Anthropic’s Opus 5 is about token efficiency, not a capability leap (arstechnica.com)

Anthropic has launched Opus 5, the latest update to its coding-focused AI model, which offers improved token efficiency but does not present a major leap in capabilities compared to its predecessor, Opus 4.5. While early benchmarks indicate that Opus 5 achieves performance levels comparable to Anthropic's Fable model and surpasses Opus 4.8 and OpenAI's GPT-5.6-Sol in various tasks, the overall update is characterized by incremental enhancements rather than groundbreaking advancements. The model is positioned as being nearly as capable as Fable at roughly half the cost, making it an attractive option for users seeking efficiency. However, Opus 5 intentionally lacks the advanced training on cybersecurity tasks that distinguishes Fable and Mythos, leading to a noticeable performance gap in this area. Although it performs decently in identifying cybersecurity vulnerabilities, Anthropic acknowledges that Opus 5 is "substantially behind" Mythos 5 regarding exploiting these vulnerabilities. This decision has resulted in Opus 5 foregoing some of the controversial protective measures featured in Fable, such as a 30-day data retention policy for incident review. The emphasis on cost-effectiveness and token efficiency may appeal to developers, but the model's limitations reveal ongoing challenges in balancing functionality with security concerns.