Google Launches Gemini 3.5 Pro With Two Million Tokens
At the opening of the World AI Conference in Shanghai, Google DeepMind unveiled Gemini 3.5 Pro, a model with a two-million-token context window and a new Deep Think reasoning mode reserved for the Ultra tier.
- Google DeepMind released Gemini 3.5 Pro on July 17, 2026, the opening day of the World Artificial Intelligence Conference in Shanghai.
- The model ships with a two-million-token context window, enough to take entire books, code bases or research papers in a single prompt.
- The new Deep Think reasoning mode is available only on the Ultra plan, priced at $250 a month, with no free tier.
- The launch followed the scrapping of an earlier model and a retrain after structural problems.
Google DeepMind unveiled Gemini 3.5 Pro on July 17, 2026, the opening day of the World Artificial Intelligence Conference in Shanghai, describing the model as the product of a full architectural redesign. The company highlighted two headline changes: a context window of two million tokens and a dedicated reasoning mode reserved for its most expensive subscription tier.
The announcement lands at a moment when the leading model developers are competing precisely over context length, the part of the system that determines how much text a model can process in a single pass. Two million tokens, according to the company, means a user can fit entire books, large code bases or several scientific papers into one prompt without splitting them into pieces.
What two million tokens actually changes
A token is the basic unit of text a model processes, roughly a word or part of a word. A larger context is tied to a model’s ability to follow longer documents and to retain information from the start of a conversation, which has been one of the main limitations of large language models so far. Google reports notable improvements in coding and in long-horizon reasoning tasks, areas where earlier models often lose the thread across a longer chain of steps.
Alongside this, the company introduced a new mode called Deep Think, described as a way of working aimed at more complex, multi-step reasoning. Unlike the standard mode, Deep Think is not part of the free offering. It is available only to subscribers on the Ultra tier, which costs $250 a month. With this, Google draws a clear line between everyday use and the most demanding tasks, placing its most advanced reasoning behind a paywall.
It is worth noting that these are technical specifications the company states itself. Independent performance checks, which would confirm how much of the promised gains hold up in practice, are yet to be published.
Why did the launch slip?
The road to release was not a straight line. The launch followed the scrapping of an earlier version of the model and a retrain after structural problems, among them in recursive tool calling and in graphics generation. This refers to situations in which the model calls external tools, or itself, in a loop during operation, so errors in that chain can accumulate. Rather than ship a model with such weaknesses, the company chose to rebuild the base version, which pushed back the release date.
The choice of Shanghai and the timing, the opening day of a major international AI conference, also reads as a signal about where the race for dominance in artificial intelligence is being run. But the event itself says nothing about how the model will behave in users’ hands, or whether a price of $250 a month will find enough people for whom the most advanced mode is worth that much.
For now, only what Google has stated is certain: a larger context, a new reasoning mode and a heavier subscription for access to the top of the range. Does a larger context window really change how you use tools like this, or is the limit still in how much you can trust them?
