OpenRouter’s 100 Trillion Token Study Reveals AI Shifts
Malika Aubakirova and Anjney Midha at a16z put out a report on over 100 trillion tokens of real-world LLM usage from OpenRouter. They pull data from the platform to show how developers work with AI. Check the full a16z piece here.
OpenRouter’s Scale
OpenRouter reaches over 5 million developers. It sends traffic to 300+ models across 60+ providers. Token volume jumped from 10 trillion a year to more than 100 trillion by mid-2025. Last week alone, it hit over 1 trillion tokens per day. OpenAI’s full API did 8.6 trillion daily in October, for scale.
How They Did the Study
The team looked at this massive real-world data to track changes in AI use. OpenRouter’s view covers industries, places, and model types. The full report breaks down open vs. closed models, geography, task categories, and user retention over time.
Main Results
Agentic workflows grow fastest. Developers run models in loops: plan, call tools or APIs, revise, repeat. Prompts get longer, chats have more back-and-forth, and reasoning or tool-focused models take bigger shares.
This picked up after OpenAI’s o1 launch on December 5, 2024. o1 added built-in multi-step thinking, unlike one-shot text generators.
Other patterns:
- Open-source picks up speed, especially reasoning models like DeepSeek R1 and Kimi K2. Cost and flexibility help.
- Coding and creative tasks eat up the most tokens.
- Users switch models for big capability jumps and stay put.
Key Takeaways and What Comes Next
AI moves from chat replies to work partners that handle tools and long tasks. Builders win by focusing on orchestration and reliability there.
For researchers, the data raises points like why roleplay rules across models, tool patterns for new designs, and retention signals for hits.
Grab the full study on OpenRouter for the deep cuts.
