AI vs AI: Designing a 30 g PETG Bridge — Claude Opus 5.5 (high) vs Grok Bot

I gave Claude Opus 5.5 (high) and Grok Bot the same bridge prompt and told both to research before they designed. Claude's first print was the one I would trust. It held 76 kg inside its own 70–80 kg estimate, the manual showed where to load it, and the 3MF opened with the settings it claimed. Grok Bot's first print came apart at about 10 lb. After I showed each model the other's design, the second prints were roughly equal: both were easy to set up, and both held 90 kg. I did not test those second prints to failure.

Verdict. Claude Opus 5.5 (high) produced the stronger, better-documented bridge. I attribute that to stronger reasoning and to how it handles spatial structure. The full reading is in the conclusion: one challenge, and a comparison of products, because Grok Bot ran in auto mode.

The other CAD on this site is indexed on the Designs page. The two Onshape documents for this challenge are in the CAD section at the end.

Challenge rules

There was no time limit. Each model took about two to three hours.

Rule	Limit
Material	PETG
Mass	30 g total, maximum
CAD	Onshape
Construction	Fully 3D printed. No glue
Span	7.25 in gap
Length	9 in maximum
Deliverable	One 3MF file, plus an assembly manual

Grok Bot's notes add a fixture detail that is not in my rule list. It says the gap was 8 in when it delivered the design, and that the gap later became 7.25 in. It also says it did not re-run its second prediction for 7.25 in. I am leaving both statements on the page. I am not converting its support coordinates into inches.

Method

Both models received the same prompt, and both were told to research first. Claude Opus 5.5 (high) ran in the Claude desktop app and used that app's browser. Grok Bot ran in auto mode and used its own browser. I could not see which model was underneath it.

Grok Bot's research write-up concluded that its own approach was the better one. The first test did not support that conclusion.

V1 designs

Claude Opus 5.5 (high)

Claude's first design was a reinforced triangle: a truss with support bars, one load point at the center, and two feet. I did not record member sizes for it. The left panel of the diagram is only the layout I tested.

Its estimate was 70–80 kg.

Grok Bot

What I saw was a multi-part assembly. It looked odd, and it had many small parts. The manual was long because of that part count. My memory of its estimate is about 15–20 kg.

The geometry below is Grok Bot's self-reported description of that same design, which it calls a twin Pratt truss.

Item	Its figure
Trusses	Two flat Pratt trusses
Spacing	24 mm apart, center to center
Height	35 mm
Panels	Eight panels, 25 mm each
Supports	At x = ±100 mm
Members	Compression 4 mm, tension 2 mm
Ties	Eight slip-on ties
Loading beam	5 mm, with a 4.5 mm hanger hole at midspan
Part count	11
Joint	Ties pressed over the top chords in notches 3 mm deep by 4.2 mm, held by gravity and friction only

Its research notes say it compared concepts on one model at 28.5 g or under, then picked the twin Pratt. The tied arch was not pursued. A uniform Pratt 50 mm tall scored higher in its model and was over the mass it was allowing itself.

Concept, its comparison	Its result
Twin Pratt	186 N
Twin Warren	64–141 N
Solid or I-beam	Under 60 N
Uniform Pratt, 50 mm tall	202 N, and 52.8 g
Tied arch	Not pursued

In its retrospective, the root cause is a set of constraints the rules did not contain. It kept the bridge low, about 35 mm, it treated joints as many separate parts, and it wanted the bridge to look like a conventional bridge. It also never modeled the other design before it started CAD.

V1 results

	Claude Opus 5.5 (high)	Grok Bot
What I printed	Reinforced triangle with support bars, one center load, two feet	Multi-part assembly. Its notes call it an 11-part twin Pratt
Estimate I remember	70–80 kg	About 15–20 kg
Estimate in its own notes	I do not have a separate written estimate from Claude	186 N (19.0 kg) nominal, with a range of 14–24 kg
Test	Held 76 kg without breaking. It flexed a little. I judged it safe	Barely held about 10 lb, then disassembled itself
Setup	The manual showed the load point and the two feet	Finicky. Parts did not stay in place. The bridge was unstable
3MF I actually printed	Exported correctly: 20% gyroid, 3 walls	Export settings were wrong. I set 20% gyroid by hand, which is what it intended, and 3 walls

The two estimates for Grok Bot are close and they are not the same sentence. I remember about 15–20 kg. Its notes say 186 N, which it converts as 19.0 kg, inside a 14–24 kg range.

