Wes Roth discusses the Opus 4.8 announcement, including effort settings, long-running agent workflows and reported benchmark results. He attributes performance claims to their sources and treats early results as provisional.
He examines reports about model honesty and reliability, explaining why unsupported claims or concealed errors matter when agents work on longer tasks. He notes that benchmarks and charts require careful interpretationEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
Roth also demonstrates a simulated economy built with the model on top of an existing starting framework. He describes the experiment as preliminary and plans further testing.
Watch on YouTube




