The 2-Million-Token Promise: Inside Gemini 3.5 Pro's Long Wait and What It Means for AI's Context Wars

At Google I/O on May 19, Sundar Pichai made a commitment that drew an audible groan from a room full of developers: Gemini 3.5 Pro, with its headline 2-million-token context window and new “Deep Think” extended reasoning mode, would be available by the end of June. “Give us until next month,” he said.

That next month is now ending, and Gemini 3.5 Pro remains in a limited preview for select Vertex AI enterprise customers. Prediction markets give the June 30 deadline roughly 50 to 55 percent odds — a number that reflects something significant about the state of frontier model development in mid-2026. Building the most capable AI systems and shipping them on schedule have become, in practice, nearly incompatible objectives.

What Gemini 3.5 Pro Actually Is

The model’s defining feature is its context window: 2 million tokens, the largest confirmed for any production frontier model announced to date. To put that in concrete terms — 2 million tokens is approximately 1.5 million words, or the combined text of five to eight full-length novels. In practice, it means developers and enterprises can feed entire large codebases, complete year-long email threads, or comprehensive legal discovery datasets into a single session without truncation or chunking workarounds.

Gemini 3.5 Flash, which launched at Google I/O and went live on May 19, already set a new benchmark: 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas, outperforming the previous Gemini Pro generation in most evaluated categories. The 3.5 Pro is expected to extend those gains substantially, particularly on the tasks where Flash still lags — hard mathematical reasoning, multi-step scientific problem-solving, and long-horizon retrieval across massive document sets.

Deep Think: Google’s Extended Reasoning Play

The second major feature, Deep Think, is Google’s answer to the extended reasoning capabilities that OpenAI introduced with its o-series models and that Anthropic subsequently refined. The approach is conceptually straightforward: instead of generating a response in one forward pass, the model is allowed to spend additional compute time evaluating the problem, checking its own intermediate reasoning, and revising before producing a final output.

The practical gains are significant on the class of problems where extended reasoning matters. For hard math competitions, multi-step logic puzzles, and complex scientific derivations, reasoning models consistently outperform their standard counterparts by substantial margins. Google claims Deep Think will push Gemini 3.5 Pro to frontier performance on ARC-AGI-3 and similar hard reasoning evaluations.

There is, however, a significant caveat. Deep Think is restricted exclusively to the $250-per-month Ultra subscription tier. Standard Gemini Pro subscribers at $20 per month will have access to the 2-million-token context window and the full model capability — but not the extended reasoning mode. This tiered access strategy has drawn criticism from developers who argue it creates an artificial capability ceiling for a feature that is fundamentally computational rather than qualitatively distinct.

The Competitive Pressure Google Is Navigating

Google is not releasing Gemini 3.5 Pro into a vacuum. Since May 19, the frontier model landscape has continued its relentless acceleration.

Anthropic’s Claude Opus 4.8, released May 27, currently holds the top position on the Artificial Analysis Intelligence Index with a score of 61.4 — the first model to break above 60 by a clean margin. OpenAI’s GPT-5.5 Instant became the default ChatGPT model in April and continues to define the practical benchmark for everyday AI use at scale. And GPT-5.6, expected imminently, is rumored to target a 1.5-million-token context window with sharper long-horizon coding — territory that directly overlaps with Gemini 3.5 Pro’s core value proposition.

If GPT-5.6 ships before Gemini 3.5 Pro becomes publicly available, Google risks the awkward situation of having its flagship model’s headline feature — the 2-million-token context window — arrive in a market where the competitive bar has already moved again.

Pricing and the Flash Comparison

Gemini 3.5 Flash is already available in the API at $1.50 per million input tokens and $9 per million output tokens — a price point that makes it competitive with any mid-tier model on the market and dramatically less expensive than GPT-5.5 for high-volume use cases. The 3.5 Pro is expected to follow the historical pricing ratio of roughly 10× Flash, putting it around $15 per million input tokens and $60 per million output tokens, comparable to Claude Opus 4.8 and GPT-5.5 pricing.

The question facing enterprise buyers is whether the 3.5 Pro’s additional capability justifies the cost premium over Flash for their specific workloads. Gemini 3.5 Flash has already been described by several enterprise evaluators as “good enough for 80 percent of use cases at 10 percent of the cost.” A 2-million-token context window matters enormously for specific enterprise tasks — comprehensive code review, long-document synthesis, contract analysis — but is irrelevant for the conversational and task-execution workloads that dominate most actual deployments.

The Broader Pattern: Slipping Schedules

Gemini 3.5 Pro’s delay is not unique. OpenAI’s GPT-5.6 has been “coming in June” for weeks without a firm date. Anthropic ships models with limited advance notice specifically to avoid this dynamic. The pattern reflects a structural reality of 2026-era frontier model development: the final stages of safety evaluation, red-teaming, and infrastructure hardening are inherently unpredictable. A model can be technically complete but spend weeks in pre-launch review.

For Google, the stakes of this delay are reputational as much as competitive. Google has spent years trying to shake the perception that its AI products ship late and underperform relative to announcements. Gemini 1.0’s troubled December 2023 launch, the Gemini Ultra delays in early 2024, and now the 3.5 Pro extension are points on a line that observers are increasingly connecting.

Google’s defenders note that the Flash launch at I/O was genuinely on time and genuinely capable. The underlying quality of the 3.5 architecture is not in serious dispute. What’s in question is whether Google has solved the organizational problem of making its capability improvements legible to the market on predictable schedules — a problem that, in the current environment, is as strategically important as the underlying technology.

What Comes Next

Three scenarios define the next ten days. First: Gemini 3.5 Pro ships before June 30 with the 2-million-token context, Deep Think on Ultra, and API access at the expected pricing. This recovers the Google I/O promise and positions the model strongly against GPT-5.6 in the enterprise market. Second: the release slips to July, accompanied by an updated capability set that extends the 2M context window or adds new modalities to justify the additional time. Third: GPT-5.6 ships first, and Google’s release becomes a reactive catch-up rather than a leading announcement.

For developers building applications on the frontier model stack, the practical advice is the same in all three scenarios: evaluate Flash for your current workloads and build in abstraction layers that let you switch underlying models without rewiring your application. The model that looks best today is rarely the model that looks best in ninety days.

What Gemini 3.5 Pro will eventually deliver — a 2-million-token context window at frontier performance levels — is a genuinely meaningful capability improvement. The question is only whether it arrives soon enough to define the market moment, or gets published into a landscape that has already moved on.