Introduction

AI use is changing fast as people shift from simple chats to full task workflows. This study shows how 100 trillion tokens reveal that change.

What's the problem it solves?

Most talk about AI is based on hype or small tests. This report shows how people actually use AI at massive scale so builders know what to focus on next.

Quick Summary

The study looks at more than 100 trillion tokens of real use from OpenRouter, a platform that sends requests to over 300 models for 5 million developers. The data shows a clear shift. People are not using AI just to answer questions anymore. They are using models to plan steps, call tools, fetch data, and work through tasks over many turns. This rise in reasoning and agent-like behavior is the fastest-growing trend on the platform.

The model market is also shifting. Open-source models that are strong at reasoning are growing fast because they are cheaper and flexible. Coding and creative work still drive most token use. Some models stay popular even if they do not top benchmarks because users like their style and feel. When a model unlocks a new power, people switch to it and rarely switch back.

Overall, AI is moving from a chat box to a working partner. The next wave of winners will build around long tasks, better control, and models that can act in steps instead of stopping after one reply.

Key Takeaways

• AI use is shifting from single answers to multi-step reasoning and tool use.
• Agent-like workflows are the fastest-growing pattern in real traffic.
• Open-source reasoning models are gaining share because of cost and freedom.
• Coding and creative tasks still use the most tokens.
• User preference and model personality matter more than benchmarks suggest.
• Breakthrough features cause users to switch models fast and stay loyal.
• Real usage data shows where the industry is going, not just where benchmarks point.

What to do

• Build features around multi-step tasks, not just single replies.
• Add tool use and planning flows into your product early.
• Test open-source reasoning models for cost and speed wins.
• Watch user behavior, not just benchmarks, when picking models.
• Track breakthrough features that could shift user loyalty.
• Design for long, ongoing sessions so the model can act like a helper, not a chatbot.

‍