2 Million Tokens, $250 a Month: Do You Really Need That?

Since July 17, 2026, Gemini 3.5 Pro has been generally available worldwide. Google DeepMind is promoting the new model with a number that makes even seasoned AI decision-makers sit up: a 2 million token context window – according to several sources, the largest of any production-available frontier model as of July 2026. To use this capability to its full extent, however, you need the Google AI Ultra plan for $249.99 a month, roughly €219 per user per month. The question that immediately arises for every Mittelstand business: do we really need this – or are we buying capacity we'll never fully use in everyday operations? In this article, we break down the numbers and show where the investment genuinely pays off and where a much cheaper solution works just as well.

What's actually new about Gemini 3.5 Pro?

The core of the announcement is quickly summarized. Gemini 3.5 Pro offers a context window of 2 million tokens – roughly twice the size of what Claude Opus 4.8, Claude Sonnet 5, or GPT-5.5 currently deliver at around 1 million tokens. The Codex variant of GPT-5.5 manages only 400,000 tokens. Gemini 3.5 Pro also comes with an extended reasoning mode called "Deep Think," which can work through more complex thinking tasks in multiple steps before producing an answer. Both innovations sound impressive – with one important caveat: Deep Think is reserved exclusively for subscribers to the Google AI Ultra plan. No free usage tier for Gemini 3.5 Pro has been announced so far. Access is only available through paid subscriptions or the API.

Important context: at the time of this research, conflicting information about the Ultra pricing structure was circulating. Several more SEO-oriented sources report an allegedly new tariff structure introduced in May 2026, with Ultra tiers at $100 and $200 instead of a single $250 plan. The more credible sources – Google's own blog and the established tech outlet BGR – consistently cite $249.99 a month for the Ultra plan with Deep Think. There is temporarily a 50 percent discount for new customers in the first three months. For this article we go with the more consistent, better-documented figure of $249.99, but we explicitly note that the exact structure should be checked again on Google's official page before any concrete purchase decision, since pricing models from AI providers tend to change frequently.

What do you actually get for $249.99 a month?

Google AI Ultra is not a pure AI-model subscription but a bundled product. For $249.99 per user per month, alongside access to Gemini 3.5 Pro and Deep Think you also get 30 terabytes of Google One storage, access to the video model Veo 3.1, Project Genie, and the Gemini Agent – the latter available only in the US and in English in some cases. For private users or creative teams who want to use storage, video AI, and a text model together, this package can be genuinely attractive. For a Mittelstand business primarily interested in a capable language model for specific business processes, you're paying for a lot of features that simply go unused in day-to-day work. 30 terabytes of storage, for instance, is hardly a selling point for most businesses that already run their own IT infrastructure or use a different cloud provider.

To put the pure model costs in perspective: official API prices specifically for Gemini 3.5 Pro were not clearly confirmed on Google's official pricing page at the time of research. The predecessor, Gemini 3.1 Pro, cost around $2 per million input tokens and $12 per million output tokens via the API up to 200,000 tokens of context, and $4 and $18 per million tokens respectively beyond that. For comparison, the much cheaper Gemini models from the same family: Gemini 3.5 Flash runs $1.50 for input and $9.00 for output per million tokens, Gemini 3 Flash $0.50 and $3.00, and Gemini 3.1 Flash-Lite just $0.25 and $1.50 per million tokens. The difference between a flat-rate consumer subscription at $250 per head and usage-based API access with one of the Flash models is therefore substantial – and this is precisely where the real cost calculation for businesses begins, which we work through in detail below.

How big is 2 million tokens, really?

To make the number tangible: 2 million tokens correspond roughly to 1.5 million words, or about 3,000 to 4,000 book pages – roughly 10 to 14 average books that Gemini 3.5 Pro could theoretically process simultaneously in a single prompt window. This conversion, however, is based on English-language text. German text tends to tokenize more densely due to compound words and umlauts, so the actual page equivalent for German-language documents is likely somewhat lower than this rule of thumb suggests. Even with a safety discount for the German language, the scale remains impressive.

In practice, this means: a complete contract archive, several years of annual financial statements, or an entire support history could theoretically be analyzed in one pass, without having to manually break documents into chunks – a process experts call "chunking," which becomes necessary with smaller context windows because the model otherwise can't "see" all the relevant information at once. For very specific, rare use cases, that's a genuine advantage, since the manual prep work is eliminated. But the decisive question is: how often does this specific use case actually occur in your business – once a quarter, once a month, or several times a day? That frequency determines whether a permanent premium subscription is worthwhile at all.

A 2-million-token context window is like a truck driver's license for someone who moves furniture once a year. Impressive, expensive to acquire – and completely oversized for the daily commute to work.

A concrete calculation illustrates the ratio even better: a typical customer support inquiry rarely involves more than a few thousand tokens – the email history, the order data, maybe a product manual. Even a generously sized 500-page internal handbook would fill only a fraction of a 2-million-token window. Whoever books a model with such a huge context for tasks like these is essentially paying for a capacity reserve that's never called upon in normal operation – comparable to buying a company car built for occasional long-distance trips just to cover the daily five-kilometer commute to the office.

