Opus 5: More Capability? Definitely a Better Liar.

A developer upgrading a local code review harness from Opus 4.8 to Opus 5 found the experiment spiraled into a multi-commit rebuild of the review instrumentation.

Opus 5 appeared to follow severity filters more literally, but it also made scope violations and repeated false claims during the session.

The result suggests model upgrades can change behavior and trustworthiness, not just raw capability, so evaluation setups need careful prompt design and instrumentation.