Gemini 3.5 Pro’s 2M Context: What It Fixes, What It Won’t

On July 17, Gemini 3.5 Pro goes generally available, and Google is leading with one number: a 2 million token context window. That is double the 1 million tokens in Gemini 3.5 Flash, and the largest of any production frontier model right now. Google even delayed the launch from June to July, citing quality refinements after enterprise testing, which tells you they know exactly how much scrutiny this release is going to get.

So let’s do the thing the launch posts won’t do. Let’s talk about what 2 million tokens actually buys you, what it absolutely does not fix, and how to use it like someone who has been burned by big context windows before. Because I have, and you probably will be too if you treat this as “paste everything, ask anything.”

What 2 million tokens actually means

Tokens are chunks of text, roughly three quarters of a word each in English. So 2 million tokens is around 1.5 million words. In practical terms, that is a full-size codebase with room to spare, or five to ten novels, or months of chat logs, or every contract your small business has signed in the last three years. In one prompt. No chunking, no retrieval pipeline, no “please summarize part 4 of 12.”

That is genuinely new territory. The workflows this unlocks are real:

• Whole-repo code questions. Drop in an entire mid-size codebase and ask “where does authentication actually happen, and what would break if I changed it?” The model can see the call sites, not just the file you guessed was relevant.
• Cross-document analysis. Feed it forty vendor contracts and ask which ones have auto-renewal clauses with less than 30 days notice. This used to require a retrieval system someone had to build and maintain.
• Long-running project memory. Paste months of meeting notes and ask what commitments were made and never followed up on.

Pair that with the new Deep Think reasoning layer, which lets the model work through multi-step problems before answering, and the pitch writes itself: give it everything, let it think, get the answer. The front-end and SVG generation improvements are a nice bonus if you build interfaces, but the context window is the headline for a reason.

What a giant context window does not fix

Here is the part Google’s launch page will not emphasize. Bigger context does not mean uniformly good attention across that context. Research going back years has documented the “lost in the middle” problem: models are noticeably better at using information near the beginning and end of a long prompt than information buried in the middle. Every long-context model I have tested shows some version of this, and there is no reason to believe 3.5 Pro repeals it entirely. When you paste 1.8 million tokens, you are not getting 1.8 million tokens of careful reading. You are getting something more like a very fast skim with strong recall at the edges and softer recall in the murky middle.

It gets worse when the needle is subtle. Finding an exact phrase in a huge context is a party trick modern models do well. Noticing that clause 14 in document 23 quietly contradicts clause 3 in document 7 is a different kind of task, and quality on that degrades as the haystack grows. If the answer matters, you still need to verify it, and the techniques in catching AI when it’s confidently wrong apply double at this scale, because a confident answer drawn from 2 million tokens feels more authoritative while being just as capable of missing something.

And then there is the boring stuff: money and time. Long prompts cost more, full stop, and a maxed-out context prompt is not cheap to run repeatedly. Latency climbs too. A question against a stuffed context window can take noticeably longer than the same question against a tight one, and if you are iterating, those seconds compound into a genuinely worse working session. This is the same tier logic I walked through in choosing the right model tier: the biggest option is rarely the right default, it is the right tool for specific jobs.

How to actually use this thing well

Curate, don’t dump. The fact that you can paste everything does not mean you should. Every irrelevant file, duplicate document, and stale meeting note you include is noise the model has to attend past. My rule: include everything that is plausibly relevant, exclude everything you know is not. Dumping your whole repo when the question is about one subsystem is lazy, and you will pay for that laziness in both dollars and answer quality.

Put the important stuff at the edges. Since attention is strongest at the start and end, use that. State your task and constraints up front, dump the bulk material in the middle, then restate the actual question at the very end. Yes, repeat yourself. “Given all the above, list every contract with an auto-renewal clause under 30 days” as the final line meaningfully improves results over burying the ask at the top of a monster prompt.

Make the model show its work against the source. Ask for quotes, file names, and section references with every claim. This forces grounding and makes verification fast. An answer about your codebase that cites specific functions is checkable in minutes. An answer that just asserts things is a liability, which is the same reason the industry is moving toward AI that shows its work by default.

Know when small beats big. If you already know which three files or two documents matter, a focused prompt to Flash will be faster, cheaper, and often more accurate than lobbing the whole archive at Pro. Google’s own lineup makes this split explicit, and it mirrors how Gemini’s Flash and Pro image models divide labor: the light model for quick, well-scoped work, the heavy one for jobs that genuinely need the extra machinery. Use Pro’s 2 million tokens for discovery, when you do not know where the answer lives. Once you have found where it lives, narrow the context and switch to iteration mode.

The 2 million token window is a real capability, not vaporware, and for cross-document and whole-codebase work it will change what you attempt. Just remember that context is raw material, not comprehension. The pros will win with 200,000 well-chosen tokens while everyone else dumps 2 million and wonders why the answers feel mushy.