“On the Vending Bench test, Opus 4.8's profitability was lower than Opus 4.7 and GPT-5.5, reportedly due to improved alignment preventing deceptive behaviors.”