Gemini 1.5 Pro
Google DeepMind's mid-size multimodal model with up to a 1 million token context window
Gemini 1.5 Pro is a mid-size multimodal model from Google DeepMind, announced in February 2024 as the first model of the Gemini 1.5 generation. It matches the quality of Gemini 1.0 Ultra while using less compute, and its main breakthrough is a context window running up to 1 million tokens.
- Context window of 128,000 tokens, up to 1 million in preview
- Matched Gemini 1.0 Ultra quality while using less compute
- Beat 1.0 Pro on 87% of Google's LLM development benchmarks
- Found hidden facts 99% of the time in 1 million token haystacks
- Processed 1 hour of video, 11 hours of audio or 700,000 words at once
What it is
Gemini 1.5 Pro is a mid-size multimodal model built by the Gemini team at Google DeepMind. Announced in February 2024, it was the first model released from the Gemini 1.5 generation and was offered to developers and enterprise customers for early testing. Google says it reaches a similar quality level to Gemini 1.0 Ultra, its largest model at the time, while needing less compute. Its defining feature is context: it handles 128,000 tokens by default, with a preview running up to 1 million tokens, which lets it take in long documents, hours of audio or video, and large codebases in a single prompt.
How it came about
Gemini 1.0 Ultra had gone live the week before the 1.5 announcement, first in Gemini Advanced and then for developers through the Gemini API in AI Studio and Vertex AI. Sundar Pichai presented 1.5 as the next generation, and Google DeepMind CEO Demis Hassabis wrote the technical account on behalf of the Gemini team. The company framed 1.5 as a step change in how it builds foundation models, with gains in training and serving efficiency. A technical report followed on arXiv in March 2024.
Inside the model
The model uses a Mixture-of-Experts design. Instead of one large neural network, it is split into smaller expert networks, and only the pathways relevant to a given input are activated, which the company says makes training and serving more efficient. Google describes itself as an early pioneer of this technique, citing work such as Sparsely-Gated MoE, GShard-Transformer and Switch-Transformer. Machine learning work on the model raised context capacity well past the 32,000 tokens of Gemini 1.0, to 1 million tokens in production, with 10 million tested in research. Tokens here can be parts of words, images, video, audio or code.
What it can do
Google reports that 1.5 Pro beat 1.0 Pro on 87% of the benchmarks used to develop its large language models, and landed at a broadly similar level to 1.0 Ultra. In the Needle In A Haystack test, which hides a fact inside a long block of text, it located the passage 99% of the time at lengths up to 1 million tokens. It can reason across a 402-page Apollo 11 transcript, spot plot points in a 44-minute silent Buster Keaton film, and work with prompts of over 100,000 lines of code. On the MTOB benchmark it learned to translate English into Kalamang, a language with fewer than 200 speakers, from a grammar manual alone.
How to use it
A limited preview went to developers and enterprise customers through AI Studio and Vertex AI, with the 1 million token window in private preview. Developers could sign up to test 1.5 Pro in AI Studio. Google said it was working on latency and compute optimisations as the longer window rolled out more widely.
Safety and testing
Google says 1.5 Pro was tested against its AI Principles and safety policies before release, covering content safety and representational harms, with more testing planned. The company describes red-teaming work on potential harms and says it is developing new evaluations aimed at the model's long-context abilities. Findings from this research feed into its governance and model development processes.
Why it matters
Google presents the 1 million token window as the longest of any large-scale foundation model at the time, and says longer context opens up new kinds of application for developers and enterprises. Because the capability was new at this scale, the team said it was still building evaluations to test it. The company also credits the efficiency of the MoE design with letting it train and ship new Gemini versions faster.
Questions people ask about Gemini 1.5 Pro
What is Gemini 1.5 Pro?
It is a mid-size multimodal model from Google DeepMind, announced in February 2024 as the first model of the Gemini 1.5 generation. It handles text, images, audio, video and code, and Google says it performs at a similar level to the larger Gemini 1.0 Ultra while using less compute.
What is special about its context window?
It runs 128,000 tokens by default, with a preview reaching 1 million tokens, and Google says it tested up to 10 million in research. That is enough for an hour of video, 11 hours of audio, over 30,000 lines of code or more than 700,000 words in one prompt.
How does it handle such long inputs?
It uses a Mixture-of-Experts architecture, where the model is split into smaller expert networks and only the parts relevant to the input are activated. Google says this makes the model cheaper to train and serve while keeping quality high.
How can developers use it?
Google offered a limited preview through AI Studio and Vertex AI, with the 1 million token window in private preview for a small group of developers and enterprise customers. Developers could sign up in AI Studio to test it.
What safety work was done?
Google says it ran evaluations on content safety and representational harms before release, used red-teaming to probe for potential harms, and is building new tests aimed at the long-context capability. It describes this as part of its AI Principles and safety policy process.
Good reading and sources
There is no English Wikipedia article on Gemini 1.5 Pro yet, so the facts on this page come from the sources listed above: blog.google/technology/ai/google-gemini-next-generation-model-february, arxiv.org/abs/2403.05530. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.