Anthropic — AI Safety Level: ASL-3 (Opus 4+), ASL-2 (Sonnet/Haiku)

NoteThe claim states 'ASL-3 (Opus 4+), ASL-2 (Sonnet/Haiku) (as of 2025-05)' with a note about 'First activation of ASL-3 for Claude Opus 4; precautionary/provisional.' The source directly confirms all three elements: (1) Claude Opus 4 is deployed under ASL-3 as of May 22, 2025; (2) Claude Sonnet 4 remains ASL-2; (3) Haiku implicitly remains ASL-2 as part of the baseline; (4) the activation is explicitly described as 'precautionary and provisional.' The date matches (2025-05 = May 2025). The 'Opus 4+' notation in the claim is reasonable given the source's language about Opus 4 being the first model to trigger ASL-3, though the source does not explicitly discuss future models.

NoteThe source directly confirms the claim's core assertions: (1) Claude Opus 4 is deployed with ASL-3 protections as of May 22, 2025, (2) this activation is explicitly described as 'precautionary and provisional', (3) Sonnet (Sonnet 4 specifically mentioned) does not require ASL-3 and remains at ASL-2 baseline, and (4) Haiku is not mentioned but the source confirms Sonnet does not meet ASL-3 threshold. The date matches the claimed 2025-05 timeframe. The only minor gap is that the source doesn't explicitly mention Haiku's ASL level, but it does confirm Sonnet's ASL-2 status and Opus 4's ASL-3 status with the precautionary/provisional qualifier as claimed.