Reflecting on AI in 2025

best model counter: 14

Ts is not worth a 45-minute video. TLDR: The 2025 "State of AI" study by OpenRouter highlights several major shifts in artificial intelligence, most notably the transition from single-pass pattern generation to multi-step deliberation through reasoning models like OpenAI’s o1 and DeepSeek R1. Open-weight models have reached over 30% of total token usage, with Chinese labs such as DeepSeek and Qwen leading this expansion and creating a highly fragmented, competitive market. Surprisingly, creative roleplay has emerged as the top use case for open models, accounting for roughly half of all usage, while programming remains the second-largest category. The industry is also seeing a massive rise in agentic inference, with over 50% of tokens now dedicated to reasoning processes rather than simple text output. Ultimately, the ecosystem has moved toward a multi-model reality where developers and users choose specific models based on their unique strengths in reasoning, tool calling, or cost efficiency.

you need to start time stamping your videos if you're gonna make these long form videos.

What is considered roleplaying? If I tell an AI to pretend that it is a senior software engineer and ask it about software architecture it should be classified as roleplaying?

1 new ai article = 1 new theo video

ntroduction to 2025 AI Trends (0:00-0:40) Open Router's Unique Insights and Data (0:40-1:07) Sponsor Break: WorkOS (1:07-2:36) Open Router's Research Methodology (2:36-4:02) Understanding Data Categorization (4:02-6:07) Open vs. Closed Source Models Token Share (6:07-7:35) Rise of Chinese Open-Weight Models (7:35-8:29) Underrated GPT OSS Model and its Speed (8:29-12:02) Overall Token Usage Growth on Open Router (12:02-13:08) Key Open-Source Players by Token Volume (13:08-14:10) Impact of Free Models on Usage (14:10-15:00) Decline of DeepSeek's Dominance and Market Diversification (15:00-17:40) Model Size and Usage Trends (Small, Medium, Large) (18:55-21:10) Dominance of Roleplaying in Open-Weight Model Usage (21:10-23:29) Roleplaying and Programming Usage in Chinese Open-Weight Models (23:29-25:27) Roleplaying Breakdown for Open-Source Models (25:27-26:58) Rise of Agentic Inference and Reasoning Tokens (26:58-30:06) Top Used Models with Reasoning (30:06-32:02) Rising Adoption of Tool Calling (32:02-34:48) Anatomy of Prompts: Increase in Token Length (34:48-35:56) Agentic Inference as the New Default (35:56-37:00)

Happy Holidays everyone!

I'm literally in the middle of developing a roleplaying app with a lot of extra, cool features. Nice timing with this one, lol.

.25% of 100 trillion tokens means this study analyzed 250 BILLION tokens. Amazing..

Cool analysis Theo, always appreciate your sharp LLM evals, more pls!