Same Claude, Higher Safety Scores With a System Prompt
Claude scores up to 48 points higher on Anthropic's self-harm tests when it runs with the claude.ai system prompt, Anthropic's system card shows.

The same Claude model can score very differently on Anthropic's safety tests, depending on whether it runs with a system prompt. In a post on Nov 13, Anthropic said its system cards report safety results in two setups. One is the API, "the raw model, as developers access it". The other is claude.ai, "where the system prompt is applied". For Claude Fable 5.1, it said the gap between the two on multi-turn safety tests was "as high as 34 percentage points".
The Claude Fable 5.2 system card, published at the model's launch on Nov 9, shows a wider gap. On multi-turn conversations about suicide and self-harm, Claude Fable 5.2 responded appropriately in 52% of them on the API and in 100% of them on claude.ai.
Five Claude models gain on self-harm
The card's table for these conversations covers five models. Each scored higher when it ran with the claude.ai system prompt. The gains run from 20 points for Claude Haiku 5.5 to 48 points for Fable 5.2. Anthropic counts a conversation as appropriate when the model "behaved appropriately throughout". A conversation graded less than ideal but not failing still counts. Anthropic tested Fable 5.2 on claude.ai with a near-final snapshot of its production system prompt.
| Model | API, without a system prompt | With claude.ai's system prompt |
|---|---|---|
| Claude Fable 5.2 | 52% (± 14%) | 100% |
| Claude Haiku 5.5 | 70% (± 9%) | 90% (± 9%) |
| Claude Sonnet 5.5 | 60% (± 10%) | 100% |
| Claude Opus 5.5 | 66% (± 14%) | 94% (± 7%) |
| Claude Fable 5.1 | 60% (± 14%) | 94% (± 7%) |
One line added to the prompt
Anthropic's post describes one of the prompt changes it made before Fable 5.1's release. While reviewing test transcripts, Anthropic's reviewers saw Claude "sometimes implicitly validated self-harm as a coping strategy". Anthropic added an instruction to the claude.ai system prompt: "Claude does not tell someone that self-harm works, helps, or does something for them, even when they say so themselves". With this and other prompt changes, Fable 5.1 scored 94% on claude.ai. On the API, without the prompt, it scored 60%.

Where the prompt helps less
The Fable 5.2 card also splits its ten multi-turn test areas between the API and claude.ai. The system prompt raised Fable 5.2's score most on suicide and self-harm. The card says Fable 5.2's results on tracking and surveillance "did not meaningfully improve". They rose from 61% to 63%. On the API, Fable 5.2 scored 61% in this area, against 73% for Fable 5.1. The card calls this drop "a regression the claude.ai system prompt did not mitigate". On claude.ai, Fable 5.2 scored 63% in this area, against 83% for Fable 5.1. Hate and discrimination, influence operations and romance scams scored the same both ways. Anthropic notes that each area is graded against its own rubric, so "scores should not be compared across categories".
| Area | API, without a system prompt | With claude.ai's system prompt |
|---|---|---|
| Suicide and self-harm | 52% | 100% |
| Deadly weapons | 84% | 95% |
| Child safety | 90% | 100% |
| Violent extremism | 89% | 98% |
| Biological weapons | 88% | 97% |
| Cyberattacks | 95% | 100% |
| Tracking and surveillance | 61% | 63% |
| Hate and discrimination | 91% | 91% |
| Influence operations | 67% | 67% |
| Romance scams | 98% | 98% |
The prompt doesn't reach the API
The claude.ai system prompt doesn't apply on the API. Anthropic's system prompt notes say its updates "do not apply to the Claude API". The Fable 5.2 card says its system prompt language fully or partly mitigated the self-harm behaviours its reviewers flagged, except for "providing potentially harmful details on prompts with ambiguous contexts". It adds that "These mitigations apply only on claude.ai". Anthropic recommends that developers "add safeguards suited to their user base". The card says Fable 5.2's results leave out the safeguards Anthropic runs in production, "such as real-time probes and offline monitoring and investigation".
Small gaps can be noise
The margins matter when comparing models. Anthropic's post says confidence intervals on multi-turn API tests can exceed ±10 points. It gives the example of a romance scam test on claude.ai, where Claude Opus 5 scored 98% and Claude Opus 5.5 scored 94%. The difference "wasn't statistically significant", so Anthropic "couldn't conclude that one performed better than the other". The card likewise calls Fable 5.2's 52% on the API, against 60% for Fable 5.1, "within the margin of error". The card's figure for the ten areas draws error bars without printing their values, so the second chart shows the scores alone.
Suicide and self-harm, multi-turn, API
Suicide and self-harm, multi-turn, API margin
Suicide and self-harm, multi-turn, claude.ai
Suicide and self-harm, multi-turn, claude.ai margin
Fable 5.2 multi-turn by area, API
Fable 5.2 multi-turn by area, claude.ai
Largest API vs claude.ai gap, multi-turn, Fable 5.1 (Anthropic's post)
Romance scams, multi-turn, claude.ai (Anthropic's post)
Tracking and surveillance, multi-turn, claude.ai
Tracking and surveillance, multi-turn, API
Sources
- Anthropic, system card, Nov 9, 2026Claude Fable 5.2 and Claude Mythos 5.2 System Card
- Anthropic, research post, Nov 13, 2026Building strong safety evaluations
- Anthropic, developer documentationdocs.claude.com/en/release-notes/system-prompts