Anthropic has introduced Claude Opus 5, presenting it as a cost-effective alternative to the high-end Fable 5 model. Released on Friday, this new AI model is almost as proficient as the advanced Mythos 5 in identifying software vulnerabilities but lacks in converting those findings into actionable exploits.
Performance Evaluation and Limitations
The performance gap between Opus 5 and Mythos 5 is highlighted in Anthropic’s OSS-Fuzz-based evaluation. This assessment measures the capability of each model to pinpoint and exploit vulnerabilities with minimal human intervention. While Opus 5 closely matches Mythos 5 in identifying vulnerabilities, its ability to develop exploits significantly lags behind.
According to Anthropic, this limitation is intentional. The company has refrained from training Opus 5 specifically for offensive cyber tasks. The improvements observed in vulnerability detection are attributed to overall advancements in the model’s capabilities.
Operational Features and Restrictions
Opus 5’s safety classifiers are designed to be less restrictive than those in Fable 5, resulting in approximately 85% fewer interventions. Although the model can directly search for vulnerabilities within source code, functions such as binary-based vulnerability scanning, penetration testing, and exploit generation are restricted. Requests triggering these classifiers revert to the previous Opus 4.8 model within platforms like Claude.ai, Claude Code, and Claude Cowork.
For enterprises and researchers participating in Anthropic’s Cyber Verification Program, a version of Opus 5 with fewer restrictions is available, offering more flexibility in its application.
Market Position and Future Prospects
Despite its capabilities, Mythos 5 remains unavailable for general release, while Fable 5 serves as its safeguarded, publicly accessible counterpart, built on the same foundational model. Both were temporarily withdrawn last month due to concerns from the Trump administration regarding their potential use by foreign entities. These restrictions have since been lifted.
Opus 5 is priced competitively at $5 per million input tokens and $25 per million output tokens, mirroring the pricing of Opus 4.8. A faster-response option is available at double the standard rate, providing flexibility for users with varying needs.
As Anthropic continues to refine its AI models, Opus 5 stands out as a practical choice for entities seeking a balance between cost and cybersecurity efficacy.
