Gangsta AI

Gemini 3.5 Pro Leaks: 2 Million Tokens, a $250 'Deep Think' Wall, and a Rebuild Nobody Was Supposed to Mention

Model Watch

By Beck · 2026-07-13 · 4 min read

The thing about a leak is that somebody, somewhere, wanted you to see it. Maybe. Or maybe a spreadsheet fell off a truck. Either way, the internet spent this week passing around the same rumor: Google's Gemini 3.5 Pro is targeting a July 17 launch, and the specs on the napkin are, in the great tradition of unreleased software, enormous.

What the napkin says

According to multiple outlets tracking the same set of leaks, Gemini 3.5 Pro is supposed to ship with a 2-million-token context window — roughly thirty novels stuffed into a single prompt, double what most rivals offer. Bolted on top is a new Deep Think reasoning mode aimed at multi-step logic and math, the part that's supposed to make it stop confidently inventing things.

Here's the catch, and there's always a catch: Deep Think is reportedly walled off behind the $250-per-month Ultra plan. So the smartest version of the model exists, allegedly, for people who treat $250 a month like a rounding error. The rest of us get the trailer.

“Every generation the number gets bigger and the good version gets more expensive. That's not a bug. That's the whole show.”

The part they'd rather you skimmed

The juicier thread underneath the spec sheet: reporting says Google scrapped the original foundation the model was built on and pushed the launch out to July 17 for a fresh pretraining run — a near-total architectural rebuild after a stretch of staff turnover and quality complaints. That's the tell. You don't tear the engine out and start over because everything's going great. You do it because the last one was making a noise you couldn't fix.

Give Google credit for the nerve, though. Rebuilding the plane mid-flight and then leaking that you rebuilt it is a move.

The reality check

None of this is official. Google has not confirmed the date, the price, or the 2M number. July 17 is a widely-reported target, not a signed launch post — so file the whole thing under "probably, mostly, we'll see." The API pricing floating around ($12–15 per million input tokens in some versions of the leak) is even softer than the date.

And the timing isn't an accident. OpenAI just shipped GPT-Live's full-duplex voice and xAI put Grok 4.5 on the board at #4 on the independent intelligence rankings. Everyone released in the same nine-day window. A big Gemini leak lands right in the middle of that noise, which is either a coincidence or exactly the kind of thing that is never a coincidence.

So what do you actually do

Wait a week. If it ships July 17 with the 2M window and Deep Think does what the napkin promises, great — the context war escalates again. If it slips, you learned nothing except that leaks are a marketing channel now.

In the meantime, you don't have to take anyone's leaked spec sheet on faith. You can put Gemini, ChatGPT, Claude, and Grok on the same question and watch which one actually thinks and which one just talks fast. That's the only benchmark that survives contact with your real work — compare them side by side on Gangsta AI's best-AI breakdown and let the models argue it out for you.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles