Gemini 3.5 Pro Lands: 2M-Token Context, Deep Think at $250

This is the week the AI calendar overflowed. Frontier launches stacked on top of one another, and the same morning a major model is watched for, Shanghai opens the world’s largest AI conference. The quiet question underneath all the noise is not who can hold the most, but who can understand what they hold. What arrives today puts that question in sharp relief.

The Lead: Google’s Gemini 3.5 Pro Reaches General Availability With a 2-Million-Token Context Window

Google’s Gemini 3.5 Pro is expected to reach general availability today, carrying a reported 2-million-token context window and a premium Deep Think reasoning mode gated behind a $250-a-month Ultra subscription.

A 2-million-token context window is roughly 1.5 million words, several full books held in mind at once. In plain terms, you could drop an entire folder of PDFs, a year of meeting notes, or a whole manuscript into one prompt and ask across all of it, without stitching the pieces together yourself. If the spec holds, it is the largest context of any production frontier model.

Hold the specifics loosely for now. As of this week, the July 17 arrival date, the 2-million-token figure, and the pricing were widely reported industry targets rather than a confirmed Google launch post, and official API pricing was still unpublished. The one consistent consumer-facing fact is the gating: Deep Think, the model’s extended reasoning mode, sits behind the $250-a-month Ultra tier.

The editorial point is the shape of the offer, not the benchmark. A model that holds more is being sold as a model that understands more, and those are not the same claim. It lands in the most crowded single week the field has seen, the same day Shanghai’s World AI Conference opens (TechTimes).

What It Means for You

The tools you already pay for are quietly getting more capable this week, without you changing a single setting.

Perplexity opened its Computer agent to every Pro subscriber, not just the top tier. The full suite of twenty-plus models, prebuilt and custom skills, hundreds of connectors, and Slack workflows now reaches the $20-a-month plan, while a new $200-a-month Max tier sits above it. If you already pay for Perplexity Pro, you woke up with an agent you did not have last week.

The same upgrade-by-default pattern shows up inside Office. OpenAI made GPT-5.6 the preferred model inside Microsoft 365 Copilot across Word, Excel, PowerPoint, and Chat, so the AI help millions already use gets the newer engine without anyone flipping a switch.

Anthropic, meanwhile, reworked how Claude remembers you, replacing its old daily-summary memory with individual categorized entries that carry richer context from one conversation to the next. The thread across all three is the same: more memory, more reach, more capability arriving silently in tools you already trust.

“The upgrade you did not ask for is still an upgrade you now depend on.”

What’s Moving Underneath

Underneath the consumer apps, the market is deciding who gets to go public and who builds their own silicon. OpenAI, which filed confidentially for an IPO, is now reportedly weighing a delay into 2027 after a rocky SpaceX debut cooled the market, having earlier eyed a valuation near a trillion dollars.

Anthropic is moving on both fronts at once. It is in talks with Samsung to build a custom AI chip tuned to its Claude models, its own answer to the in-house silicon strategies at Google, Amazon, Meta, and OpenAI, and is reportedly preparing to file for an IPO as early as October, backed by locked-in long-term compute deals.

None of this reaches your desktop this week. But the same morning Gemini is watched for, Shanghai’s World AI Conference opens, a reminder of how globally synchronized and crowded this race has become. The IPOs, the chips, and the conferences are the scaffolding that will decide which tools reach you next year, and on whose terms.

“The storehouse keeps growing; the question of what it is all for does not.”

One Tool Worth Knowing

Gemini 3.5 Pro

Gemini 3.5 Pro’s headline pitch is that 2-million-token context, the ability to hold an entire body of material in one conversation. Before you let that reshape how you work, ask whether your real problem is that your tools forget too fast, or that you have never organized what you feed them. A bigger window rewards a clear question and punishes a vague one, because more room to hold is also more room to get lost.

If you write code, drop a full repository into a single prompt and ask for one specific cross-file change, then check whether the model actually reasoned across the whole codebase or only its edges. If you do not touch code, gather a year of notes or a long PDF into one file and ask a question that can only be answered by reading all of it, then judge honestly whether the answer shows understanding or just fluent retrieval. Either way, resist paying $250 for Deep Think until a real task proves you need it.

Wisdom Speaks

“But where shall wisdom be found? Where is the place of understanding?” Job 28:12, WEB

Job 28 is a poem about human reach. It catalogues how far we will go to pull value out of hidden places, sinking shafts into the dark, bringing every buried thing to light, and then it stops and asks the question above. All that extraction, and still: where is binah, the understanding that knows what a thing is for? A model holding two million tokens has enlarged its storehouse, not its judgment.

The poem answers its own question twice. Understanding “can’t be gotten for gold, neither shall silver be weighed for its price” (Job 28:15, WEB), which is worth remembering the day depth of reasoning is put behind a $250 paywall. Then the resolution: “the fear of the Lord, that is wisdom; and to depart from evil is understanding” (Job 28:28, WEB).

You can hold more this week than you ever could before. The older question is whether you understand any more of it, and where you will go to find the understanding that no context window contains.

Wednesday’s digest: OpenAI Retires the Atlas Browser and Debuts GPT-Live Voice, on how fast the labs discard their own tools. Earlier this week: Apple Sues OpenAI Over Alleged Trade-Secret Theft, on the legal fight over how AI reaches your devices. Today’s 2-million-token arrival extends the same thread: capacity keeps swelling, and the harder question is whether more holding ever becomes more understanding.