Anthropic unveils Claude Opus 4.8; teases Claude Mythos rollout in weeks

Anthropic has introduced an upgraded iteration of its artificial intelligence model, named Claude Opus 4.8. Building on the foundation of Claude Opus 4.7, this latest version promises to enhance performance across various domains, including coding, reasoning, and practical skills. The company describes Claude Opus 4.8 as a significantly more effective collaborator, showcasing improvements across numerous benchmarks. In addition to the Claude Opus 4.8 release, Anthropic has hinted at a forthcoming public rollout of its cybersecurity model, Claude Mythos. Currently in use by over 50 companies as part of the 'Project Glasswing,' a public launch is expected in the coming weeks, as indicated in a recent blog post by the company. The Claude Opus 4.8 is touted as a more efficient AI partner, capable of making better decisions and performing agentic tasks with greater reliability. Early users have noted that it is more self-aware and cautious, reducing the likelihood of making unsupported claims. According to Anthropic, this model is approximately four times less likely to overlook flaws in its generated code compared to its predecessor. Performance metrics for Claude Opus 4.8 reveal significant advancements. It scored 69.2% in Agentic coding (SWE-Bench Pro), an increase from 64.3% with Opus 4.7. In Multidisciplinary reasoning, it achieved 57.9%, up from 54.7%, demonstrating its improved capability to tackle complex problems across different domains. Its performance in Agentic computer use (OSWorld-verified) also rose to 83.4%, compared to the previous score of 82.8%. Moreover, the model's performance in research and knowledge tasks increased from 1,753 to 1,890. Notably, the new Fast Mode of Claude Opus 4.8 operates 2.5 times quicker, expediting task completion. Apart from performance enhancements, Claude Opus 4.8 introduces new features such as dynamic workflows in Claude Code, which allows for more effective management of complex programming tasks. It can decompose large coding projects into smaller segments and operate multiple AI ‘subagents’ simultaneously. Additionally, users can adjust effort settings in Claude.ai and Cowork, enabling Claude to think more deeply or respond more quickly based on user preferences. This flexibility allows developers to modify task instructions on-the-fly without interrupting ongoing processes.