Tested it all morning, and so far it feels close to Opus 4.7 (still needs further testing). Using M3 to write code, with Opus 4.8 + GPT5.5 for adversarial code reviews—pretty decent results. 1 PR completed already.