Key Takeaways

  • Claire Vo tested Anthropic's Fable 5.2 against an unlabelled flat-pack furniture delivery containing 25 wooden pieces, an open bag of screws, and zero instructions.
  • Every existing frontier model failed the test days earlier, unable to map the physical assembly sequence or geometric relationships.
  • Fable 5.2 derived accurate step-by-step build diagrams from low-resolution floor photos containing Chinese labels.
  • Vo verified the output in the real world: a professional assembler built the unit by following the model's exact sequence.
  • The test highlights a split in multimodal performance: strong spatial reasoning and 3D geometry paired with continuing weaknesses in unassisted web design and image generation.

The Drop-Ship Test That Broke Every Frontier Model

Benchmarking multimodal models on synthetic vision tests rarely exposes whether software understands the physical world. Real understanding shows up when the input is messy, uncurated, and missing documentation.

Vo encountered this problem when an order arrived at her home. “I got a very shady pile of furniture straight drop shipped from China delivered to my house and it showed up with an open bag of screws, 25 pieces of wood and no instructions,” (play from 11:21) Vo said. Days before Fable 5.2 launched, she ran the problem through every leading model on the market. She photographed the parts scattered across her floor: wooden boards marked only with Chinese text, along with loose fasteners. Every model failed to produce a coherent plan.

Most visual models operate as pattern-matching engines over 2D pixels. They identify a board, a screw, or a bracket, but they fail to determine how those parts constrain one another in three dimensions. When forced to deduce physical relationships without a manual, earlier models produced hallucinated hardware counts and physically impossible joints.

Solving Assembly Order Through Geometric Constraints

When Vo uploaded the same low-resolution photos to Fable 5.2, the model decoded the Chinese labels and mapped the structural joints. More importantly, it determined the correct dependency tree for assembly.

“If you've ever done complex IKEA or flatpack furniture, you know that the order in which you put stuff together is really important,” (play from 11:56) Vo explained. Fasten the exterior frame too early, and you cannot seat the interior shelving dowels. Secure the base panel backwards, and the structural load fails.

Fable 5.2 generated step-by-step diagrams that matched physical build rules. The validation was practical rather than theoretical. Vo hired a professional to build the furniture, and they tested the AI-generated sequence on the floor. “I followed these instructions with the guy that came to my house to assemble it. It did it in the right order,” (play from 12:07) she said. The model succeeded from low-grade visual inputs where Opus 5.5 and other predecessors stalled.

This spatial capability sits alongside clear limits. While Fable 5.2 excels at 3D rendering, video animation, and geometric deductions, Vo noted it still struggles with unassisted web design and raw image generation. Builders should treat spatial deduction as a dedicated capability rather than assuming uniform general intelligence across design tasks.

What to Do With This

Replace your synthetic computer vision benchmarks with an internal physical test suite this week. Photograph five disassembled hardware products or mechanical components on an office floor, stripped of manuals. Run these images through your current multimodal pipeline to test whether your models can infer mechanical dependencies and correct build sequences before deploying vision agents into physical workflows.