Anthropic has unveiled Opus 5, a new version of its AI model that offers capabilities close to its top-performing Fable 5 model but at half the cost. According to Anthropic, Opus 5 excels in knowledge work, shows significant improvement in complex coding tasks compared to Opus 4.8, and is better suited for scientific research, particularly in predicting protein sequence variations.

Opus 5 also boasts enhanced resistance to manipulation and exhibits lower rates of deceptive behavior. Unlike its more powerful counterpart, Mythos, Opus 5 was not trained on cyber-related tasks, positioning it as a more accessible tool for general work rather than vulnerability exploitation. Anthropic aims for Opus 5 to be the go-to model for everyday tasks.

A key advantage over Fable 5 is Opus 5's less restrictive safeguards. While Fable 5's restrictions sometimes made it unusable, Opus 5's safeguards are designed to allow beneficial uses in cybersecurity and biology, with classifiers expected to intervene around 85 percent less often than with Fable 5. Opus 5 is not included in Anthropic's new data retention policy.

API costs for Opus 5 are set at $5 per one million input tokens and $25 per one million output tokens. Users can also adjust an "effort" setting to prioritize thoroughness or speed. This pricing positions Opus 5 competitively, even against models like China's Moonshot Kimi K3.