Sonnet 5.5 vs Opus 5.5: Blender animation test shows Sonnet's rigs break on playback
Given the same object description, Sonnet 5.5 and Opus 5.5 build and rig it in Blender automatically; on playback Sonnet's joints and weights visibly break, so the author routes anything that moves to Opus.
Anthropic’s Sonnet 5.5 costs half of what Opus 5.5 does, thinks longer on max effort, and its builds still break once Blender hits play. A still frame flatters both models. Only the timeline shows which one understood how the object is meant to move. > BRIEF - one identical object description, both models set to max effort > BUILD - mesh, materials and hierarchy assembled in Blender with no human edits > PLAY - every rig run through a full animation instead of judged on frame one > CLOCK - wall time and spend logged for each build, prompt to final file On frame one the outputs were hard to tell apart: matching silhouettes, clean topology, sensible material slots. Motion split them within seconds. Sonnet’s joints bent where the object never would, and weights bled into rigid parts. Max effort did not close that gap. Sonnet spent more time reasoning on every build and still landed a weaker rig. Effort is a dial for search time, not for capacity. Turning it up let Sonnet look longer, never deeper. Static work is where Sonnet makes sense. Props and scenery that never move came back close enough at half the bill. So the routing rule writes itself: anything that moves goes to Opus, anything that sits still can go to Sonnet. Scene by scene, with timings and the real cost of each build, the full scorecard follows.
This post quotes another post: x.com/0xcodio/status/2108134587633221885