Google Gemini is testing with 10 million tokens

Google has launched Gemini 1.5, a major upgrade that will allow its general AI model to analyse dramatically more content than before. To put that into numbers, it’s gone from 32,000 tokens – slices of content, such as part of an image – to 128,000 tokens in one swoop. The company even has 10 million tokens in its sights.

In Google’s words, tokens are “the smallest building blocks, like part of a word, image or video”. Or you can think of them as subdivisions of content that AIs analyse to devise a response to a prompt.

Google is pitching Gemini as a multipurpose tool that companies could deploy for enterprise use cases from speeding up coding to improving productivity, while consumers could put the model to work as an aid to creativity or scheduling. 

Released in December 2023, Google claims the AI was designed to be equally content working across different types of information, including text, images, audio and video, as well as code. 

Gemini was originally released in three different variants: Ultra, Pro and Nano. Nano was designed for tasks carried out on PCs and phones (including Google’s Pixel 8 Pro, naturally), while Pro is available for better scaling, and Ultra for the most complex types of requests. 

The first version of 1.5 released will be Pro, and Google claims it’s already outperforming Gemini 1.0 Ultra released late last year. 

That isn’t surprising. Gemini 1.0 used a context window of 32,000 tokens; with 1.5 Pro, that window is up to 128,000 tokens. That’s the same context window as GPT-4 Turbo. 

Related reading: What is Sora? Even the AI experts aren’t sure

Google Gemini beta tests 1 million tokens

Google is currently testing a 1-million token window with a limited group of customers, including developers and enterprise users. For context, 1 million tokens is about the same as analysing 30,000 lines of code or 11 hours of audio simultaneously. The company is warning users to expect longer latency times when using the 1-million token window as it’s still in testing. However, the company is working on cutting that latency as well as reducing the compute resources needed to run Gemini 1.5 and improving the user experience. 

And Google looks like it has no plans to stop at a 1-million token window – the company said it has already tested a 10-million token window.

To showcase the latest version of the AI, Google demoed how the latest Gemini iteration can be used to query the 400-page transcript of the Apollo 11 mission or a 44-minute Buster Keaton movie, as well as learn to speak the Kalamang language (current speakers: 200) from a grammar book.  

Perhaps more useful for enterprise users, however, it can also now better digest large blocks of code and offer suggestions for improvements. 

Google has already embedded Gemini into a number of its own products, including Workspace and Cloud, and claims “hundreds of thousands” of developers and corporates are now using the model. 

How much does Gemini Advanced cost?

Earlier this month, Google announced it would be rebranding its Gemini 1.0 Pro-powered chatbot, Bard, to Gemini.

At the same time, it launched Gemini Advanced, a chatbot based on Ultra 1.0. This costs US$19.99/£18.99/AU$32.99 a month to access via a subscription to Google One AI Premium Plan. 

In addition to the Gemini Advanced on the web, this will soon be integrated into Gmail, Docs, Slides and Meet. One AI Premium also gives you 2TB of storage, a VPN and Google Workspace “premium features”. Think longer calls on Meet, call recording and live streaming of Meet calls on YouTube.