Anthropic released Claude Opus 5 on Friday, positioning it as a lower-cost alternative to its flagship Fable 5 model. On cybersecurity benchmarks, the company says the new model comes close to matching Mythos 5, its most capable system, at identifying software vulnerabilities, but trails significantly when it comes to developing those findings into functional exploits.

The comparison comes from Anthropic’s own OSS-Fuzz-based evaluation, which measures how effectively a model can locate and then weaponize vulnerabilities with minimal human guidance. According to the company, Opus 5’s vulnerability-detection rate approaches that of Mythos 5, but its exploit-development score lags considerably behind.

Anthropic describes the gap as intentional rather than a shortfall. The company says it has not trained Opus 5 directly on offensive cyber tasks, and that any exploit-related gains the model shows are a side effect of general capability improvements rather than targeted training.

Loosened Safety Classifiers, Same Guardrails on Offense

Opus 5’s safety classifiers are tuned to intervene less often than those on Fable 5, with Anthropic expecting roughly 85 percent fewer interventions overall. The model is allowed to search for vulnerabilities directly within source code, but binary-based vulnerability scanning, penetration testing, and exploit generation remain blocked by default.

Requests that trigger those restrictions automatically fall back to the older Opus 4.8 model across Claude.ai, Claude Code, and Claude Cowork. Organizations and researchers enrolled in Anthropic’s Cyber Verification Program can access a version of Opus 5 with those limits relaxed further.

Mythos 5 Still Withheld

Mythos 5 remains unavailable for general release. Fable 5 is described by Anthropic as the safeguarded, publicly accessible counterpart built on the same underlying model. Both Mythos 5 and Fable 5 were briefly taken offline last month amid concerns from the Trump administration over their potential use by foreign nationals, restrictions that have since been lifted.

Pricing for Opus 5 holds steady at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8. A faster-response mode is available at double the base rate.

For security teams evaluating AI-assisted vulnerability research tools, the distinction between detection and exploitation capability remains a key factor in assessing both defensive value and potential misuse risk.