Hi everyone,
It feels like Gemini 3 is getting closer every day.
Yesterday afternoon I was browsing X and saw this:
It said Gemini 3 could already be accessed via Canvas on mobile.
But no one knows whether it is Gemini 3 Pro or Gemini 3 Flash.
I tried it on Canvas. I cannot guarantee it is Gemini 3, but the output is seriously impressive.
It is clearly stronger than the current Gemini 2.5 Pro.
Here are a few cases I tested.
SVG
My view on SVG tests:
It does not fully represent model capability, but it is a clear visual way to show model quality.
Here is a SVG example:
SVG Animation of nintendo switch playing pokemon
Honestly, this is the most impressive case I have seen.
Compare with Claude 4.5 Sonnet:
And GPT-5.1 Thinking:
The difference is huge.
Another frequent test is an Xbox controller SVG. After testing many models, only the unreleased Gemini 3 does it best.
Draw a svg of xbox controller
Compare with Claude 4.5 Sonnet:
I suspect Gemini 3 is in a league of its own for SVGs.
Beyond SVG, overall front end aesthetics also improved a lot.
Web design
A simple case, ask it to introduce itself:
Write a html page to introduce yourself including model name, knowledge cutoff date
The impressive part is that I did not specify colors or style, but Gemini still produced a strong result.
Using the same prompt on Claude shows the gap.
Another web design example:
Generate a website for 5090 gpu
The data is correct, and the color palette matches Nvidia even though I did not specify it.
A MacOS recreation example:
The terminal command effect is especially impressive.
A high end design style example:
Try a new version with not gradients and super super high end and beautiful, unique and modern design
The same prompt on Claude Sonnet 4.5:
Claude looks good too, but the interaction animation gap is still noticeable.
Creative writing
When I tested Web Arena, some people asked about Gemini 3 writing quality. I tested it in Canvas too.
A simple case:
Write a passage that is impressive enough to wow me
It is an open prompt.
Now compare with Claude, which is also strong in writing.
I personally prefer Claude here. The topic is easier to follow, while the Google one feels more abstract.
Another case: literary translation.
Translate this line in the writing style of Eileen Chang: Love is like a war: easy to begin but very hard to stop. Only one version, no extra explanation.
Note: I used Canvas but it did not explicitly show the model, so I cannot guarantee this is the stronger model.
Now Claude:
I prefer Gemini for this translation. I asked GPT-5.1 to judge and it also picked Gemini.
Gemini 2.5 was already decent at writing, so the improvement might not be immediately obvious.
Final note
Please release Gemini 3 soon, Google.