GPT-6 Sol vs Claude Opus 5.5: Which New Model Should You Use?

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: GPT-6 Sol vs Claude Opus 5.5: Which New Model Should You Use? · Channel: Zero to MVP · 09:24 · Watch on YouTube

Topics: AI evaluation and benchmarks, Claude Code, Software engineering with AI, AI models

Summary

The video compares OpenAI’s GPT-6 Soul and Anthropic’s Claude Opus 5.5 in their native coding environments, focusing on three difficult tasks rather than a comprehensive benchmark. Both models successfully created a printable ASCII pyramid with a 3D effect, though the presenter preferred Opus’s result. They also both built visually appealing genetic-algorithm creature simulations, while Opus produced the preferred code review and identified more issues, albeit taking longer. The presenter names Opus the overall winner based on these tests, but plans to use both models longer before drawing final conclusions.

Key points

The presenter tests GPT-6 Soul in Codex and Claude Opus 5.5 in Claude Code.
The video describes both models as mid-tier options for everyday use and reviews their creators’ benchmark claims.
The first task asks each model to create ASCII-style artwork that forms a 3D pyramid when printed and viewed at an angle; Opus wins this round.
Both models successfully build an appealing simulation of creatures evolving through a genetic algorithm, resulting in a tie.
The final task is to analyze the presenter’s open-source Markdown editor and report potential code problems.
Opus takes around eight minutes for the code review, compared with five minutes for Soul, and provides more findings and detail.
The presenter chooses Opus as the overall winner, while emphasizing that the tests provide only an initial impression.

Chapters

Start	Chapter	Summary
00:00	Model releases and benchmark claims	The presenter introduces Soul and Opus, discusses their positioning and benchmark claims, and outlines a direct comparison.
02:06	Testing a printable 3D pyramid	The models create printable ASCII pyramids, and the presenter prefers Opus’s more convincing 3D effect.
04:41	Simulating evolution	Both models build attractive genetic-algorithm creature simulations and receive a tie.
06:45	Code review and final verdict	Opus finds more issues in the Markdown editor’s code and is named the winner, with the presenter noting the comparison is not comprehensive.

Quotes

In this test, I would give the win to Opus. — 04:41

So, both models performed excellently in this test, which means we can say they both won. — 06:14

I’d say that for me, Opus 5.5 won this round. — 08:21

In my opinion, today’s winner is Opus. — 08:53

I just wanted to get an initial impression of the models for myself. — 08:53

Concepts and entities

GPT-6 Sol, Claude Opus 5.5, Codex, Claude Code, OpenAI, Anthropic, algoritmo genético, ilusão 3D com caracteres, GitHub

Full transcript (original language)

00:00 GPT-6 Soul and Opus 5.5, two new models from the leaders of the AI race. They were released this week. So, today I want to figure out what to use end-to-end and also how significant the progress is compared to the previous versions. I have a set of benchmarks that I am going to use to test these models today. I’ll be testing OpenAI’s GPT-6 in Codex and Opus 5.5, of course, in Cloud Code. So, each model will be tested in its native environment. But

00:31 before I move on to the testing, let’s quickly look at the releases themselves and see what the creators of these models are claiming. So, both of these are mid-tier models positioned as everyday models for regular use. OpenAI says that its strongest model is Astra, while Soul is now the mid-tier option. So, the model lineup now looks like Astra, Soul, and Luna. Terra and Eros, on the other hand, have been removed. If you look at the benchmarks provided by OpenAI itself, we can see that while

01:03 Soul falls behind Astra, it stays very close to it. And considering the price difference, in my opinion, that makes it the optimal choice. Now, let’s take a look at Anthropic’s new Claude Opus 5.5 model. For Anthropic, this isn’t their largest model either. Fable remains their flagship. At the same time, however, the company claims that they have managed to achieve Fable-level intelligence at a lower cost. And judging by their benchmarks, that does

01:34 seem to be the case. Opus 5.5 outperforms Fable 5.1 across all of their tests. They also compare their new model with GPT Astra, and it beats it in almost every benchmark, both in term of performance and price. However, Anthropic and OpenAI compare their new models with the previous generation of their competitors. What I want to do now is compare them directly against each other. So, let’s move directly to the testing. For the tests, I’ll use my repository

02:06 with a collection of prompts. They grouped from simple to difficult, and in this case, there’s no point in his using the simple tasks because they’re designed for testing local models. The models we are testing today would handle them without any trouble. So, I’m going straight to the difficult prompts. By the way, this repository is available on GitHub, so you can download it and experiment with it yourself. The link will be in the description. Okay, the first task will be an unusual one. It will test how well the model can

02:37 navigate and understand the physical world around us. In this task, the models need to create an image of a pyramid using text characters, which, when printed on a regular sheet of paper, should create a 3D illusion. In other words, the pyramid should appear three-dimensional when viewed from an angle. All right, let’s start with GPT-6 Soul. I give the model the prompt and wait. I’m not going to show the entire response generation process because

