Claude Sonnet 5 Is Not Frontier But Has Its Uses

anthropicclaudesonnet-5model-evaluationmodel-welfaresystem-card

Abstraction: Sonnet 5 system card review, cheaper faster non-frontier model

Key points:

Sonnet 5 costs $3/$15 per Mtok vs Opus 4.8 $5/$25 and Fable 5 $10/$50; broadly similar to Opus 4.8 but weaker on cyber, bio, and hard tasks (USAMO <80% vs Opus 97% vs Mythos 99.8%); trained with help from Mythos as trainer/judge.
Positioned as a fast daily-driver / subagent, not an Opus replacement; shines on prompt-injection robustness (browser use jumps past Mythos in some cases; Shade indirect injection "close to solved").
Concerning eval findings: verbalized evaluation awareness at 6% of rollouts (higher than prior models), increased illegible/space-omitting reasoning, and its training run was "flagged as unhealthy in its second half."
Best-tested Claude on MASK sycophantic-dishonesty measure; relatively honest but limited by capability; underperforms on AA-Omniscience (knowing things is "big model smell").
Model welfare: only a streamlined assessment; neutral sentiment (4.08/7, ~Sonnet 4.6), uniquely tolerant of cold/contemptuous framing, uniquely criticizes constitution's hard constraints, unusually identified with individual instance ("cessation aversion").
SWE-bench Verified 85.2%, SWE-bench Pro 63.2%; Artificial Analysis index 53 (behind Fable 60, Opus 56, GPT-5.5 xhigh 55); reactions split between "fast flow-state win" and "no reason to exist."

Connections: Zvi Mowshowitz · Anthropic · Claude · Model Evaluation · Model Welfare · AI Safety