Opus 5.5 vs GPT-6 Sol (Blender F1 Car Test)

Author: Better Stack · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Opus 5.5 vs GPT-6 Sol (Blender F1 Car Test) · Channel: Better Stack · 06:16 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI content generation, AI models

Summary

The video compares Opus 5.5 and GPT-6 Soul in Blender using identical prompts to recreate a Ferrari F1 car and animate its construction and a pit stop, then compares both with GPT-6 Astra and Fable 5.1. Opus 5.5 produces a cleaner, more detailed result than Soul, though Astra is a close competitor and wins the audience and staff poll for its pit stop animation. The video also compares cost and completion time, finding Soul much cheaper but less polished, while Astra takes longer and costs more than Opus. The presenter names Opus 5.5 as their favorite overall and new daily driver.

Key points

Opus 5.5 ran in Claude Code and Soul in Codex, both on extra high, using identical prompts in the same order.
The three prompts asked each model to recreate a Ferrari 2026 F1 car in Blender, animate its assembly, and create a pit stop animation.
Opus 5.5’s build animation was praised for its clean blueprint style, lighting, and assembly, although the front wing and some details were inaccurate.
GPT-6 Soul’s result was less detailed and polished, with disconnected parts and visible issues in the pit stop animation.
The presenter reports Opus 5.5 cost $27.86 and took 1 hour and 7 minutes; Soul cost $4.50 and took 48 minutes and 44 seconds.
Astra was considered very close to Opus, with more realistic-looking details and the preferred pit stop animation in the poll.
Fable 5.1’s car was described as well done, but its render had artifacting and its livery was less successful.
The presenter chose Opus 5.5 as their favorite overall, while noting Astra was also a strong option.

Chapters

Start	Chapter	Summary
00:00	Test setup and prompts	The presenter introduces the model comparison and explains the shared Blender prompts and setup.
01:03	Opus 5.5 and Soul results	The video reviews the two models’ build and pit stop animations, revealing Opus 5.5 as the stronger result.
03:10	Results, costs, and Astra comparison	The presenter identifies Opus and Soul, compares their costs and times, then evaluates Astra against Opus.
04:44	Fable 5.1 and final verdict	The video reviews Fable, gives the final cost and time figures, and summarizes the presenter’s preference for Opus 5.5.

Quotes

Spoiler alert, one of these new models absolutely dominates the other. — 00:00

It’s obviously not a one-to-one of the real thing, but it’s crazy to be complaining about this. — 01:34

So, in my opinion, Opus 5.5 is the big winner. — 03:10

the winner for both of them was Astra’s animation — 04:44

My personal favorite was Opus 5.5 — 05:45

Concepts and entities

Opus 5.5, GPT-6 Sol, GPT-6 Astra, Fable 5.1, Blender, Claude Code, Codex, Anthropic, Better Stack

Full transcript (original language)

00:00 We had two new models drop last week, Opus 5.5 and GPT-6 Soul. So, I put both of them head-to-head on Blender to see which one is best and if they’re better than Fable and Astra. Spoiler alert, one of these new models absolutely dominates the other. >> [music] >> So, my setup for this was running Opus in Claude Code and Soul in Codex, both of them on extra high with the exact same prompts in the same order. The first prompt was recreate the Ferrari 2026 F2 car in Blender, high detail, and use the web to gather images for reference. I actually made a mistake in

00:32 this prompt by typing F2, and every single model caught that for me. So, I guess I’m the one who hallucinates now. Once these prompts kicked off, I noticed that every single model used the same official Ferrari launch photos, most of them from the Formula 1 website. So, these should give the models a really good chance as it covers all of the angles. Now, at first I just wanted some static images, but once I got the output, I was really impressed. So, I wanted a little bit more from them. So, I sent the next prompt to animate the car being built from all of its parts, and then another animation of a pit stop where the car drives in and out of frame. So, we had three prompts in total

01:03 for each model. After sending them, it was just a waiting game to see the final result. So, there’s still time for you to hit that subscribe button. And now that you’ve done that, we can see the final product for each of the tests, and I’ll tell you which model made each one after you see the results. We’ll start out with model number one here. This build animation is super clean in my opinion. I really like the blueprint style and the glow that it gave to the individual parts. And it has a really clean reflection on the ground. And also, the parts fly in without any real glitchiness or overlap issues. I must say this is one of, if not the best render I’ve ever had a model make. It’s

01:34 definitely not perfect. The front wing is massive and not very aerodynamic. Real F1 wings are pretty big, but a lot more flat with end plates. I really like what it’s done with the tires and the rear wing. I think it’s done a pretty good job there. And it has tried to include the jagged edges on the fin of the car, but they’re actually the wrong way around compared to a real F1 car. So, as you start nitpicking, you will find a few issues. It’s obviously not a one-to-one of the real thing, but it’s crazy to be complaining about this. We’ve gotten so spoiled now. I couldn’t make this today in the same time that the AI model did, and a year ago, I didn’t even think it could make a wheel.