03:08 nothing particularly interesting happens. The model is just thinking. So, let’s jump straight to the result. Soul generated a number of files, and it looks like one of them is exactly what we need. Okay, in principle, everything matches the task. ASCII-style graphics made from text characters, perspective distortion. It looks like the model both understood what I wanted correctly and produced the expected result. But, let me actually print this image and see how convincing

03:38 the 3D illusion is. When printing, I’m setting the scale to 100%, so that what ends up on the paper is exactly what the model created without any additional distortion. So, I’d say there is at least something resembling a 3D effect. At the very least, when I change the viewing angle, the image starts to look like a normal pyramid. Okay, now let’s move on to Opus 5.5. I deleted all the files created by Soul, so that both models are tested under the same

04:09 conditions. I give it the exact same prompt. Once the model is finished, let’s look at the files. Yeah, Opus also generated a PDF file. Let’s open it. In my opinion, the image looks better here. It even has a background. I also set the print scale to 100% and print it on paper. So, let’s look at the result. I’m trying different viewing angles, and yes, in my opinion, the 3D effect looks more convincing here than with Soul.

04:41 But, let’s also compare them side by side. Overall, I think both models completed the task successfully, but I still prefer the version from Opus. So, in this test, I would give the win to Opus. All right, I think everything is clear with the first task. Let’s move on to the second one. Here, I want to give the model a task that combines coding with elements of physical world simulation. They will need to build a simulation of generations of creatures evolving using

05:11 a genetic algorithm. And [clears throat] one of the requirements is that the simulation should also look visually appealing. Just like last time, Deputy 6 Soul gets the task first. Since the task is described in detail in a separate file, I’m not going to copy the whole text into the prompt. I’ll just tell the model that the project is described in the file. Again, I’ll skip the whole generation process and go straight to the result. So, I’m opening the final file in the browser. All right, it looks

05:42 like everything launched and is working. As you can see, the evolution process works. The generation are changing and overall everything behaves as required. And on top of that, it looks very good. Great result. Open AI’s new Soul model handled this task really well. Okay, now let’s move on to Anthropic’s model Opus 5.5. Everything is exactly the same. I launch Claude code and I ask it to implement the project described in the file and wait. A few minutes later,

06:14 the result is ready. By the way, Soul also spent a few minutes on this task. It looks like everything is fine with Opus as well. The simulation launched and is working. The creatures gradually evolve. It looks a bit different from Soul’s version, but it also looks very good. Definitely an excellent result. Opus handled the task really well, too. So, both models performed excellently in this test, which means we can say they both won. Or in other words, this one is a tie. Okay, I think everything is clear

06:45 with the second task as well. So, it’s time to move on to the final third test for today’s model. But before we get to it, I want to remind you that I recently launched a Patreon, where I post a lot of interesting and useful things that don’t make it onto this channel. Definitely check it out if you want more experiments, extra materials, and content that doesn’t make it to YouTube. So, the third task will be to analyze the code of one of my projects and find potential problems in it. The project

07:17 will be my markdown editor, which is free, open source, and is available on GitHub. I clone it into the folder where I run in today’s tests and ask Soul to analyze it. Then I ask it to save the results into a markdown file. After a few minutes of work, or more precisely, after 5 minutes, the result is ready. Let’s take a look. The funny part is that I am reviewing the feedback on the editor’s code inside the editor itself. You could say we ended up with a kind of recursions here, but let’s look at the

07:48 findings. The model found a number of issues and grouped them by severity. Great. From what I can tell at a quick glance, the findings seem valid. I also like that each issue includes the exact location in the code. Excellent. Now, let’s see how Opus 5.5 handles this task. Of course, I deleted the file created by Soul so that Opus couldn’t peek at it. Okay, Anthropic’s model took longer, around 8 minutes, which is almost 40% more than Soul took. But,

08:21 let’s take a look at the result. Okay, at the beginning of the file, there’s a summary table with all the issues it found. Not bad. I can also already see that the number of problematic areas it found is much higher. Great. Each issue also has a slightly more detailed description than what Soul provided. Yes, overall, I like this version more. I’d say that for me, Opus 5.5 won this round. All right, I think after three rounds, we can already sum things up. In

08:53 my opinion, today’s winner is Opus. It handled both the first and the third tasks better. However, Soul is also very good. OpenAI’s model performed slightly worse, in my opinion. But, it didn’t fail any of the tasks. I simply preferred the results from Anthropic’s model. But, it’s important to note that this was not meant to be any kind of comprehensive benchmark. I just wanted to get an initial impression of the models for myself. Over the next few weeks, I am going to keep using both of

09:24 them, and after that, I’ll be able to draw some final conclusions and share them with you. So, don’t forget to subscribe to the channel. And that’s all for today. Also, don’t forget to subscribe to my weekly newsletter about IT and AI. The link in the description. See you soon. Take care.