Nickson reports that Claude's bridge held an estimated 130 lb, ahead of Meta's 26.5 lb and OpenAI's estimated 17.5 lb. Grok and Kimi did not complete the challenge.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [X](https://x.com/rpnickson/status/2104234974350111108)
Why it matters #
The exercise tests whether AI-generated designs can be printed, assembled and bear a load. In Nickson's reported comparison, Claude's bridge led on both load capacity and printing efficiency; the results describe these designs under his test conditions, not a universal ranking of the models' engineering ability.
Anthropic's Claude Opus 5.5 won Roberto Nickson's (@rpnickson) five-model 3D-printed bridge challenge, according to the results he reported in a September 27th post on X. Its bridge held an estimated 130 lb, compared with 26.5 lb for Meta's design and an estimated 17.5 lb for OpenAI's. The Grok and Kimi designs did not complete the challenge.
https://x.com/rpnickson/status/2104234974350111108 The lineup was Anthropic's Claude Opus 5.5, Moonshot AI's Kimi K3, Meta's Muse Spark 1.3, OpenAI's GPT-6 Astra and SpaceXAI's Grok 4.7. Nickson listed all five at "High" reasoning effort.
The rules and results
Each bridge had to span two feet, use no more than 500 grams of filament and take less than 18 hours to print. Nickson said the task also specified the type of weight and where it would be placed on the bridge.
His reported results were:
- Claude Opus 5.5: Held an estimated130 lb . The bridge took 9 hours and 11 minutes to print, used 17 parts and consumed 441 grams of filament.
- **Meta [Muse Spark 1.3](https://runtimewire.com/models/azure/muse-spark-1.3):** Held**26.5 lb** . Printing took 13 hours and 12 minutes, with 49 parts and 478 grams of filament.
- **OpenAI [GPT-6 Astra](https://runtimewire.com/models/openai/gpt-6-astra:batch):** Held an estimated**17.5 lb** . Printing took 15 hours and 44 minutes, with 29 parts and 442 grams of filament.
- SpaceXAI Grok 4.7: Did not finish. Nickson said the assembled bridge could not stand on its own. It took 12 hours and 22 minutes to print 29 parts, using 460 grams of filament.
- Kimi K3: Did not finish. Nickson said the design could not be assembled because it was not engineered correctly. Printing took 11 hours and 19 minutes, with 35 parts and 446 grams of filament.
Claude's design led on more than load capacity: among these five entries, it also had the shortest print time, fewest parts and lowest filament use. Nickson named it the winner of the second episode of "AGENT WARS."
These are Nickson's reported results for the particular designs and printing conditions in this challenge. The Claude and OpenAI load figures were explicitly described as estimates. A physical build adds assembly and fabrication constraints that a text-only evaluation misses, but this comparison alone does not establish a general ranking of the models' engineering abilities.
The creator's disclosure
Nickson addressed possible ties to the model makers in a reply, saying he had worked "in a paid capacity" with all of the companies represented except Kimi. He said the episodes would be "completely honest tests." His post does not specify what work he did for the four companies or whether any of them supplied models, equipment or production support.
Replying to another user, Nickson said the first episode tested models on marketing websites and took place "a few months ago"; he said GPT-5.6 Sol won that round and linked an Instagram Reel of the episode. The bridge challenge changes the deliverable from a website to a physical object that must be printed, assembled and loaded.
Nickson has built and promoted consumer technology before. In a 2014 Washington Business Journal profile, he was identified as founder and CEO of Arlington startup PicLab. The publication reported that PicLab's photo-editing app reached Apple's U.S. top 50 despite having no marketing budget; Nickson was preparing to release its video app, VidLab.