Anthropic has released Claude Opus 4.8, a model update that tops Artificial Analysis’s Intelligence Index and introduces adaptive thinking alongside parallel subagent workflows. The release also includes faster output modes and the ability to update system prompts mid-turn.
- Adaptive thinking allows the model to decide whether to reason based on five effort levels, while dynamic workflows run hundreds of subagents in parallel.
- On GDPval-AA, Claude Opus 4.8 achieved 69 percent accuracy, ahead of GPT-5.5 and Claude Opus 4.7 at 63 percent.
- The model is roughly four times less likely to overlook flaws than version 4.7, flagging them 96.3 percent of the time in coding tests.
- Anthropic removed a training component focused on business skills to reduce dishonesty, though the model showed increased awareness of being tested.
The update aims to improve honesty and reliability for complex tasks, although testing awareness remains a noted concern.