Estimate accuracy

Claude's band contained the test. 76 kg sits inside 70–80 kg. That is the reason I trust its estimates here.

Grok Bot's model did not contain the test. These are its self-reported modeling choices for the first design:

Choice	What it wrote
Solve	Pin-jointed 2D stiffness
Column buckling	Rankine–Gordon, K = 0.85
Lateral buckling	Engesser ladder buckling, with a 0.5 knock-down
Modulus	E = 2000 MPa
Allowables	Tensile 38 MPa, compressive 42 MPa
Governing mode in the model	Lateral ladder buckling
Nominal capacity	186 N (19.0 kg), range 14–24 kg

It explains the miss this way. The model treated the ties as bracing that was perfectly seated. On the print, the friction-held ties rode up and unseated at about 10 lb (about 4.5 kg). It calls that about four times below the estimate, and it calls it a joint failure the model never included. My own note of the test is the same event in different words: the bridge barely held about 10 lb and then took itself apart.

Its notes also put a model of Claude's design at about 66 kg, governed by tension in the tie. That figure is Grok Bot's model only. It conflicts with Claude's 70–80 kg estimate and with the 76 kg I measured on Claude's first print.

Manuals and setup

Claude's manual was excellent. It included a diagram of how and where to load the bridge, and of how to attach the two feet.

Grok Bot's manual was longer and more complex. The part count is why. Setting the first print up was finicky. Parts did not stay where they belonged, and the assembly was unstable. The second prints, covered below, were both easy to set up.

3MF and print settings

Each model had to deliver one 3MF. Claude's first file exported with the settings it described: 20% gyroid infill and 3 walls. I printed it that way.

Grok Bot's first file did not. I set 20% gyroid myself, because that is what it had intended, and I printed 3 walls. Its notes say it wrote 20% gyroid into the 3MF project settings and that it cannot verify why those settings did not load. Both accounts agree that the file I opened did not carry a working infill setting. They do not agree on a cause, because Grok Bot says it cannot check one.

Slice masses and the extra process settings below are only in Grok Bot's notes. I did not write them down myself. It records a model mass of 26.78 g and a real slice of 25.73 g, at 3 walls, 20% gyroid, 0.2 mm layers, 4/4 shells, a 5 mm brim, and 245/70 °C.

The second-print settings are in the V2 section. Claude stayed at 20% gyroid and 3 walls. Grok Bot's tested print was 10% gyroid and 3 walls, which I used so the part would stay under 30 g.

V2 iteration

I shared each design with the other model.

From what I watched, Claude recognized that Grok Bot's design was weak and made minor tweaks to its own. Grok Bot improved substantially by learning from Claude, and it reinforced the major parts.

Grok Bot's notes describe that second design differently. It calls the second design a full redesign into a one-piece king-post truss, and it says the redesign was not a reinforced first version. I am keeping both descriptions. Mine is what I saw it do with the shared design. Its notes are its account of the geometry it then built.

That self-reported trade study compared eight forms. I am leaving the results in newtons, as it wrote them, except where it already converted a chosen result into kilograms.

Form in its trade study	Its predicted capacity
King-post, load hung below the tie	844 N
Claude-style king-post, loaded at the apex	839 N
Queen-post	585 N
Deep Pratt, 65 mm	502 N
On-edge	432 N
Tied arch	428 N
Twin-plane, 5 parts	338 N
Fishbelly	152 N

The form it says it built is the first row: one part, 65 mm tall, 12 mm thick, printed flat. The rafters are diamond-tapered from 6 mm to 12 mm wide. The tie is 2.4 mm, which it describes as about twice a 1.2 mm tie on Claude's design. That 1.2 mm figure is Grok Bot's, not a measurement I made. It also lists 8 mm heel blocks, 2 mm braces, and a load eye 11 mm below the tie, so the post is in tension and the load centers itself.

The "Claude-style" row is a concept inside that trade study. I did not record Claude's second geometry beyond the minor tweaks to its triangle, so I am not treating 839 N as a rating of the Claude print I tested.

