跪拜 Guibai
← Back to the summary

Leaked Gemini 3.5 Pro Shows One-Shot Frontend Generation That Rivals Fable 5

After two months of holding back, Gemini 3.5 Pro is finally coming! One-shot frontend code generation, rumored to debut on July 17.

What could overturn Fable 5 might just be this card Google hasn't fully laid on the table yet.

Over the past couple of days, leaks about Gemini 3.5 Pro have been flooding developer circles. Rumors suggest this long-delayed 'super-sized' model from Google might officially debut on July 17.

What's truly making people restless isn't the release date itself, but its performance in frontend generation: one-shot completion, high visual fidelity, and some are already comparing it head-to-head with Fable 5.

More critically, leaked information shows that Gemini 3.5 Pro has started to outperform Fable 5 on certain visual and frontend tasks.

Gemini 3.5 Pro's Frontend Prowess

Overtaking Fable 5

Google has indeed held back for quite a while this time.

At I/O, Google unveiled the Gemini 3.5 series in one breath. But the only model that actually went live that day was the lightweight Gemini 3.5 Flash; the much-anticipated Pro version was never released.

Pichai's line on stage at the time was: 'Give us another month.'

June came and went, and Pro still hadn't appeared. Now Google has shifted the window to July, but officially, no specific date has been given.

Precisely because of the official silence, leaks have become the main entry point for the outside world to observe Gemini 3.5 Pro.

From the fragments that have surfaced so far, the portrait of Gemini 3.5 Pro is very distinct: it's not an all-around crushing model, but one that shows a clear leap in 'frontend and visual code generation'.

Synthesizing feedback from multiple developers on X, the leaked version shows very obvious progress in several dimensions:

During testing, developer QASIM-livelifewithai even exclaimed: Gemini 3.5 Pro generated an SVG image highly resembling Logan Kilpatrick, the head of Gemini.

Another comparison image between Gemini 3.5 Pro and Fable 5 has also been widely discussed. Looking purely at the artistic conception and atmosphere, the former is indeed more eye-catching.

For another example, with just a single prompt, it can generate a complete UI.

A hand-drawn SVG mechanical orrery, also completed with a single sentence.

Gemini 3.5 Pro can also generate a floating steampunk island using Three.js.

In developer circles, a slightly playful term has even started trending to describe Gemini 3.5 Pro's frontend performance:

'mogging'.

Simply put, it means dominating in visual and frontend generation.

Hardcore Reasoning

Still Lags Behind Fable 5 and GPT-5.6

However, dominating the frontend doesn't mean dominating everything.

In LM Arena's blind tests, Gemini 3.5 Pro indeed outperformed many competitors on a batch of visual, web, and frontend-related tasks.

But when it comes to harder Agent tasks, repository-level software engineering, and long-horizon task execution, it's still not the strongest.

Fable 5 remains very capable in 'repository-level debugging, complex engineering modifications, and high-ceiling architectural tasks,' especially showing clear advantages in benchmarks like SWE-Bench Pro.

GPT-5.6 also firmly holds its place in the top tier for multi-step reasoning and complex task chains.

So where does Gemini 3.5 Pro stand on this front?

The leaker's own statement is quite direct: even with a new base, it still can't beat Fable 5 or GPT-5.6 on the hardest agent tasks and long-chain tasks.

An earlier leaked benchmark also gave a similar conclusion: Gemini 3.5 Pro ranks behind Fable 5 and GPT-5.6, with its main weaknesses concentrated in reasoning, coding, and long-term task execution.

So it's more like a specialized but sharp machine: brilliant in frontend, visual, and web generation; but still needing to catch up in heavy engineering, long chains, and complex Agent tasks.

Old Base Replaced, Retrained from Scratch

More noteworthy than the frontend performance is why Gemini 3.5 Pro was delayed.

According to the leaks, the reason Google pushed Pro to July wasn't just a simple round of parameter tuning or fine-tuning, but something much heavier:

A complete redo of the pre-training, no longer using the older base.

In other words, Gemini 3.5 Pro is not simply a scaled-up version of Gemini 3.5 Flash, but a rebuild on a new foundation.

Besides the model itself, the leaks also included an easter egg:

Google is preparing an image model named Nano Banana Pro based on this new Gemini 3.5 Pro base, aiming to more directly compete with OpenAI's GPT-Image 2.

If true, this means Google isn't just trying to fill a product line gap, but wants to push forward on both 'text + image' fronts simultaneously.

One retraining, feeding two product lines: one targeting frontend code, the other targeting image generation.

Google's calculation is clear: use the same new foundation to fill gaps in two battlefields simultaneously. Now, the entire web is waiting for July 17.

And this is precisely the truest picture of the 2026 large model competition: every company is accelerating, but no one can completely shake off the others in one breath.

Looking back at this half-year, the spotlight on AI's biggest stage has mostly been taken by OpenAI and Anthropic in turns.

But those who know Google are aware it has always been good at holding back big moves. Staying silent for a while, then releasing a concentrated round of key updates, is its familiar playbook.

This time, Gemini 3.5 Pro might be following the same script: using a flagship model with a retrained base to catch up on lost ground.

Whether July 17 really arrives, and whether it truly matches the leaks upon arrival, only Google can reveal.

But one thing is already clear enough:

In this large model race, no one dares to stop and catch their breath.

Using Gemini in China:

Entry link: https://geminiai.asia/list/#/home