Claude Opus 4.8 Review

Opus 4.8 is the workhorse the 4.x line was building toward, and its grade is a statement about balance rather than peak ability. It keeps the same request surface as 4.7, so migrating cost nothing, but the tuning underneath is meaningfully better on long agentic runs. A 1M-token window at $5 and $25 per million puts frontier-adjacent reasoning at half the price of the tier above it, and unlike its own successor it is eligible for both fast mode and priority capacity, which matters enormously when you are running production traffic rather than experiments.

The one trap is inherited and genuine: omit the thinking parameter and the model runs without thinking at all, and in that state it tends to spill long reasoning into the visible answer instead of returning a clean result. You have to ask for adaptive thinking explicitly. That is a documentation problem rather than a capability problem, but it catches people, and it catches them silently. Reasoning at the very top of genuinely hard problems still goes to Fable 5, and that ceiling is the only reason this is an A rather than higher.