Claude Sonnet 5 upgrades are not uniform across every skill.
Rohan Paul Twitter · Rohan Paul (@rohanpaul_ai) · 2026-06-30
Claude Sonnet 5's capability improvements over Sonnet 4.6 are domain-uneven, with a notable regression on CyberGym vulnerability-discovery benchmarks that Anthropic attributes to deliberately omitting targeted cyber training in favor of general reasoning.
Appears in
Extraction
Topics: claude-sonnet-5model-evaluationcybersecuritybenchmark-regression
Claims
- Claude Sonnet 5 performs worse than Sonnet 4.6 on CyberGym, which evaluates vulnerability discovery and exploit-finding behavior.
- Anthropic explicitly stated that Sonnet 5 was not deliberately trained for cyber tasks.
- Sonnet 5's cyber benchmark performance derives from general reasoning ability rather than targeted optimization for exploit skills.
Key quotes
Sonnet 5 upgrades are not uniform across every skill. e.g. its weaker than Sonnet 4.6 on CyberGym
Anthropic also explicitly said in its announcement blog that Sonnet 5 was not deliberately trained for cyber tasks, so its cyber ability likely comes from general intelligence rather than targeted optimization.
So Sonnet 5's performance on CyberGym comes from general reasoning rather than specialized exploit skill.