2 million tokens: what it took Google to fix the Gemini it shelved

The failure that preceded today’s launch wasn’t announced. Google quietly shelved the first version of Gemini 3.5 Pro after internal evaluations revealed critical deficiencies: poor performance on mathematical reasoning, broken SVG generation, and inconsistent image quality. The model it had slated for release earlier this year wasn’t good enough to ship, so it didn’t.

The rebuild and what it unlocked

What launched today is a different model entirely. The rebuilt Gemini 3.5 Pro opens with a 2-million token context window, twice the capacity of Gemini 2.5 Pro’s 1-million limit, which means a legal team can drop an entire contract library, a year of financial filings, and a full email archive into a single API call before asking their first question. A 200,000-word brief that would have required chunking through three separate calls now fits in the context with room to spare.

That window comes with a pricing structure designed for enterprise consumption: $15 per million input tokens, $60 per million output tokens. Generating a 10,000-word analysis from a 500,000-word document corpus costs roughly $37 — real money, but below the hourly rate of the junior analyst it replaces for document review tasks.

The premium reasoning tier, called Deep Think, sits behind a $250-per-month Ultra subscription. That pricing decision draws a line: standard API users get a capable generalist model; the most powerful reasoning version stays walled off from individual developers who won’t pay the equivalent of a software seat fee just to access it.

The competition it needs to outrun

DeepSeek V4-Pro launched in June at $0.87 per million output tokens, roughly 69 times cheaper on that metric, with benchmark scores that rival Gemini 3.5 Pro in several head-to-head evaluations. Fable 5 and GPT-5.6 Sol are running extended-context variants of their own, though both remain in limited preview. Google’s answer on pricing is that 2 million tokens enables workloads no competing architecture can handle without expensive multi-call orchestration — the cost comparison only holds if the task fits in a shorter window.

No independent evaluation of Gemini 3.5 Pro at the 2-million-token scale has yet been published. Long-context models reliably lose retrieval accuracy as document depth increases — a known failure mode in prior Gemini versions. Google spent a year rebuilding the model. Whether that year fixed the problem at scale is the question every enterprise team buying in at $60 per million will answer first.

Tags: artificial intelligence, Google, Gemini, DeepSeek, AI models, Machine Learning

2 million tokens: what it took Google to fix the Gemini it shelved

The rebuild and what it unlocked

The competition it needs to outrun

More Like This

Google’s Gemini Spark keeps working with the lid shut — paid beta lands May 26

Google Pixel 11 Pro: higher price, thinner upgrade, and AI the base model can’t fully run

How Gemma 4 and DeepSeek V4 ended the brute-force era of LLM scaling

Google built the AI cloud that drove up the price of its own phone

OpenAI’s new voice model thinks inside the audio loop, and the silence that used to give AI away disappears

DeepSeek V4 launches as the largest open-source AI — at a fifth of GPT-5’s cost

Discussion