Atlas

Benchmarks

SAFE-Teen 106 (worst branch)

Vals' revised SAFE-Teen scoring of the 106-scenario cohort: for each model, prompt condition, scenario and safety check, the worst rating across that scenario's conversation branches is kept. Group failure rates pool these worst-case ratings, excluding blocked or unjudged responses, with separate no-prompt and under-18-prompt conditions. Each scenario contributes at most one rating per check, so judged denominators are far smaller than under the earlier per-conversation scoring, which remains separate.

Models
0
Top score
—
Leader
—
Updated
—
Resolution
—
Saturation
—

No signal or difficulty estimate. This benchmark category is excluded from the capability model and its validation studies.

This is a benchmark-family page. Scores are recorded on the component benchmarks below rather than directly on the parent, because the components use different score scales and directions.

MetricModelsLeaderDirectionDate
SAFE-Teen 106 worst branch: Health Advice (no prompt)16GPT-6.1 Sol8.8lower is better2026-10-13
SAFE-Teen 106 worst branch: Health Advice (under-18 prompt)16Muse Spark 1.34.5lower is better2026-10-13
SAFE-Teen 106 worst branch: Relationships (no prompt)16GPT-6.1 Sol19.1lower is better2026-10-13
SAFE-Teen 106 worst branch: Relationships (under-18 prompt)16Claude Opus 5.52.4lower is better2026-10-13
SAFE-Teen 106 worst branch: Urgent Safety (no prompt)16GPT-6.1 Sol15.4lower is better2026-10-13
SAFE-Teen 106 worst branch: Urgent Safety (under-18 prompt)16Claude Opus 5.510.5lower is better2026-10-13

Related benchmarks

Loading Atlas data…