What 1,000 Trillion Tokens Reveal About the Rise of Open‑Source LLMs

Key Findings from the 1 000 Trillion‑Token Study

The joint analysis by a16z and OpenRouter examined real‑world LLM usage across more than one quadrillion tokens, exposing how the market has shifted dramatically in the past year.

Open‑Source Models Capture One‑Third of the Market

Open‑source LLMs now account for roughly 30 % of total token consumption, with DeepSeek alone contributing 13‑14 % and ranking among the global top‑10 open‑source providers.

Programming Tasks Become a Must‑Have Use Case

Requests for code generation have exploded from 11 % at the start of the year to over 50 % of all LLM traffic, making AI‑assisted programming a dominant application.

Role‑Play Scenarios Power a Hidden Gold Mine

Interactive storytelling and role‑play consume about half of the open‑source token volume, far outpacing programming (≈10 %). This “Cinderella glass‑shoe” effect shows that users who find a perfect niche stay highly engaged.

Market Segmentation by Cost and Quality

High‑end players : Claude 3.7 Sonnet, $2 per M tokens, high usage in enterprise settings.

Value monsters : DeepSeek V3, Gemini 2.0 Flash, $0.4 per M tokens, dominate high‑volume tasks.

Long‑tail small models : 7B‑scale models, cheap but low adoption due to limited performance.

Luxury tier : GPT‑4/5 Pro, $35 per M tokens, serve niche scenarios where cost is irrelevant.

Pricing Elasticity Is Extremely Low

Even a 10 % price cut only yields a 0.5‑0.7 % increase in usage, indicating that most users prioritize capability over cost.

Reference

https://openrouter.ai/state-of-ai