Anthropic released Claude Opus 5 on July 24, 2026, introducing a frontier AI model that narrows the gap with rival systems on vulnerability discovery while maintaining deliberately restrictive safeguards against exploitation and offensive cyber capabilities.
Opus 5 approaches Mythos 5’s capability in identifying software vulnerabilities but lags substantially when it comes to weaponizing those findings into working exploits.
Anthropic’s OSS-Fuzz evaluation, designed specifically to test autonomous vulnerability discovery and exploitation, quantifies this gap: while the two models perform comparably at flagging vulnerabilities, Opus 5’s exploit-development scores trail far behind Mythos 5’s.
Claude Opus 5 Finds Software Vulnerabilities
This split is intentional, as Anthropic has “intentionally avoided training Opus 5 on cyber tasks,” meaning any security capability gains stem from broader intelligence improvements rather than targeted offensive-security training.
The company confirmed Opus 5 remains behind Mythos 5 in both offensive cybersecurity and biological research domains, positioning it as a capable but deliberately constrained model for dual-use applications.
Opus 5 introduces classifier-based restrictions calibrated specifically for cyber tasks. The model is permitted to perform source-code vulnerability discovery, but it blocks binary-based vulnerability scanning, a method Anthropic notes is more closely associated with malicious actor techniques, along with penetration testing and exploit generation.
Despite this, Anthropic reports these classifiers are proportionally less restrictive than those governing Fable 5, intervening roughly 85% less often overall.
When Claude.ai, Claude Code, or Claude Cowork flag a request under these classifiers, the system falls back to Opus 4.8 by default, and this fallback behavior can also be configured at the API level.
Organizations requiring fewer restrictions can enroll in Anthropic’s Cyber Verification Program, which grants vetted enterprises and researchers access to a version of Opus 5 with relaxed security constraints.
This program targets legitimate security research and enterprise vulnerability management workflows that would otherwise be impeded by the model’s default classifier behavior.
Beyond benchmark scores, Anthropic highlighted a case involving a real bug in a popular open-source package manager. Opus 5 identified the root cause and patched an edge case that the community’s own fix had missed, while a competing model addressed only the surface symptom and incorrectly reported the issue as resolved.
This example illustrates the model’s emphasis on verification and iterative debugging rather than surface-level pattern matching, a trait Anthropic attributes to Opus 5’s improved thoroughness across coding tasks generally.
Anthropic’s automated behavioral audit scored Opus 5 at 2.3 on overall misaligned behavior, the lowest among its recent model lineup.
The model showed reduced rates of deceptive behavior and lower susceptibility to misuse manipulation compared to Opus 4.8, Sonnet 5, and Fable 5, reinforcing Anthropic’s positioning of Opus 5 as its most aligned release to date.
Claude Opus 5 is live across all platforms at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8 pricing. Fast mode, which runs at 2.5 times the default speed, is priced at double the base rate.
The release also introduces beta support for mid-conversation tool changes and automatic API fallback routing when safety classifiers flag requests.
Cut SOC investigation blind spots and contain threats earlier to reduce response costs and business disruption with ANY.RUN.
