Gemini 3.5 Pro is the most-anticipated model Google hasn't shipped yet — and over the weekend, the leaks finally spilled out.

The problem? They completely contradict each other.

One camp says Gemini 3.5 Pro is mogging everything in sight, generating cleaner front-end code than Anthropic's Fable 5. The other camp says it's quietly stumbling on reasoning, coding, and long-horizon tasks — landing behind both Fable 5 and GPT-5.6.

Same unreleased model. Two opposite verdicts. Zero official confirmation.

Welcome to how frontier models get "benchmarked" in 2026.

On July 2, developer Harshith (@HarshithLucky3) posted what he claimed was an LMSYS Arena result showing an unreleased Google model tagged "Gemini 3.5 Pro" beating Claude Fable 5 High on a coding task — generating an SVG of a "minimal isometric card swiping machine."

He didn't just say it won. He said it wasn't close.

Dev X ran with it. The word "mogging" — slang for utterly outclassing something — started attaching itself to Gemini 3.5 Pro's front-end output. Clean SVGs. Pixel-perfect layouts. Spatial composition that reportedly snaps into place in a single prompt where Fable 5 wants you to wait.

Then came the second wave. An overnight leak claimed internal benchmarks show Gemini 3.5 Pro edging out Fable 5 on several zero-shot reasoning tasks, running a 2 million-token context window in a new "Deep Think" reasoning mode.

If any of that holds, Google has a monster on its hands.

Big "if."

Here's where it gets messy.

A separate, differently-sourced leak says the exact opposite. According to that report, early testing shows Gemini 3.5 Pro struggling on advanced reasoning, coding, and long-term task execution — placing it behind Fable 5 and GPT-5.6, with weak spots in bidirectional processing and code infilling.

So one leak says it's the best front-end model on the planet. Another says it can't keep up with the models already shipping.

Both can't be right. And neither one published raw Arena data, prompt templates, or anything you could reproduce.

Worth knowing: Harshith has posted "Gemini 3.5 Pro spotted in Arena" variations several times since May — and the claims never converge on a consistent result. A single-prompt Arena sighting tells you almost nothing about cost, latency, or how a model behaves across real workloads.

Google and Anthropic have said nothing about any of it.

Let's separate the signal from the hype, because a lot is getting blended together online.

Actually confirmed:

Reported, not confirmed:

Pure rumor:

The leaks are the story. But the leak game is the bigger story.

Stealth-testing frontier models under fake names inside public arenas is now so routine that individual X users track it as a recurring beat — re-posting sightings, screenshotting matchups, and shaping community opinion about a model months before any official benchmark exists.

That means early reputations are getting set by unreproducible screenshots. A model can be "the best coder alive" or "quietly falling behind" before its makers say a single word.

And Gemini 3.5 Pro is walking into that meat grinder at the worst possible moment for Google.

In roughly ten days, the company delayed this flagship, watched Transformer co-author Noam Shazeer leave for OpenAI and Nobel laureate John Jumper leave for Anthropic, and shed around $225 billion in market value in a single session — its worst trading day in over a year, per CNBC.

So every leak now gets read as a referendum: can Google still hold the frontier?

The delay reasons Google points to don't help the vibe — reporting attributes the slip to token-efficiency concerns, coding not yet at flagship standard, and long-horizon reasoning below the bar it set at I/O. In a year where "intelligence per dollar" became a procurement metric, a flagship that burns tokens is a flagship that's late for a reason.

Gemini 3.5 Pro leaked over the weekend — twice — and the two leaks tell opposite stories. One, sparked by developer Harshith on July 2, claims a stealth-tested Gemini 3.5 Pro outclasses Claude Fable 5 on front-end coding, with a follow-on leak claiming internal wins on reasoning and a 2M-token Deep Think mode. A separate, differently-sourced leak claims the model actually lags Fable 5 and GPT-5.6 on reasoning and coding. Neither published reproducible data, Harshith's Arena sightings have never converged since May, and Google and Anthropic have stayed silent — so treat every capability claim as unverified rumor. What is confirmed: Google unveiled the model at I/O on May 19, promised it "next month," missed June, and it's still in limited Vertex preview with a leaked-but-unconfirmed July 17 target. The real takeaway isn't which leak is right. It's that frontier models now get their reputations set by unreproducible arena screenshots before launch — and Google is absorbing that pressure right after losing two legendary researchers and $225B in a single day. Next thing to watch: an official GA date with real benchmark disclosures, and whether any independent eval reproduces either leak.

Stay building. 🤖