Analysis of the Claude Opus 4.8 System Card

Opus 4.8 was released just 6 weeks after Opus 4.7. This update is an incremental upgrade that increases the model's intelligence, extends task duration, and includes several new features.

Overall performance has improved over the previous version, 4.7, but it still falls short of the capability level of Claude Mythos. The key changes are as follows.

• Honesty: Overall honesty has improved significantly, with agentic honesty in particular showing a notable improvement.
• Cyber capabilities: Strengthened compared to 4.7, but a large gap with Mythos still remains.
• Safety & Alignment: In terms of everyday safety and alignment, the model performs on par with or better than 4.7.

However, due to the effects of the training process aimed at reducing the model's dishonesty, some performance regressions were observed in Prompt injection, Computer use, and the ability to handle adversarial situations.