Harvard and Chutes Released Massive AI Dataset

Researchers published data on over 6 billion LLM requests to analyze user behavior and model efficiency.

Updated on Sept. 21, 2026 in Artificial Intelligence

Isometric editorial illustration showing a vast, dense grid of millions of uniform translucent cubes representing large-scale data. — Harvard University and Chutes have released a comprehensive dataset of over 6.12 billion AI requests to improve understanding of model efficiency and user interaction. AI Illustration. Upload story photo >

Live Poll

Are you using automated AI systems to perform specific, targeted tasks more frequently?

Harvard University and Chutes have released a comprehensive public dataset containing more than 6.12 billion large language model requests. The information covers over 9,000 models and was collected over a one-year period.

Why it matters

The findings provide unprecedented insight into how users interact with AI models, revealing that nearly all repeat requests happen in rapid succession. This data helps developers better understand token usage patterns and request frequency in real-world environments.

The dataset spans 35.8 trillion input tokens and 2.52 trillion output tokens. Researchers noted a significant shift in efficiency, as response lengths decreased from hundreds of tokens to fewer than 100 throughout the study year.

The players

Harvard University

This Ivy League research institution co-published the dataset and hosted it via cloud infrastructure.

Chutes

This organization collaborated on the collection and release of the large-scale LLM request dataset.

University of Chicago

This academic institution acted as a collaborator on the methodology used to structure the findings.

The details

The data, available via GitHub and a Harvard S3 bucket, captures activity from April 2025 to April 2026. To protect privacy, user identifiers in the set rotate every three months.

Dataset collection began on April 11, 2025.
Data contribution opt-in occurred from March to July 2026.
The collection period concluded on April 12, 2026.
The dataset was officially released on September 21, 2026.

The Tech Race

This release follows the integration of decentralized computation networks like Bittensor Subnet 64 to host large-scale AI research. The availability of this data represents a shift from proprietary silos toward open analysis of model efficiency and user demand.

Developers and AI enthusiasts can use this dataset to optimize model latency and understand how shortening output lengths improves system responsiveness. The findings suggest that users frequently repeat queries within minutes, highlighting the need for efficient caching in AI-powered tools.

The takeaway

The rapid drop in response lengths over the year demonstrates a move toward more concise and efficient AI interactions. Understanding these patterns is essential for anyone building applications that rely on consistent, high-speed LLM performance.