How the latest AI models are good at generating things in SVG format: the case of Capybara SVG
As SVG can get very clear and scale its content very beautifully and has a very small amount of memory usage compared other file formats, it is becoming one of the best options to create visualizations in software engineering. Therefore, the capability of the latest frontier LLM models in creating those is becoming one of the benchmark points in the field. Therefore, I have done my own interesting experiment comparison.
I have given the following the same prompt for all of the latest frontier LLM models with their best efforts, they kind of generated they are their own version of the sample PNG image of Capybara. I had generated this image of Capybara in an image generation model and then attached the following prompt with this image to all of the models below and had quite different outputs as you can see below:
so i want you to create svg version of this same image. if you create it in a very big scale i am fully sure that you will be able to create it in this form:
animated like a 3D object, adding all of the small details making the final output fully the same as this image. remember you have to make the final result fully like this, you can use unlimited token, you can create it much bigger as creating it bigger gives you opportunity to use smaller details easily that looks smooth and better when the image is viewed in smaller size.
this image was attached to prompt
several frontier models has different results while some fully failed to make something at one shot.
Claude Opus 4.8 xhigh:
the result of Opus 4.8 with xhigh effort in Cowork
GPT-5.5 extra high:
result of GPT-5.5 extra high in codex app
Gemini 3.1 Pro High:
Gemini 3.1 Pro High in Antigravity Agent is trash, it could not do anything
Gemini 3.5 Flash High continued the degradation, and created an HTML version of … guinea pig:
and antigravity app does not have AN HTML VIEWER!!!! and it does not have option to open the html file in a browser nor it does offer a browser in it. WHAT A SHAME GHOOOOOGLE!
AND i opened it from finder folder in the chrome:
it is pig, not mouse
and that famous GLM 5.2 with Max effort?? :
GLM 5.2 Max in the chat.z.ai platform
Not sure why but it had the worst result among all.
again GLM-5.2 Max, but this time from Cursor app and it was better, imo:
GLM-5.2 Max from Cursor app
Composer 2.5 Fast:
Composer 2.5 Fast in Cursor app
End finally long awaited model, FABLE, THEY MAJESTY. if you were expecting something great, you were wrong. it is in the league of the Opus in terms of that specific SVG generation:
Fable 5 xhigh effort in Cowork (it is not Potato!)
As a conclusion, I can say that SVG is one of the greatest visualization options that exist in software engineering, and being really good at creating those makes all along the model exceptional. From my experience, generally most of the models are really good these days, but this was just an experimental, different approach of creating something in the SVG format. Thanks for reading this.
Related case study
This article stands alone for now, but the contact invitation below is always open.
Start a project
Need this level of product thinking in your build?
Reach out if you need someone who can shape the product and ship the interface, system, and iteration loop.
Start a project