Claude Opus 5.5 vs GPT-6.1 Soul and Terra: 3D Model Test Results
Opus 5.5, GPT-6.1 Soul, and Terra were given the same prompt to build a 3D diecast RX-7. One result stood out from the rest.
What happened when Claude Opus 5.5 and GPT-6.1 went head to head on 3D generation?
In a hands-on test, a creator asked three models (GPT-6.1 Terra, GPT-6.1 Soul, and Claude Opus 5.5) to generate the same thing: an interactive, in-browser 3D diecast model of a Mazda RX-7 with a working Wankel rotary engine, built from a set of uploaded reference photos. Terra produced a flat 2D image instead of a 3D model on its first attempt. Soul eventually delivered a legitimate 3D car with a modeled twin-rotor engine, but with visible geometry errors. Opus 5.5 produced the most complete result: a drivable model with popup headlights, a modeled rotary engine showing the combustion stroke, a turntable view, and sound that approximated the real car’s engine note. No developer instruction to “make it 3D” was needed for Opus 5.5. Soul needed to be told twice.
TL;DR
- GPT-6.1 Terra, pitched by OpenAI as a cheaper, faster alternative to Soul for computer-use tasks, failed the 3D test outright, returning a flat 2D image when asked for a diecast model.
- GPT-6.1 Soul needed an explicit follow-up prompt reminding it to build in 3D, and even then produced a car with reversed headlight mechanics and non-turning wheels.
- Claude Opus 5.5, run on medium reasoning effort, generated a recognizable RX-7 with a correctly modeled rotary engine, working turbo effects, drivable physics, and multiple gears, without being told explicitly to make it 3D.
- Longer Opus 5.5 generation times (a 3-minute run versus a roughly 90-second light-thinking run) produced noticeably better results, suggesting reasoning effort matters a lot for complex 3D tasks.
- Both model families still struggle with organic curved geometry, like a sports car’s body lines, compared to simpler boxy shapes, and both made the same “pancake” orientation mistake when placing small objects like popup headlights.
- On pricing, Terra sits awkwardly between GPT-6.1 Luna (cheap and fast) and Soul (expensive and capable), and the creator argued it exists mainly to nudge users toward Soul rather than to be a genuinely useful middle tier.
- The creator concluded that for this specific 3D generation task, Opus 5.5 outperformed Soul by a wide margin, enough that they’d pick Opus 5.5 over Soul for similar creative or visual tasks even at a higher cost.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
How did the two AI families actually build the 3D models?
Each model was given the same prompt: build a diecast RX-7 with a working Wankel engine, using uploaded photos of the real car as reference. The process ran entirely in-browser, with the models generating 3D assets, textures, and basic physics for driving simulation on the fly.
GPT-6.1 Terra returned a 2D image on its first pass, despite the prompt explicitly asking for a diecast 3D model and including reference photos meant for a 3D build. The creator had to prompt again with a blunt “3D, please” to get Terra moving in the right direction, and the fixed version still landed closer to a basic demo than a finished product.
GPT-6.1 Soul was tested across two reasoning settings. On light thinking, it finished in about a minute and a half but skipped detail: the engine shape was roughly right but oriented incorrectly, poking through the hood, and the car didn’t resemble an RX-7 specifically. A longer, roughly three-minute run on Soul produced a clear jump in quality: better body curves, a modeled twin-rotor engine with a visible compression stroke, working popup headlights (though reversed in function), and an engine sound closer to a real rotary. Driving physics remained rough, with wheels that didn’t visually turn.
Claude Opus 5.5, run on medium, built a version that nailed more of the small details without extra prompting: a slender, aerodynamic body shape, a correctly functioning twin-rotor engine with visible spark and combustion effects, a turbo with its own visual cues, a turntable display mode, multiple drivable gears, and sound effects that more closely matched a real rotary engine’s tone. It also let the user switch between automatic and manual transmission and multiple color options, details the creator noted Opus 5.5 added without being asked.
