Malika Aubakirova & Alex Atallah & Chris Clark & Justin Summerville & Anjney Midha — a16z & OpenRouter - State of AI An Empirical 100 Trillion Token Study with OpenRouter - December 2025

Free

An empirical 100 trillion token study of real-world LLM usage

FreeFree tier

Type

Open Source

Company

OpenRouter

Visit Website GitHub (38 stars)

About Malika Aubakirova & Alex Atallah & Chris Clark & Justin Summerville & Anjney Midha — a16z & OpenRouter - State of AI An Empirical 100 Trillion Token Study with OpenRouter - December 2025

An empirical research study analyzing over 100 trillion tokens of real-world large language model (LLM) interactions, conducted by researchers from a16z and OpenRouter. The study leverages OpenRouter's platform to examine usage patterns across tasks, geographies, and time, revealing key trends such as the rapid adoption of reasoning models (e.g., OpenAI's o1), the outsized popularity of creative roleplay and coding assistance, and the rise of agentic inference. It also identifies the 'Cinderella Glass Slipper effect' in user retention, where early cohorts exhibit persistently higher engagement than later ones. The findings offer actionable insights for model builders, AI developers, and infrastructure providers.

Key Features

Analysis of over 100 trillion tokens of real-world LLM interactions

Leverages OpenRouter platform data across 70+ providers and 400+ models

Observes adoption of reasoning models like OpenAI's o1

Highlights creative roleplay and coding assistance as dominant use cases

Identifies 'Cinderella Glass Slipper effect' in early user retention

Covers task diversity, geographic distribution, and temporal trends

Provides implications for model builders, AI developers, and infrastructure providers

Pros & Cons

Pros

Large-scale empirical data with over 100 trillion tokens provides robust statistical power
Covers a wide variety of LLMs from 70+ providers, reducing single-model bias
Identifies novel behavioral phenomena (Cinderella Glass Slipper effect)
Includes diverse task categories beyond typical productivity tasks
Free and openly accessible PDF download
Actionable insights for developers and researchers

Cons

Based solely on OpenRouter platform usage, which may not represent the entire LLM ecosystem
Focuses primarily on data from 2024–2025, limiting historical comparison
May have selection bias toward users who choose a multi-provider gateway
Does not include user demographic details beyond geography
Lacks controlled experimental validation for observed trends

Best For

Understanding real-world LLM usage patterns for product and model designAnalyzing adoption of reasoning models and agentic inferenceStudying user retention and cohort behavior in AI platformsIdentifying dominant tasks (creative roleplay, coding, productivity) for market insightsBenchmarking model performance and provider ecosystems

FAQ

What is the Cinderella Glass Slipper effect?

It refers to the finding that early user cohorts exhibit persistently higher engagement and retention compared to later cohorts, analogous to a perfect fit that diminishes over time.

What data was analyzed in the study?

Over 100 trillion tokens of real-world LLM interactions from OpenRouter, covering a wide variety of models, tasks, geographies, and time periods from late 2024 through 2025.

Who conducted the study?

The study was led by Malika Aubakirova and Anjney Midha from a16z, along with Alex Atallah, Chris Clark, and Justin Summerville from OpenRouter.

How can I access the full study?

The PDF is available for free download on the OpenRouter State of AI page at https://openrouter.ai/state-of-ai.