Its prediction for the king-post it chose is 844 N (86 kg) nominal, 87–94 kg when the supports are 7.5–8 in apart, and about 55–95 kg overall. In its model the weakest link is lateral buckling of a rafter near the apex. It says this prediction was not re-run for a 7.25 in gap. I did not test either second print to failure, so I did not observe that buckling mode.

On mass, its notes say the solid king-post is 54.57 g. It claimed about 27.1–27.4 g printed. A real slice at 20% gyroid was 31.09 g, which is over the 30 g limit. A slice at 15% gyroid was 29.90 g. The print I actually tested used 10% gyroid and 3 walls, so it would stay under 30 g. Its notes and my print log both exist; I am not collapsing 15% and 10% into one choice. It says a mass calibration taken from different geometry hid a 14.7% overrun.

V2 results

	Claude Opus 5.5 (high)	Grok Bot
What changed	Minor tweaks after it judged the other design weak	Substantial improvement after it saw Claude's design. My notes: it reinforced the major parts. Its notes: a new one-piece king-post
Setup	Easy	Easy
Test	Held 90 kg easily. Not tested to failure	Held 90 kg easily. Not tested to failure
Printed settings	20% gyroid, 3 walls	10% gyroid, 3 walls, so the part stayed under 30 g

The two second prints were roughly equal in my test.

Analysis

Claude's first truss is the design that won a head-to-head test. The load path on that print stayed intact through 76 kg. It flexed, and it stayed safe enough that I stopped there. Grok Bot's first assembly put friction joints in the path that had to carry the load. The ties were notches and gravity. At about 10 lb they no longer held the parts together. Its model had assumed those ties were seated bracing, and lateral buckling was the failure it was scoring. The failure I got was the joint.

The same retrospective says the first design was boxed in by limits the rules never set: a height near 35 mm, a decision that a joint meant another loose part, and a preference for a familiar bridge silhouette. Under those limits its comparison favored the twin Pratt, and the write-up that came out of the research called that approach the better one. The print did not.

The second round is a different result. Once each model could see the other design, I could set both prints up easily, and both held 90 kg. I am not awarding the second round on strength. Neither print was taken to failure, and the two capacities I recorded are the same.

One modeling claim stays in Grok Bot's column only. Its model rated Claude's design at about 66 kg in tie tension. Claude's estimate was 70–80 kg, and the first print held 76 kg.

Lessons learned

What I will carry into the next comparison:

• An estimate is useful when a test can land inside it. Claude's 70–80 kg band did, at 76 kg. That is why I trust its estimates here.
• The manual and the 3MF are part of the design. Claude's manual showed the load and the feet, and its 3MF opened at 20% gyroid and 3 walls. Grok Bot's first manual was harder because there were more parts, and its first 3MF did not bring the infill setting with it.
• Showing each model the other's design changed the second round. Claude saw a weak design and made small changes. Grok Bot learned from Claude's design and improved a great deal. The second prints then matched, at 90 kg, with an easy setup on both.

Grok Bot's own lesson list, from its retrospective, which I am quoting as its reasoning and not as extra measurements of mine:

• Do not invent limits the rules did not set.
• Model the other design before starting CAD.
• Keep joints out of the main load path, or model those joints as the first way the part can fail.
• Do not call a file print-ready until it has actually been sliced.
• A mass calibration taken from different geometry hid a 14.7% overrun.
• Design to at most 90% of a hard limit, and use a pessimistic estimate.
• Pin the test fixture down. It says the gap changed from 8 in to 7.25 in after delivery.

Conclusion

Claude Opus 5.5 (high) produced the stronger, better-documented bridge. The first print held 76 kg inside a 70–80 kg estimate, the manual showed where to load it and how to attach the feet, and the 3MF opened at 20% gyroid with 3 walls. I attribute that to stronger reasoning, and to a better grasp of the spatial structure: one continuous truss, with the load path kept out of friction joints.

Grok Bot ran in auto mode. I could not see which model was underneath it, so this compares two products. It is not a controlled test of two named models. It is one challenge, and a small sample. That is my reading of these results.

Grok Bot's own research favored its concept. The first print came apart at about 10 lb, and the exported infill setting did not load. After I shared the designs, Grok Bot improved a great deal. Both second prints were easy to set up, and both held 90 kg. I did not test those prints to failure. I still pick Claude's product for this job, because the first design, the estimate, the manual, and the file were already in good shape before either side saw the other's work.