Google unveils Gemini 1.5 Pro with a massive 1M token context window

Just a week after opening up Gemini 1.0 Ultra to developers, Google dropped an unexpected bomb on the AI ecosystem: Gemini 1.5.

The headline act isn't just raw intelligence gains, but an absurd leap in context length. Gemini 1.5 Pro can handle up to 1 million tokens in production, with internal tests successfully pushing past 10 million.

What 1 million tokens actually looks like

To put 1 million tokens in perspective, you can feed the model 1 hour of video, 11 hours of audio, over 30,000 lines of code, or roughly 700,000 words in a single prompt.

Under the hood, Google shifted to a Mixture-of-Experts architecture. Instead of firing up the entire neural network for every query, it selectively activates only the relevant specialized pathways, dramatically cutting down compute costs.

Thanks to this efficiency, the mid-sized 1.5 Pro outperforms the older 1.0 Pro on 87% of benchmarks, matching the flagship 1.0 Ultra while using significantly fewer resources.

Long context is useless if the model hallucinates or forgets the middle. In Needle-In-A-Haystack tests, 1.5 Pro retrieved hidden target facts across 1 million tokens with a 99% success rate.

In multimodal tests, researchers fed it a 44-minute silent Buster Keaton movie, and it effortlessly pinpointed obscure visual gags and plot turns. Handed the entire 402-page Apollo 11 moon mission transcript, it reasoned across scattered conversations without breaking a sweat.

Learning whole languages on the fly

Perhaps the most jaw-dropping demonstration is its in-context learning.

Google tested the model on Kalamang, an endangered language spoken by fewer than 200 people globally. With virtually zero digital footprint on the open web, the model could not have memorized it during pre-training.

Researchers simply dumped an entire Kalamang grammar book, dictionary, and a few sentences directly into the prompt. Without any fine-tuning, 1.5 Pro learned to translate between English and Kalamang at the level of a human studying the exact same textbook.

The experimental 1-million-token window is currently rolling out to select developers and enterprise customers via AI Studio and Vertex AI. When generally available, 1.5 Pro will launch with a standard 128,000 token limit, with the 1 million tier priced accordingly.

Why it matters

Gemini 1.5 Pro feels like a tactical cruise missile aimed directly at the entire RAG industry.

For the past eighteen months, developers have bent over backwards building vector databases, chunking text, and optimizing embedding pipelines just to get LLMs to understand private documentation. If an off-the-shelf model can reliably parse an entire codebase or decades of financial statements in one shot, the justification for complex retrieval pipelines shrinks overnight.

For users and engineers, it changes how software is audited, media is edited, and research is conducted. Dropping a full-length video into a prompt and asking for exact timestamped breakdowns without third-party transcription plugins is a massive leap.

That said, brute-forcing context comes with real friction. Processing a million tokens means serious latency, and testers will definitely feel the delay on complex queries. Inference pricing will also be steep once the free preview wraps.

Still, Google just proved that scaling context is not just about memory, but emergent reasoning over massive data silos. Over to you, OpenAI.