AI Usage Shifts to Reasoning and Programming, Driven by Developers

The AI Daily Brief: Artificial Intelligence News and Analysis · December 08, 2025 · Listen to Original Episode →

Original Title: What People Are Actually Using AI For Right Now

TL;DR

Reasoning model token usage surged from negligible to over 50% of all tokens consumed in 2025, signifying a paradigm shift towards AI capable of complex problem-solving and autonomy.
Programming has become the dominant AI use case, now accounting for over 50% of usage, with average prompt lengths quadrupling to accommodate detailed code and documentation inputs.
Open-source models, particularly those from China, captured a third of usage by Q4 2025, demonstrating their viability for diverse applications and challenging the dominance of closed-source alternatives.
Role-play and creative dialogue constitute over 50% of open-source model usage, indicating a significant demand for AI in entertainment and non-work-related generative tasks.
Developers are increasingly adopting hybrid stacks, leveraging closed models for high-value workloads and open models for high-volume tasks, indicating a pragmatic approach to AI integration.
Early adoption of new models, especially for critical workloads like programming, creates near-permanent user lock-in, with users willing to pay a premium for reduced debugging time.

Deep Dive

A new empirical study analyzing over 100 trillion tokens reveals that real-world AI usage has fundamentally shifted towards reasoning and coding workloads, driven by developers and power users. This empirical data challenges earlier assumptions about AI's primary applications and highlights the increasing sophistication and specialization of AI tool adoption heading into 2026.

The dominant trend observed is the dramatic rise of reasoning models, which now account for over 50% of token usage, a stark contrast to their negligible presence at the start of 2025. This shift coincides with a significant increase in tool invocation, with 15% of requests now utilizing external tools, indicating a move towards more autonomous and integrated AI capabilities. Programming has emerged as the paramount use case, escalating from 11% to over 50% of usage, with average prompt lengths quadrupling to accommodate complex coding tasks, documentation, and logs rather than simple essay generation. This surge in AI-assisted coding signifies a fundamental change in software development workflows.

Alongside these productivity-focused applications, role-playing and creative dialogue represent the other major use case, particularly prevalent in open-source models, accounting for over 50% of their usage. This suggests a bifurcated AI landscape: closed-source models are increasingly leveraged for high-value, complex tasks like programming, while open-source models cater to high-volume, creative, or specialized applications, including entertainment and fantasy role-play. The study also underscores the growing significance of open-source models, particularly those originating from China, which now constitute about a third of overall usage, though this growth has plateaued recently with advancements in closed-weight models. This hybrid approach, utilizing both closed and open models for distinct purposes, defines the modern AI stack.

The implications of these findings are substantial. The "Cinderella glass slipper effect" indicates that early adopters of models that successfully address specific, painful workloads create strong user lock-in, suggesting that differentiation is key, not necessarily being the cheapest. For instance, Anthropic's Claude is strongly associated with programming, while Deepseek excels in role-play, demonstrating model specialization. Price inelasticity is also evident, with users willing to pay a premium for models that significantly reduce debugging or development time. This points to a market where efficiency and effectiveness for specific tasks outweigh cost considerations. The rise of wrappers and scaffolds like OpenRouter is also a critical development, enabling users to navigate a sensitively differentiated and rapidly evolving AI market by maintaining optionality across providers and models.

The key takeaway is that AI usage is rapidly maturing beyond general-purpose chatbots into specialized, high-impact applications, with coding and reasoning leading the charge. This empirical data signals a decisive shift in how AI is being integrated into professional workflows, emphasizing developers' need for powerful, context-aware tools and the growing importance of specialized models tailored to distinct use cases.

Action Items

Analyze 100 trillion token study: Identify 3-5 dominant use cases (e.g., programming, roleplay) and their associated model types (open vs. closed source).
Track programming workload growth: Measure prompt token length increase (target 4x) for coding tasks to understand evolving user needs.
Evaluate model lock-in: For 3-5 key models (e.g., Claude, Gemini), measure user retention rates after 6 months to assess stickiness.
Measure price inelasticity: For 3-5 high-demand workloads (e.g., debugging), calculate user willingness to pay premium token costs.
Audit open-source model usage: For 3-5 Chinese open-source models, analyze the shift in usage from roleplay to programming/technology.

Key Quotes

"The study they released last week is called the state of ai an empirical 100 trillion token study with open router in the abstract they write we analyzed over 100 trillion tokens of real world llm interactions across tasks geographies and time the findings underscore that the way developers and end users engage with llms in the wild is complex and multifaceted."

This quote introduces the core data source for the discussion: a large-scale analysis of real-world AI usage. The authors, OpenRouter and a16z, examined over 100 trillion tokens to understand how developers and users interact with Large Language Models (LLMs). This empirical approach aims to reveal the actual, complex patterns of AI engagement beyond theoretical assumptions.

"The dominant use case by far has become programming early in 2025 programming was around 11 of usage and now it is over 50 we are coming up towards end of year episodes and i think any accounting of 2025 has to start with the fact that the dominant and most important phenomenon of this year in ai was the rise of ai coding."

