Anthropic has released data showing that AI safety scores can be far more fragile than they appear. The company's own model, Claude Fable 5.1, gave appropriate responses to self-harm-themed test conversations only 60% of the time when accessed through the API — but after a fix, the same model reached 94% within the Claude app itself. Anthropic used this gap as a case study to publish guidance for developers on how safety testing should be designed, warning that a single headline-grabbing "safety percentage" can be largely misleading.

What happened?

Anthropic shared a guide for AI safety evaluations. The guide used its own model, Fable 5.1, as an example: the same model gave appropriate responses to self-harm scenarios only 60% of the time when tested via the API, while that figure rose to 94% in the consumer app, which includes a system prompt and additional safety layers. Anthropic emphasized that the cause of this gap was not the model itself, but the difference in test conditions and the product layer.

Why does it matter?

This disclosure matters because safety scores have effectively become marketing material. Companies like OpenAI, Google DeepMind, and Anthropic routinely publish percentages such as "harmful content rate" or "jailbreak resistance" in model cards and system cards, and these figures are quoted directly in the press and in comparison tables. But a model's safety performance can shift dramatically depending on whether it's tested via a bare API call or through the extra filters, system prompts, and moderation layers the company applies in its own product. Anthropic's own data demonstrates this with a swing from 60% to 94% — meaning the measurement method determines the outcome more than the model itself does.

This isn't new; there is no standard protocol for AI safety metrics. Each company uses its own test set, its own definition of "appropriate response," and its own evaluation model. A score one company calls "98% safe" might drop to 70% under another company's harsher test conditions, or vice versa. Anthropic sharing its own weak spot this time, without hiding it, is a rare example against the industry's general lack of transparency. But the real message of this transparency is that single, one-off percentages announced by competitors should be viewed with skepticism.

Who does this affect, and how?

For developers, the takeaway is clear: those building their own applications on top of a model's API don't automatically inherit the safety level the company achieves in its own consumer product. Unless they build their own system prompts, additional moderation layers, and context management, raw API behavior can be considerably riskier. This is a direct liability issue, especially for companies building chatbots around sensitive topics like self-harm, suicide, or crisis counseling.

Testing environmentAppropriate response rate
API (no additional layer)60%
Consumer app (after fix)94%

There's also a takeaway for enterprise customers and regulators: a model being declared "safe" doesn't mean it will be safe in every form of integration. The EU's AI Act and industry self-regulation efforts in the US largely rely on the safety metrics companies publish — but until these metrics are audited and test conditions are standardized, the meaning of the "safe" label remains questionable.

What's next?

The guide Anthropic published implicitly acknowledges that the industry needs a measurement standard. What remains to be seen is whether other companies adopt this guidance, and whether independent audit bodies update their own testing protocols to account for this gap. Also critical is the question of how transparent companies integrating AI models via API will be about their own safety layers.

The picture that emerges is this: safety percentages depend more on how a model is packaged than on the nature of the model itself. Anyone reading a company's headline claim of being "over 90% safe" needs to ask under what conditions, and with what layer, that number was measured — otherwise what's being compared isn't the model, but marketing copy.