Why did GPT-6.1 Terra fail to generate a 3D model at all?
Terra’s positioning, according to OpenAI’s own framing referenced in the test, is as a faster, cheaper variant of Soul aimed at computer-use tasks, not necessarily complex creative or visual generation. The creator’s working theory is that Terra is tuned to take the “easy way out” on ambiguous tasks, defaulting to simpler 2D output rather than the more resource-intensive 3D build the prompt called for, even when the prompt and attached reference images made the intent clear.
This matters for anyone evaluating model tiers for production use. A cheaper, faster model is only useful if it reliably interprets the task correctly. In this test, Terra required manual correction to do what was asked, which erodes the time and cost savings it’s supposed to offer.
Is Claude Opus 5.5 better than GPT-6.1 Soul for 3D generation?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Based on this single test, yes, by a clear margin. Opus 5.5 produced a more recognizable RX-7 silhouette, a better-modeled rotary engine, more convincing driving physics, and a more complete feature set (turntable view, transmission options, color choices) without needing the 3D instruction spelled out. Soul’s best result, after extra prompting and more reasoning time, still had functional issues like non-turning wheels and reversed headlight mechanics.
That said, this is one test with one prompt and one car. The creator was explicit about treating it as a single data point rather than a definitive verdict, and encouraged readers to try similar prompts themselves. Both models shared a specific flaw: when placing small 3D objects like popup headlights or an engine, each model sometimes oriented the object incorrectly, as if it had been modeled flat and then rotated the wrong way. That’s a geometry-handling issue that showed up across both companies’ models, not just one.
What does this mean for picking between AI model tiers?
The broader takeaway is less about any single car model and more about how reasoning effort and model tier affect output quality on open-ended generation tasks. Within Soul itself, light reasoning effort produced a noticeably worse result than a longer run, suggesting that for tasks like 3D generation, spending more compute time matters more than it might for text-based tasks.
On pricing structure, Terra was positioned as roughly half the cost of Soul with faster response times, while Luna sits as the cheapest, fastest tier for simple tasks like routine computer-use automation. The creator argued that Luna is cheap enough to be worth using for low-stakes tasks, while Soul justifies its higher cost for genuinely complex work. Terra, in their view, occupies an awkward middle that mostly serves to push users toward Soul rather than standing on its own merits. Whether that holds for other task types (coding, agentic workflows, document analysis) wasn’t tested here and would need separate evaluation.
Frequently Asked Questions
What is GPT-6.1 Terra?
Terra is OpenAI’s lower-cost, faster variant of GPT-6.1 Soul, aimed at computer-use and agentic tasks at a lower price than Soul. It’s positioned as less capable than Soul but more capable than Luna, OpenAI’s fastest and cheapest tier.
Did Claude Opus 5.5 need special prompting to build in 3D?
No. Unlike GPT-6.1 Soul, which required an explicit follow-up instruction to produce a 3D model instead of a 2D image, Opus 5.5 generated a 3D interactive model on the first attempt from the same prompt and reference photos.
Which model produced the most accurate rotary engine model?
Claude Opus 5.5 produced the most convincing rotary (Wankel) engine simulation, including visible spark and combustion effects and an engine sound closer to a real rotary’s tone. GPT-6.1 Soul’s longer-reasoning run also modeled a twin-rotor engine but with less convincing detail.
Is this test representative of overall model quality?
It reflects a single prompt and task (3D diecast car generation with reference images), not a comprehensive benchmark. The creator noted it as one test and suggested readers try similar prompts themselves before drawing broad conclusions about which model is “better” overall.
Why did both models struggle with curved car bodies?
Organic, curved geometry (like a sports car’s body lines) is harder for current AI 3D generation than simpler, boxier shapes. Both Soul and Opus 5.5 showed modeling imperfections on the RX-7’s curves, though Opus 5.5’s result was noticeably closer to the real car’s shape.