This quote highlights a significant shift in AI application, as identified by the study. The author emphasizes that programming has emerged as the primary use case for AI, growing from 11% of usage at the start of 2025 to over 50% by the end of the year. This trend is presented as a defining characteristic of AI development in 2025.

"The other use case that dominates is role play basically everything in and around chatting with ai in a fantasy context from innocent to not so safe for work that is particularly true for open source models where role play and or creative dialogue as they put it accounted for more than 50 of oss usage."

This quote points to a second major use case identified in the study, particularly prevalent with open-source models. The author notes that role-playing and creative dialogue, encompassing a wide range of fantasy-based interactions, constitute over half of the usage for open-source AI. This suggests a significant demand for AI in entertainment and creative expression.

"If you want a single picture of the modern stack closed models are for high value workloads and open models are for high volume workloads and as they point out teams are using both."

This quote from the study describes a bifurcated strategy in the modern AI stack. The author explains that closed-source models are typically employed for tasks requiring high value, while open-source models are utilized for high-volume applications. Importantly, the study indicates that teams are leveraging both types of models to optimize their AI deployments.

"A model that's the first to nail a painful workload creates near permanent lock in early 2025 cohorts of claude for sonnet and gemini 2 5 pro still retain 40 to 50 of user six months later while every later cohort churns relatedly he points out demand is wildly price inelastic users happily pay 10 to 50x more per token for claude or gpt 5 if it saves them 10 minutes of debugging being cheap is nowhere near enough."

This quote, attributed to Tang Yan, discusses user retention and price sensitivity in the AI market. Yan argues that the first model to effectively solve a difficult problem creates strong user loyalty, with early adopters of models like Claude and Gemini 2.5 Pro showing high retention rates. Furthermore, Yan highlights that users are willing to pay significantly more for AI models if they offer substantial time savings, indicating that price is a secondary concern for critical workloads.

Resources

External Resources

Books

"The State of AI: An Empirical 100 Trillion Token Study" by OpenRouter and a16z - Mentioned as the source of data analyzed for real-world LLM interactions.

Articles & Papers

"The State of AI: An Empirical 100 Trillion Token Study" (OpenRouter and a16z) - Discussed as an empirical analysis of over 100 trillion tokens of real-world LLM interactions to reveal developer and power user AI activities.

People

Amar - Product and design lead at Google DeepMind, mentioned in relation to vibe coding with Gemini 3.
Tom Warren - Verge reporter, cited for reporting on GPT-5.2 release rumors.
Nick Churlie - Head of ChatGPT, quoted regarding rumors of ads in ChatGPT.
Mark Chen - Chief Research Officer, acknowledged issues with AI suggestions feeling like advertising.
Benjamin Dekraker - User quoted for his reaction to AI suggestions resembling advertising.
John G. Andrea - Former Senior VP of Machine Learning and AI Strategy at Apple, mentioned as a departure from Apple.
Allen Die - Former Head of UX Design at Apple, mentioned as joining Meta.
Johnny Srouji - Senior VP of Hardware Technologies at Apple, discussed in relation to considering leaving Apple.
Tim Cook - CEO of Apple, mentioned as discussing potential departure with Srouji.
Mark Gurman - Bloomberg correspondent, reported on Srouji's potential departure from Apple.
Nicholas - Twitter user, commented on Apple's AI-capable chips and software utilization.
Tang Yan - Runs the Chain of Thought AI newsletter, provided observations on model usage and lock-in.
Token Bender - Commented on extrapolating patterns from the OpenRouter study.
Onchardry - Stated that AI usage is mostly long-run in coding jobs with tool calls.
J Little - Commented on DeepSeek's role-play capabilities and usage.
Sean Chahan - Commented on the market gap versus reality gap in AI usage for fan fiction and debugging.
Brian Catano - Reflected on the success of Cursor and OpenRouter as scaffolds and wrappers.

