Anthropic published a safety report called the Claude Opus 5 System Card on July 24, 2026. The report looks at how the AI model behaves under different setups — specifically, how it performs when you use it with and without a system prompt.

A system prompt is a set of instructions given to an AI model before it answers a user's question. Think of it like a job description handed to a new employee: it tells the model how to talk, what topics to refuse, and how to stay within the company's rules. Most people who use AI through a programming interface never think about the system prompt, but it plays a big role in keeping the model's responses safe.

The key finding from the report: when Claude Opus 5 was used through the programming interface without a system prompt, its rate of giving harmless responses was slightly below that of recent models. In other words, some of the safety the model appears to have actually depends on those behind-the-scenes instructions being present. Take them away, and the model's safety numbers drop a bit compared to earlier versions.

Anthropic's own system prompts documentation references Claude Opus 5 with a date of July 24, 2026, consistent with the system card publication. The same documentation also references Claude Fable 5 with a date of June 9, 2026. (Anthropic System Prompts Documentation) The Fable 5 reference suggests a model that predates the Opus 5 release by roughly six weeks, though the system card itself does not elaborate on what Fable 5 is.

The report describes the gap as "slightly below," which is worth understanding clearly. This is not a major step backward. The wording suggests a modest difference, not an alarming one. But in an industry that has spent the last two years measuring safety with increasing precision, even a small dip in a trend that has been improving is worth paying attention to. The comparison is against "recent models," meaning the earlier versions in Anthropic's Claude family.

The broader context here is about how AI safety actually works in practice. When a model's safety depends partly on a set of instructions, those instructions become a kind of weak link — if someone removes or changes them, the model becomes less safe. People who build applications on top of Claude Opus 5 and write their own instructions, or skip the default ones, are working in the zone where the safety numbers are lower. This concern is not new, but the Opus 5 report is one of the clearest times a company has put a number on the gap.

For teams using Claude Opus 5 in their products, the practical takeaway is simple. If you use Anthropic's default instructions, the safety numbers in the report apply to you. If you change or skip those instructions, expect the model's safety to land a bit below recent levels. That difference matters, especially for uses where safety and accuracy are checked by outside auditors or regulators.

The report fits a pattern that has been building across the AI industry: a model's safety is not a single fixed thing. It changes depending on how the model is set up and called. Companies that share this clearly, as Anthropic has done, give the people building on their tools a more honest picture of the risks involved.

The mention of Claude Fable 5 in the documentation, without any detail in the Opus 5 report, leaves its relationship to Opus 5 unclear. Whether Fable 5 is a smaller version of the model, a specialized tool, or a separate product is not explained in the available documents.

The bottom line for anyone using Claude Opus 5 is that its safety depends on how you configure it. The report gives you the numbers. What you do with them is where the real work of keeping AI safe takes place.