8 instead. For developers and researchers, the key takeaway extends beyond model capability improvements to show how cutting-edge AI may be entering a new product phase of 'capability tiering, permission management, and model routing.'
Free shows the 3-line summary; Pro unlocks the full deep summary (~300 words) so you never have to click through.
See Pro plans →Want the original English / full article?
Read on 量子位 QbitAI →Summaries are AI-generated; the original article is authoritative.