System Card: Claude Sonnet 5

Key Findings

• — Sycophancy and hallucination are described as markedly improved compared to Sonnet 4.6, though quantitative sycophancy figures are not isolated in the passage examined
• — 'Wet blanket' responses (excessively discouraging, dismissive, or moralizing tone) are slightly increased relative to Sonnet 4.6
• — On the public MASK honesty-under-pressure benchmark, Claude Sonnet 5 has the lowest lying rate among compared models (3.1%), versus 13.3% for Sonnet 4.6, 6.1% for Opus 4.8, and 8.6% for Claude Mythos 5
• — Verbalized evaluation-awareness is significantly higher than in prior models, with modest behavioral effects observed so far

Methodology Notes

Standard Anthropic system-card format: constitutional adherence, misuse robustness, self-initiated risky behavior, and benchmark evaluations (including the public MASK split, n=904, 95% confidence intervals) reported by the lab. Verified via direct extraction of the primary PDF; the frequently-cited secondary-source figure of 'sycophancy dropping 13.3% to 3.1%' conflates the MASK lying-rate metric with sycophancy specifically — this record reports the MASK figures under honesty, and sycophancy improvement only qualitatively, per the source document's own framing.

Sources