02:04 As for the final prompt, which was the pit stop animation, that is super clean as well. Added some really nice motion blur for the car coming in and out, and it looks like it attempts to lift the car to change the tires, but I will say the way the tires fly off and up is a bit odd. Overall though, I really like this model, especially when we compare it to our next one. I think you can just tell by the way the parts are being assembled here, that this one is just less clean and polished. It’s not bad for an AI model at all, but now that we’ve seen what model one can do, it is just worse. The car itself gives me the vibes of a single seater series that’s not F1, so

02:36 like F3 or below, and a lot of the parts aren’t even connected to the actual car itself, and you can see the IBM logo at the back is not the right way around either. It is just a lot less detailed and feels more like what I expected an AI model to be capable of before I saw what model one could do. If we take a look at the pit stop animation, it did okay given the model that it was using, although these tires fly off in an odd way, and you may have noticed when the car is driving in, the tires that are going to be attached to it are actually following along by its side, and one of them appears in the frame, and the same thing happens when the car drives off as well. It is just not polished at all.

03:10 Let me know if you agree with my assessment down in the comments below, and also pause the video here if you want to make a guess, as I can now reveal that model one was Opus 5.5, and model two was GPT-6 Soul. So, in my opinion, Opus 5.5 is the big winner. But one thing we do have to take into account is cost and time. So, if we take a look at the coding sessions, if done through the API, Opus 5.5 would have cost $27.86, and in total it took 1 hour and 7 minutes to finish everything that we just saw. If we compare that to GPT-6 Soul, that cost $4.50 and took 48

03:41 minutes and 44 seconds. Soul was actually half the price of Opus per token, but Opus came out to be about six times more expensive because it did a lot more work. It wrote about six times more output, and about $11 of its bill was just writing to its cash. So, Soul is a lot cheaper, but I’d much rather the first result. And I have to say, I did all of this on my Claude subscription, and I was really surprised that it barely made a dent in my usage. It seems to me that for Soul’s training, they didn’t really focus on 3D understanding, whereas Opus 5.5 definitely seems to have, but it’s not

04:12 over for OpenAI yet, because 3D is supposed to be one of the strong suits of GPT-6 Astra. So, let’s see how Astra and Fable did on the exact same prompts. Up first, we have Astra, and now I am definitely torn. I’m struggling to pick between Opus 5.5 and Astra. I think there’s something about this Astra one that just feels a little more real. It might be that the front wing is more accurate, same with the wheel covers and rear wing, or that the lighting seems to be a little bit better than Opus. I found the Opus one seemed to wash out the colors of the livery a bit more, but I do think Opus got the size of the car more accurate, but it is definitely very close. If we take a look at the pit stop

04:44 animation, I actually polled everyone that works here at Better Stack, and also you guys as a YouTube post, and the winner for both of them was Astra’s animation, with Opus coming in a very close second, only losing by a few votes. I think Astra just nailed the car being lifted up and the tires going on and off accurately, but I personally prefer the Opus one as an overall video, maybe because of that added motion blur. Moving on to Fable 5.1, this model actually came third in my poll, and I think it did a worse job than Opus with the build animation. I really liked Opus’s blueprint style. The car itself,

05:15 though, is very well done. I think it did a worse job on the livery, and there’s also a lot of weird artifacting on the render itself, and I didn’t seem to get that from any other model. As for the pit stop animation, a lot of people commented that it nailed the animation on this one, and I do really like it. I think my favorite part is the suspension when the car accelerates, and also the way the tires come on and off. If we take a look at the final figures for these two models, GPT-6 Astra cost $32.47 and took 1 hour 49 minutes to complete, and Fable 5.1 cost $42.44

05:45 and took 1 hour and 13 minutes. So, out of the four models, GPT-4 was by far the cheapest and fastest by a pretty large margin, but it was also the worst result and I’d avoid it for 3D work. My personal favorite was Opus 5.5, which was the second cheapest. I also think Astro was very close and it doesn’t even cost much more, either. Although, it did take about 40 minutes longer than Opus. Overall though, while making this video and through my other testing, I think this is a really solid release by Anthropic. Opus 5.5 is my new daily driver and I’ve not found the need to use fade while using it. It is such an

06:16 improvement from Opus 5. GPT-4 isn’t a bad model, either. It’s extremely competent. I think where OpenAI is doing better is on the cheaper models like Luna, which is great for those small tasks and extremely cheap in comparison. Let me know what you think in the comments down below. Right there, subscribe and as always, see you in the next one.