Claude Opus 5.5 Intelligence, Performance and Price Analysis

Commenters debated Claude Opus 5.5’s reasoning settings and practical value. Simonw and others reported the "max" reasoning setting can overthink and exhaust the 128,000-token budget on tasks like "generate an SVG of a pelican riding a bicycle," with simonw sharing a reasoning-trace transcript; samuelknight and sidewndr46 echoed that max is for benchmaxxing and not production, while mchusma and Someone1234 argued High or Medium often deliver the best intelligence-per-dollar and are more useful in real workloads. Hglaser highlighted lower cost-per-task versus Opus 5, user43928 noted Astra High’s slightly lower price, and sharktheone and firemelt offered mixed hands-on impressions about code-quality and practical improvement.

Opinion split on benchmarking reliability, model drift, and training exposure. Breckenedge and seabass-salmon warned of regressions after launch, calling for repeated evaluations; kzrdude and losvedir defended broad comparative benchmarking but onlyrealcuzzo criticized the UI/metric clarity and cost-per-task interpretation. RGS1811, croemer, and cubefox discussed whether the model’s recognition of the pelican task implies explicit bench-targeting or just pretraining exposure. Others (bkishan, meric_) described Anthropic’s quieter release strategy, while giancarlostoro framed pricing as cost-driven rather than greed.