Who does this pay off for in the Mittelstand – and who should skip it?

For the day-to-day operations of most Mittelstand businesses – email replies, invoice processing, customer service chatbots, simple reports – a context window this huge is overkill. Even a context of 128,000 to 200,000 tokens, or a targeted retrieval approach (also known as RAG, where only the actually relevant document excerpts are loaded into the request instead of everything at once) on a much cheaper model fully covers these cases. The processing power you're paying for with a 2-million-token window simply goes unused in these everyday scenarios – you're paying for capacity nobody calls on.

It looks different if you regularly need to review very large, interconnected sets of documents at a glance – for instance during a due diligence review with hundreds of contract documents, a comprehensive compliance analysis spanning several fiscal years, or a complete customer history that has grown over years and in which connections across many individual cases need to be recognized. In such concrete but rather rare special cases, the large context window can genuinely save time, because no manual pre-sorting of documents is needed and the model can spot cross-references across the entire dataset that are easily missed with a piecemeal approach.

The Deep Think feature, the only genuinely Ultra-exclusive capability besides the context window, is aimed at complex reasoning tasks with multiple intermediate steps. For operational Mittelstand processes like bookkeeping, customer service, or standard evaluations, that's rarely the core need – Deep Think shows its strengths more in research or strategy tasks, which come up daily in only the fewest businesses. Exact usage limits for Deep Think, such as a specific number of requests per day or month, also were not clearly quantified in the available sources – another point you should clarify directly with Google before purchasing, rather than relying on "unlimited use."

Before making a purchase decision, it's worth taking an honest inventory: how many people in the company would actually regularly max out a context window beyond 200,000 tokens? How often per month does the task of analyzing an extensive bundle of documents in one pass, rather than pre-sorting it selectively, actually come up? And can the respective use case not also be solved with a classic retrieval approach on a cheaper model, where only the relevant passages are loaded? In most cases we see in client projects, the answer clearly favors the targeted, cheaper solution – the exception remains the exception, not the rule.

The calculation that rarely gets done

$250, roughly €219, sounds manageable for a single subscription. The problem arises at scale: the price applies per user per month, not per company. With five employees on Ultra access, that quickly adds up to around €1,100 to €1,250 a month – roughly €13,000 to €15,000 a year, just for tool access. And that's without the actual context capability even being used in daily operations, since most tasks only require a fraction of the available context window anyway. With ten licenses, these figures double accordingly to €26,000 to €30,000 a year – a sum that, in many Mittelstand budgets, already equals an entire additional headcount.

The alternative we regularly discuss with Mittelstand businesses at Amazing Agents looks different: a specialized agent, built once, tailored specifically to the actually relevant use cases – such as contract review or support-history analysis – built on cheaper API models like Gemini 3.5 Flash or Flash-Lite. Tokens are billed only on actual demand, not as a flat rate per head per month. The effect: the relevant use cases are often covered at a fraction of the cost, without an entire team having to pay for a flat-rate subscription with unused extra features. This reasoning aligns with our recurring experience, which we also described in more detail in the article ChatGPT Work, Claude Cowork, and Co.: a flat-rate subscription per head is rarely the right lever when only a fraction of the team actually needs the full breadth of features.

There are, however, constellations where the investment pays off anyway: if your company regularly – not just occasionally – needs to analyze very large, interconnected sets of documents that don't break down well into smaller sections, a single Ultra access for the responsible specialist can make economic sense. One example: a legal department continuously reviewing complete contract folders running to several hundred pages, or a finance department analyzing multi-year reporting series in context. The decisive question here is the number of licenses. Instead of handing out five or ten Ultra subscriptions to the entire team as a blanket policy, in most cases a single access for the person or small team that actually needs the feature regularly is enough.

Conclusion: The question isn't the model, it's the use case

Gemini 3.5 Pro is technically impressive, and a context window of 2 million tokens is indeed a new order of magnitude among frontier models – roughly twice the size of Claude or GPT-5.5. But for the vast majority of daily tasks in the Mittelstand – customer communication, invoice processing, reporting, support – this capacity is far more than is needed. Whoever buys $250 per head per month as a blanket policy for an entire team is, in practice, mostly paying for features that lie dormant in day-to-day work: the video generator, the extra cloud storage, the huge context window that's rarely used to capacity, and a reasoning mode that hardly any operational process even calls for.

The more economically sound path is rarely the most expensive available subscription for everyone, but rather a differentiation between "do we need this model for a specific use case" and "does everyone on the team need the complete subscription." For the rare but important special cases involving very large document volumes, a single targeted access can make sense. For the rest of the company, a purpose-built agent that covers exactly the processes that actually cost time, billed on usage rather than flat rate, is as a rule the more economical and better-fitted solution. For more on how to sensibly weigh model choice and cost against each other in the Mittelstand, see also our article Claude Sonnet 5 in the Mittelstand.