【Guide】After a full two months of anticipation, Gemini 3.5 Pro is finally coming! Front-end code generated in one go, rumored to officially debut on July 17th.
What might overturn Fable 5 could be Google's trump card!
Today, the internet was flooded with leaks about Gemini 3.5 Pro. Rumored to be officially released on July 17th.
What truly makes people sit up is its performance in front-end generation: one-shot generation, pixel-perfect accuracy, zero errors.
Most crucially, leaked performance data shows Gemini 3.5 Pro has surpassed Fable 5.
Gemini 3.5 Pro's Front-end Prowess Surpasses Fable 5
After a full two months of holding back, Google's ace card, which they've been reluctant to release, is finally about to be born.
At the I/O conference, they unveiled the Gemini 3.5 series in one go:
But on the day itself, only the lightweight Gemini 3.5 Flash went live; the "premium" Pro was kept in the background.
Sundar Pichai promised on stage, "Give us another month."
June passed, and Pro didn't arrive; now Google has changed their timeline to "July," but still hasn't given a specific calendar date.
Thus, during this information vacuum of official silence, leaks became the sole source of information.
And the Gemini 3.5 Pro pieced together from these fragments presents a rather distinct and somewhat "specialized" picture.
Let's start with the conclusion, which is the strongest and most consistent signal in this round of leaks:
Gemini 3.5 Pro has shown a clear capability leap in "front-end and visual code generation."
Based on feedback from multiple developers on X, the leaked version has visibly improved in these areas—
- Better Design Taste: Generated interfaces are no longer "programmer aesthetic"; color schemes, whitespace, and hierarchy are closer to a professional designer's touch;
- Cleaner UI: One-shot generated front-end pages have clearer structures and less redundancy;
- Significant Enhancement in SVG Generation: This is a repeatedly highlighted area of improvement; complex vector graphics can be drawn correctly in one go with fewer missing pixels;
- Higher "Completion Degree" for Front-end: A single prompt can generate a decent-looking page, not a half-finished product.
During testing, developer QASIM-livelifewithai couldn't help but exclaim:
Gemini 3.5 Pro actually generated an SVG image identical to Gemini lead Logan Kilpatrick himself!
A comparison image between Gemini 3.5 Pro and Fable 5 shows the former slightly superior in artistic conception and atmosphere.
For example, generating a UI with just one sentence.
A hand-drawn SVG mechanical orrery was also achieved by Gemini 3.5 Pro with just one prompt.
Gemini 3.5 Pro can also use Three.js to generate a floating steampunk island.
In developer circles, a slightly jesting term has even started circulating to describe Gemini 3.5 Pro's front-end performance—
"mogging" (completely dominating).
Hardcore Reasoning: Inferior to Fable 5, GPT-5.6
Dominating in front-end doesn't mean dominating in all areas.
In LM Arena's stealth testing, Gemini 3.5 Pro indeed suppressed a host of opponents on these tasks.
However, on the hardest Agent tasks, repository-level software engineering, and long-range tasks, 3.5 Pro still can't beat Fable 5, nor catch up to GPT-5.6.
Fable 5 is an absolute beast on heavy engineering tasks like "repository-level, deep debugging, high-ceiling architectural refactoring," scoring quite well on benchmarks like SWE-Bench Pro;
GPT-5.6 similarly holds a steady lead on the most challenging multi-step reasoning tasks.
And what level is Gemini 3.5 Pro on this line?
Even the leaker themselves stated frankly that even with a new foundation, it "still can't beat Fable 5 or GPT-5.6 on the hardest agent and long-chain tasks."
Earlier last month, a leaked benchmark was even less polite, directly ranking it behind Fable 5 and GPT-5.6.
Its weaknesses are concentrated in reasoning, coding, and long-term task execution.
Old Foundation Swapped, Retrained from Scratch
More intriguing than the front-end performance is why Gemini 3.5 Pro is delayed.
According to the leak, the reason Google is dragging into July is not that they spent the extra time on "parameter tuning and fine-tuning," but rather on doing something much heavier—
Conducting another round of "pre-training," no longer relying on the older foundation.
In other words, Pro is not a simple scaled-up version of Flash, but a rebuild on a new foundation.
Apart from the model itself, the leak also contains an easter egg:
Google is preparing an image model called Nano Banana Pro based on this new Gemini 3.5 Pro foundation, aiming to directly compete with OpenAI's GPT-Image 2.
If true, this means Google wants to strengthen simultaneously on both the "text + image" fronts.
One round of retraining, feeding two product lines: one to tackle front-end code, one to tackle image generation.
Google's plan is quite clever—using the same foundation to open fire on two battlefields at once. Now, the entire internet is waiting for July 17th.
And this is precisely the reality of the 2026 large model arms race. Everyone is desperately accelerating, yet no one has managed to completely overturn their opponent in one go.
Looking back at the past six months, the buzz in the AI circle has had little to do with Google, with OpenAI and Anthropic taking turns stealing the spotlight.
But those familiar with Google know they love to hold back, accumulate a wave, and then release everything in one go.
This time with Gemini 3.5 Pro, it's likely the same strategy: using a flagship model retrained on a new foundation to reclaim lost ground in one fell swoop.
Whether July 17th will actually happen, and if it does, whether it will be as the leaks say, the answer can only come from Google itself.
But one thing is already certain: in this war of the gods, no one dares to stop to catch their breath.
References:
https://x.com/ggg78g89/status/2073783990646149488?s=20
https://x.com/HarshithLucky3/status/2073767413494284754?s=20
https://x.com/ggg78g89/status/2073852747049517231?s=20
This article is from WeChat public account "New Zhiyuan" (新智元), author: ASI Apocalypse; Editor: Peach (桃子)
