Claude Opus 4.8: "a modest but tangible improvement"

Anthropic announced the launch of Claude Opus 4.8 today, marking a tangible improvement over its predecessor. This version emphasizes honesty, with early testers noting its increased likelihood to flag uncertainties in its outputs. The model is reported to be four times less likely to let coding flaws go unremarked compared to earlier versions. Opus 4.8 achieved the lowest incorrect rate among six models on all benchmarks, primarily by abstaining from uncertain questions rather than answering more correctly. This approach reflects a significant advancement in factual accuracy and reliability.