Jay Akin on X: "@danshipper For complex coding tasks, Opus 4.8 at times assumed things that were incorrect. Those hallucinations were otherwise easy to check in the codebase. Almost never seen that with GPT 5.5. Also Opus 4.8 is harder to steer. This is probably because of the system prompts of the claude" / X
Replying to @danshipper
For complex coding tasks, Opus 4.8 at times assumed things that were incorrect. Those hallucinations were otherwise easy to check in the codebase. Almost never seen that with GPT 5.5. Also Opus 4.8 is harder to steer. This is probably because of the system prompts of the claude code harness, but still an issue, since no other harness is cost effective for Anthropic models.
1