Gemini 3.5 Pro is still in preview: 2M context and Deep Think

Google announced Gemini 3.5 Pro at I/O on May 19 and promised general availability (GA) for June. As of June 19, the model is still in a limited preview open only to select Vertex enterprise customers. It hasn't reached the general Gemini app, AI Studio, or the consumer subscription.

So there's a strong promise on the table but no product I can hold yet. This post isn't a review; it's an expectation map.

What was announced

Gemini 3.5 Pro targets the work Google's old "Ultra" tier covered: frontier-level multimodal capability, a 2 million-token context window, and an extended reasoning mode called "Deep Think." Deep Think means the model spends more time evaluating a problem before generating an answer — an approach similar to OpenAI's o3-style reasoning mode.

Sundar Pichai introduced the model on the I/O stage but told the room, in effect, to "wait another month" — reportedly drawing an audible groan from a crowd hoping for immediate access. Gemini 3.5 Flash, which shipped in May, is already the default in the Gemini app and AI Mode in Search.

What changed

Versus the previous generation, the real difference comes down to two headings: context and reasoning.

• Context: 2M tokens, aimed at feeding long codebases and large document sets to the model in one go.

• Deep Think: a promise of clear improvement on hard math, multi-step logic, and scientific problem solving. The expectation is it makes a difference on ARC-AGI-2 style tasks.

• Price: Pro is expected to run roughly 10x Flash — about $15 per million input and $60 per million output tokens. Deep Think is said to be restricted to the $250/month Ultra plan.

The positioning is familiar: when I wrote about Build Mode that Google brought to AI Studio at I/O 2026, we saw the same strategy — ecosystem first, most powerful model later.

My first impression

I don't have preview access yet, so I couldn't test Deep Think on my own prompts. What caught my eye while reading wasn't a technical feature but the timing: the model is gearing up for GA in the same week Anthropic had to pull Fable 5. So it's positioning itself in the gap people are calling the "post-Fable-5 frontier."

What interests me most is the 2M context. The wall I hit most often in solo projects is the context limit; being able to hand the model the whole codebase and say "change this" would be a practical win. But actually filling 2M tokens has a scary bill; growing the context doesn't always mean using it.

Practical impact

For TR developers and indie makers, the concrete question is: until GA arrives, is Flash enough? For most daily work, very likely yes. Pro's value will show up not in single chats but in niche jobs that demand long context and heavy reasoning.

My approach is to wait: I don't put a model on my roadmap before it has an API id and clear pricing. I wrote about that reflex the other day in the Sonnet 4.8 that never was — you plan around the shipped model, not the promised one.

Limits / concerns

The biggest uncertainty is the date. Prediction markets put "GA by June 30" at roughly 50–55%; so the lean is "yes" but not guaranteed. Pichai's "another month" message at I/O suggests Google isn't rushing this tier.

The second concern is price: $60 per million output tokens is serious for solo budgets. Locking Deep Think behind the expensive Ultra plan could leave the model's most interesting feature out of reach for most independent developers.

A note from me

What I feel watching Gemini 3.5 Pro is curiosity more than impatience. Everyone in the industry has piled onto the "context" and "reasoning" axes; those two headings have become 2026's real racetrack.

I'll wait for GA day and, as my first move, put Flash and Pro side by side on the same real task. It's not a benchmark table but behavior on my own codebase that convinces me. Until then, this is a name on my watchlist that hasn't entered production yet.