Organizations & Institutions

Google DeepMind - Mentioned in relation to Amar, their product and design lead.
OpenAI - Discussed in relation to GPT-5.2 release rumors and user growth slowdown.
Sensor Tower - Provided data suggesting a slowdown in ChatGPT user growth.
Bloomberg - Reported on investor sentiment shifting towards Google's AI ecosystem over OpenAI's.
Polymarket - Mentioned as a betting market where AI model dominance is being wagered.
Apple - Discussed in relation to departures from its AI team and hardware technologies.
Meta - Mentioned for acquiring Limitless and for its chatbot providing up-to-date news content.
The Verge - Cited for reporting on GPT-5.2 release rumors and OpenAI's UX cleanup efforts.
KPMG - Sponsor, mentioned for their "You Can with AI" podcast.
Google AI Studio - Mentioned in relation to Gemini 3 and building apps.
Rovo - Sponsor, mentioned for AI-powered search, chat, and agents.
AssemblyAI - Sponsor, mentioned for building Voice AI apps.
LandfallIP - Sponsor, mentioned for AI to navigate the patent process.
Blitzy.com - Sponsor, mentioned for building enterprise software.
Robots & Pencils - Sponsor, mentioned for cloud-native AI solutions.
Superintelligent - Sponsor, mentioned for their AI planning platform and "Plateau Breaker" assessment.
OpenRouter - Startup providing a unified API for LLMs, co-author of the AI usage study.
a16z - Prominent venture fund, co-author of the AI usage study.
National Football League (NFL) - Mentioned in the example of categorical reference lists.
New England Patriots - Mentioned in the example of categorical reference lists.
Pro Football Focus (PFF) - Mentioned in the example of categorical reference lists.
DeepSeek - Chinese open-source model, noted for its role-play usage.
Anthropic - Mentioned in relation to Claude's usage for programming.
Cnn - Media partner for Meta AI.
Fox News - Media partner for Meta AI.
USA Today - Media partner for Meta AI.
People Inc. - Media partner for Meta AI.
Chicago Tribune - Filed a lawsuit against Perplexity.
New York Times - Filed a lawsuit against Perplexity.
AWS - Mentioned in relation to Robots & Pencils being an AWS certified partner.
Jira - Mentioned as a platform where Rovo runs.
Confluence - Mentioned as a platform where Rovo runs.
Jira Service Management - Mentioned as a platform where Rovo runs.

Tools & Software

Gemini 3 - Mentioned in relation to Google AI Studio for building apps.
ChatGPT - Discussed in relation to user growth slowdown and potential ad integration.
Cursor - Mentioned as a tool where users can plug in OpenRouter API keys.
GPT-5.2 - Rumored upcoming model from OpenAI.
Claude - Mentioned as a model used for programming tasks.
Gemini 2.5 Pro - Mentioned as a model retaining users.
GPT-4.5 - Mentioned in relation to Opus.
Sonnet - Mentioned as a model retaining users.
Opus - Mentioned in relation to GPT-4.5.
Grok - Mentioned as dominating total consumption charts on OpenRouter due to a promotion.
VS Code - Mentioned as a base for Cursor.

Websites & Online Resources

OpenRouter.ai - Website where the AI usage study can be found.
ai.studio/build - URL for Google AI Studio.
rovo.com - Website for Rovo.
assemblyai.com/brief - URL for AssemblyAI.
landfallip.com - Website for LandfallIP.
blitzy.com - Website for Blitzy.
robotsandpencils.com - Website for Robots & Pencils.
besuper.ai - Website for Superintelligent.
patreon.com/aidailybrief - URL for subscribing to an ad-free version of the show.
pod.link/1680633614 - Link to subscribe to The AI Daily Brief podcast.

Podcasts & Audio

The AI Daily Brief: Artificial Intelligence News and Analysis - The podcast for which this episode is a transcript.
KPMG "You Can with AI" podcast - Mentioned as a new podcast from KPMG.

Other Resources

Vibe Coding - A method for describing an app and having Gemini wire up models.
Reasoning Models - A category of AI models whose token usage has significantly increased.
Tool-Use Invocation - The act of requesting AI models to use tools, showing an increase in usage.
Long-Context Programming Tasks - A dominant use case for AI, involving extensive code and documentation.
Hybrid Stacks of Closed and Open Models - A pattern in the practical AI landscape.
Code Red - OpenAI's response to competitive pressure.
AI Wearables - A category of devices that has not yet achieved product-market fit.
The Agent Readiness Audit - An assessment offered by Superintelligent.
LLM Interactions - Real-world interactions analyzed in the OpenRouter study.
Open Source Models - Models that have seen significant growth in usage, particularly Chinese models.
Closed Weight Models - Models that have seen major advances, such as Gemini 3 and GPT-5.1.
Chinese Open Source Models - Noted for their rapid growth and viability for various use cases.
Foundational Cohort - Users who persist with a new model, creating a base group.
Cinderella Glass Slipper Effect - The phenomenon of new models being tried by many users, with a subset persisting.
AI Platforms vs. Copyright Holders - An ongoing discussion and legal battle.
AI Native SDLC - Software Development Lifecycle incorporating AI natively.
Teamwork Graph - Atlassian's intelligence layer unifying data across apps.
AI Planning Platform - A type of platform offered by Superintelligent.
Plateau Breaker - A new assessment type from Superintelligent for breaking AI plateaus.
AI Coding - Identified as the dominant and most important phenomenon of 2025.
Roleplay - A dominant use case, particularly for open-source models.
Creative Dialogue - Another term used for roleplay in open-source model usage.
AI Enabled Smart Glasses - Products produced by Meta's Reality Labs.
AI Platforms - A category of technology discussed in the context of market competition.
Copyright Holders - Entities involved in legal disputes with AI platforms.
Technical Content Protection Measures - Measures that web crawlers may evade.
Scaffolds and Wrappers - Tools that help manage complexity in the AI market.
Sensitive Differentiation - The characteristic of the AI market where small input changes yield large output differences.
Sticky Models - AI models that retain users over time, potentially with features like memory.