one year on
Google announces Gemini 1.5 Pro with 1 million token context window, up to 10 million in research
A new Mixture-of-Experts model can process an hour of video or 700,000 words in one prompt, matching Gemini 1.0 Ultra performance while using less compute.
Google today unveiled Gemini 1.5, its next-generation AI model, anchored by a breakthrough in long-context processing. The first model available is Gemini 1.5 Pro, a mid-size model that matches the performance of Gemini 1.0 Ultra despite being more efficient to train and serve, thanks to a new Mixture-of-Experts architecture.
Gemini 1.5 Pro comes with a standard 128,000-token context window, but starting today a limited set of developers and enterprise customers can test an experimental version with up to 1 million tokens via AI Studio and Vertex AI. Google says it has also successfully tested up to 10 million tokens in research settings. A context window of 1 million tokens equates to roughly one hour of video, 11 hours of audio, codebases over 30,000 lines, or over 700,000 words.
The company showcased the model’s ability to reason across the 402-page transcript of the Apollo 11 mission, analyze a 44-minute silent Buster Keaton film, and work with codebases exceeding 100,000 lines. In the Needle In A Haystack test, Gemini 1.5 Pro found the embedded text 99% of the time across 1-million-token blocks. The model also demonstrates in-context learning: given a grammar manual for Kalamang (a language with fewer than 200 speakers), it learned to translate English to Kalamang at a level comparable to a human studying the same material.
Google CEO Sundar Pichai and DeepMind CEO Demis Hassabis jointly announced the release, emphasizing that Gemini 1.5 represents “a step change in our approach” and that the longer context windows “show us the promise of what is possible.” Pricing tiers starting at the standard 128,000-token window and scaling up to 1 million tokens are planned for wider release. Developers and enterprise customers are being invited to sign up for the limited preview starting today.
The record
The site notes the 1-million-token capability and reports that Google also tested up to 10 million tokens, positioning the model against GPT-4 Turbo and Claude 2.1.
One year later — open only if you can handle spoilers
By mid-2026, Gemini 1.5 Pro’s long-context capabilities became a standard feature across major models, but the initial moat was short-lived as competitors quickly matched the 1M-token window. The MoE architecture, however, proved influential and was adopted by many subsequent models.