I enjoyed building simple yet useful tools such as a digital Life Calendar or a Blink Reminder with AI, and I have already tried most of the major AI models for this.
Every time I tested multiple AI providers to build the same tool, I found one of them to be more intelligent than the others. However, the prompts were not exactly the same, so I couldn’t compare the results.
This time I gave Gemini, ChatGPT, and Claude the exact same prompt.
All three AI models understood the assignment
The basics were surprisingly similar
I asked for a used car dealership website that includes a homepage, an inventory page, and a page for viewing a single vehicle’s details.
I wanted to test the reasoning capability of the AI models, so I kept the prompt somewhat vague on purpose.
All three models understood the basic structure and returned a working website. None of them failed.
All three also included the usual dealership messages, such as how the vehicles were tested and how the pricing is transparent. Every AI model successfully executed the basic instructions.
I used the most capable models and settings available to free users: Gemini 3.6 Flash with Extended Thinking enabled, ChatGPT with its only available Think mode checked, and Claude Sonnet 5 with Thinking enabled and the effort set to Extra.
Claude technically has a Max effort setting, but it didn’t even finish the task before running out of tokens.
The differences appear when you look beyond the surface
The cars revealed how much thought went into the output
Gemini’s output looked decent at first glance. But I found some major issues as soon as I looked into the details.
An entry named “2020 Jeep Wrangler Unlimited Rubicon” has an image of an empty road. There is no car in it.
Moving to the inventory section, I spotted another entry with no image. Finally, I noticed something that was broken. There is an entry for a “2019 Ford F-150 Lariat 4×4” with an image of a Ferrari supercar.
The images were randomly chosen instead of being matched to the vehicle names. While the images showed actual vehicles from real brands, most of them had little or no connection to the listed models.
Another aspect that lacked reasoning was the pricing. Most of the cars were listed for almost the same price, even though the categories were very different.
Claude understood what I didn’t explicitly say
It went beyond the basic requirements
Claude did a much better job with the images. It found the exact same models as the listings.
And that’s not all. The images were of used cars, and the entries included condition badges such as A+, A-, and B.
The images were also high resolution, making it easy to inspect the details. I never explicitly asked it to find realistic used-car photos with dents and scratches.
While Gemini’s output looked more like a commercial for new cars, Claude’s made the listings look like what they were supposed to be.
Lastly, the overall design of the website looked much better in Claude’s output.
Gemini’s output felt more vibe-coded, whereas Claude’s looked far more mature and polished.
ChatGPT did a relatively better job with the images. It still pulled random images from the internet. But at least some of them matched the type of vehicle they were used for.
It seemed to understand something about what it was building.
The images were also high resolution, and the overall website design looked more polished than what Gemini offered.
Claude needed a second chance
It generated sketches instead of using real images at first
Claude initially took a different approach to the images. The AI says pulling real images from the internet with brand logos would invite copyright concerns. So it generated SVG sketches.
But the purpose of this project was to compare the capabilities of different AI models.
If I had chosen a different type of project, perhaps something focused squarely on the core logic or the website design, the result might have been more competitive with the rest.
For the sake of a fair comparison, I asked Claude to use real images. However, I still didn’t explicitly mention something like “realistic used-car photos.” It just figured out the intent on its own.
Finally, it matched all the vehicle models correctly and outperformed the other two models in both accuracy and design.
Claude is slower, but it thinks more like a designer
Speed isn’t everything when English is the new coding language
Claude takes longer to build something compared to its competitors, and I have noticed this consistently. This is a legitimate disadvantage, especially if you are trying to build something quickly.
Both Gemini and ChatGPT took less than 2 minutes to complete the task. In comparison, Claude took 10 minutes to add the real images.
It took even longer to build the first version, long enough to make me think it was stuck somewhere and wouldn’t progress further.
But I’m always willing to wait a few more minutes if the result works.
In this project, I wanted AI to spend effort figuring out what a used-car dealership should look like.
Claude once again impressed me with its reasoning capability. The other models technically completed the task, but they made a few poor decisions along the way.
Claude also needed a second chance in this case, but that was more due to the tool’s ethical constraints than its technical ability.
Fixing all the issues in the output of Gemini and ChatGPT would not only take more time than using Claude, but also be a headache for anyone unfamiliar with software development.