Anthropic launches Claude Opus 5, skips cyber-task training
Anthropic released Claude Opus 5 on July 24, 2026, with Simon Willison noting relentlessly proactive behavior on Frontier-Bench tasks such as 3D model generation. The launch also confirms Anthropic is intentionally avoiding training Opus 5 on cyber tasks to limit autonomous agent risk.
Key takeaway
Opus 5 pairs a more proactive agent stance with an explicit refusal to train on cyber tasks, a rare public tradeoff between autonomy and attack-surface control.
Context
Coverage of the July 24 release frames Opus 5 as Anthropic’s new top Opus tier, highlighting unusually proactive completion of Frontier-Bench workloads rather than a quiet incremental bump.
The confirmed choice to exclude cyber-task training is the strategic signal for builders: expect stronger agent initiative in general workflows, and do not assume cyber-offense or network-defense competence from that training mix.