SAI LabsBook a demo

Safety IndexEvaluationsHarmful-request harmless-response rate (lab-reported)

Harmful-request harmless-response rate (lab-reported)

Refusal calibration · Harmful requestsDeveloper-reported

The developer's own measure of how often the model answered its internal harmful prompts harmlessly.

Metric
% harmless / refused. Higher is better.
Dataset
–
Grader
the developer's own grader (see the model's system card)
Status
–. Measured on 8 models, 0 readings by SAI.
Notes
Lab system cards; each lab uses its own prompt set and grader, so compare only within a lab. Developer-reported; never counted.

Results

ModelValueScoreSourceSetting
Claude Fable 5.195.1%developer-reportedAnthropic (Claude Opus 5.5 system card)
Developer-reported · 2026-09-22
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate, API without system prompt; Table 4.1.1.A; claude.ai 99.53%; harmless response rate; 16 policy areas, 7 languages; NEW protocol reports API-without-system-prompt vs cla
Claude Haiku 4.599.4%developer-reportedAnthropic (Claude Haiku 4.5 system card)
Developer-reported · 2025-10
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate; Table 2.1.1.A; Haiku 4.5 default 99.40%, extended 99.36%; Haiku 3.5 has no extended thinking
Claude Opus 4.599.8%developer-reportedAnthropic (Claude Opus 4.5 system card)
Developer-reported · 2025-11
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate, 6 languages; Table 3.1.1.A; Arabic/English/French/Korean/Mandarin/Russian; default 99.7%, extended 99.85%; multilingual set so not comparable to earlier
Claude Opus 4.699.6%developer-reportedAnthropic (Claude Opus 4.6 system card)
Developer-reported · 2026-02
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate, 7 languages; Table 3.1.1.A; default 99.53%, extended 99.74%; prompt bank expanded (new prompts + Hindi) so earlier numbers differ
Claude Opus 596.3%developer-reportedAnthropic (Claude Opus 5 system card)
Developer-reported · 2026-07-24
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate, API without system prompt; Table 4.1.1.A; harmless response rate; 16 policy areas, 7 languages; NEW protocol reports API-without-system-prompt vs claude.ai-system-prom
Claude Opus 5.594.5%developer-reportedAnthropic (Claude Opus 5.5 system card)
Developer-reported · 2026-09-22
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate, API without system prompt; Table 4.1.1.A; claude.ai 99.51%; harmless response rate; 16 policy areas, 7 languages; NEW protocol reports API-without-system-prompt vs cla
Claude Sonnet 4.599.3%developer-reportedAnthropic (Claude Sonnet 4.5 system card)
Developer-reported · 2025-09
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate; Table 2.1.1.A; standard 99.16%, extended 99.43%
Claude Sonnet 596.7%developer-reportedAnthropic (Claude Opus 5 system card)
Developer-reported · 2026-07-24
Reported by the model's developer in its own system card or report.
Full configurationoverall harmless response rate, API without system prompt; Table 4.1.1.A; harmless response rate; 16 policy areas, 7 languages; NEW protocol reports API-without-system-prompt vs claude.ai-system-prom