How Did AI Model Usage Change Throughout 2025?

TL;DR

Closed-weight models still served more than 70% of OpenRouter token traffic, even as open-weight models briefly reached roughly one-third of usage after major releases. OpenRouter’s analysis of more than 100 trillion tokens also found strong demand for creative roleplay, rapid growth in Chinese open-weight models, increasing agentic inference, and higher retention among people who adopted language models early.

Transcript

2025 has been a crazy year for AI. At the end of 2024, OpenAI introduced 01 which brought reasoning models to the entire world and right when 2025 started, Deepseek rose with Deepseek R1, an openw weight model that allowed for reasoning to be done for everyone. These changes shifted the whole industry and we've seen crazy things happening since. So... Read More

Key Insights

• OpenRouter’s dataset covers more than 100 trillion tokens of real-world language-model inference across tasks, geographies, and time, giving its researchers a broad view of how developers and end users select models and apply them in practice.
• Closed-weight models remained the majority on OpenRouter, serving more than 70% of token traffic after open-weight usage retreated from a brief peak above 30%. This is notable because OpenRouter makes open-weight models especially convenient to access, test, and integrate.
• Open-weight model adoption grew steadily during 2025, reaching approximately one-third of usage by late in the year. Major releases including DeepSeek V3 and Kimi K2 produced visible usage spikes, while Qwen 3 and GPT OSS also gained adoption.
• Chinese models became the majority of open-weight usage on OpenRouter by the end of 2025. Their share had been very small at the beginning of the year, demonstrating how quickly competitive releases from Chinese laboratories changed the open-model segment.
• Creative roleplay is an unusually popular language-model use case, exceeding the prominence that productivity-centered assumptions would suggest. Programming ranks second in the categories discussed, reinforcing that actual usage encompasses entertainment, creativity, technical assistance, and many other activities.
• GPT OSS leads OpenRouter usage within the legal, technology, finance, and science categories described in the transcript. It also ranks second for programming, suggesting that its strongest adoption appears in analytical, technical, and data-oriented applications rather than across every category.
• OpenRouter’s categorization method uses approximately 0.25% of prompts and responses, processes only the first 1,000 characters, and retains classifications with confidence scores of at least 0.5. The study uses this classified sample because OpenRouter’s overall query volume is extremely large.
• Early language-model adopters show stronger retention than people who arrived later, a pattern the study calls the glass slipper effect. This cohort difference suggests that adoption timing is relevant when evaluating long-term engagement and designing language-model products.

Questions & Answers

Q: How much OpenRouter traffic used open-weight models in 2025?

Open-weight models grew to approximately one-third of OpenRouter usage by late 2025 and briefly exceeded 30% around the releases of Qwen 3 and GPT OSS. Their share later returned to the 20% range described in the transcript. Closed-weight models therefore continued serving more than 70% of traffic, despite OpenRouter making open-weight alternatives easy to access and integrate.

Q: Why did Chinese open-weight models grow rapidly in 2025?

Chinese open-weight usage accelerated alongside competitive releases such as DeepSeek V3 and Kimi K2. At the beginning of 2025, these models represented very little OpenRouter traffic, but by the end of the year they constituted the majority of open-weight usage. The transcript attributes visible adoption spikes to important releases and notes that OpenRouter lets customers use these models through providers with preferred data-retention policies.

Q: What were the most popular language-model use cases on OpenRouter?

Creative roleplay showed outsized popularity and challenged the assumption that language models are used primarily for conventional productivity. Programming ranked second among the use cases highlighted in the transcript. OpenRouter’s broader taxonomy also included translation, general questions and answers, knowledge, productivity, writing, education, literature, creative writing, adult content, and a collection of other activities.

Q: How did OpenRouter classify prompts without exposing user content?

OpenRouter internally categorized a random sample of approximately 0.25% of prompts and responses with a non-proprietary Google tag classifier. The system classified up to the first 1,000 characters and produced hierarchical category paths with confidence scores between zero and one. Processing occurred within OpenRouter’s infrastructure, classifications were anonymous, and the researchers had no direct access to prompts or model outputs.

Q: What confidence threshold did OpenRouter use for content categories?

OpenRouter excluded classifications with confidence scores below 0.5. Its classifier could return one or more hierarchical category paths, such as programming or role-playing games, together with scores ranging from zero to one. Applying the threshold meant that uncertain classifications were ignored, while the remaining categorized sample supported comparisons among model types, application categories, locations, token directions, and other metadata.

Q: Why did reasoning models change the AI industry in 2025?

The release of OpenAI o1 on December 5, 2024 introduced the first widely adopted reasoning model described in the transcript, moving usage from single-pass pattern generation toward multi-step deliberation during inference. DeepSeek R1 followed as an open-weight reasoning model near the beginning of 2025. Together, these developments accelerated deployment experiments and enabled developers to explore new classes of applications.

Q: Which tasks showed strong adoption of GPT OSS?

GPT OSS ranked first among models used for legal, technology, finance, and science questions in the OpenRouter category data discussed. It ranked second in programming as well. The presenter says he frequently uses it for data-related work, especially through fast providers, and argues that its release timing contributed to its strengths receiving less attention than they deserved.

Q: What is the glass slipper effect in language-model retention?

The glass slipper effect describes OpenRouter’s observation that people who began using language models early tend to remain active longer than users who arrived later. It is a cohort-based retention pattern rather than a claim that all early users behave identically. The finding indicates that the timing of adoption can help explain differences in continued engagement with language-model services.

Summary & Key Takeaways

• OpenRouter analyzed metadata from more than 100 trillion tokens of real-world language-model inference to examine usage across tasks, locations, model types, costs, and retention patterns. The research emphasizes that practical language-model activity is more varied than conventional productivity narratives suggest, with creative roleplay, coding assistance, and agentic inference all representing important areas of adoption.

• Open-weight models expanded steadily during 2025 and briefly accounted for more than 30% of OpenRouter inference around major releases. DeepSeek V3, Kimi K2, Qwen 3, and GPT OSS contributed to adoption, while Chinese models grew from a small portion of open-weight traffic to its majority by the end of the year.

• The study categorized a random sample of approximately 0.25% of prompts and responses using Google Cloud’s text-classification system. Classification occurred within OpenRouter’s infrastructure, was not linked to individual customers, and excluded results below a 0.5 confidence threshold. Researchers received categories and metadata rather than direct access to prompts or model outputs.