A 100 trillion token study of OpenRouter usage suggests the center of gravity in AI is moving from text generation to context-heavy, tool-using, reasoning workflows.
Source note: Malika Aubakirova, Alex Atallah, Chris Clark, Justin Summerville, Anjney Midha. “State of AI: An Empirical 100 Trillion Token Study with OpenRouter.” arXiv:2601.10088, January 15, 2026. https://arxiv.org/abs/2601.10088
Why This Paper Matters
Most arguments about AI usage are built from product launches, benchmark scores, anecdotes, or vendor dashboards that show only one slice of behavior. That leaves a gap between what models can do and what people actually route through them at scale.
This paper tries to narrow that gap. The authors analyze more than 100 trillion tokens of real-world usage metadata from OpenRouter, a platform that routes traffic across hundreds of models from dozens of providers. That vantage point is useful because OpenRouter is not one model, one app, or one provider’s consumer surface. It sees users choosing among open-weight and closed models, reasoning and non-reasoning systems, coding models, roleplay models, and cheaper utility models.
The most important takeaway is not that AI usage is growing. Everyone already knows that. The more useful point is that the shape of usage is changing. Requests are getting longer, more context-heavy, more tool-oriented, and more concentrated in workflows where the model has to reason over existing material rather than simply generate new text.
That matters for builders and operators because it changes what the AI stack needs to optimize. If AI were mostly chat, the winning system would optimize conversation quality. If AI is becoming infrastructure, the winning system has to optimize routing, context management, cost control, tool reliability, and workflow state.
The Idea in Plain English
The old mental model of AI usage was: someone asks a chatbot a question, and the model writes an answer.
The paper suggests a different model: someone gives an AI system a pile of context, asks it to reason through that context, and often expects it to participate in a larger workflow. That workflow might involve code, documents, transcripts, tool calls, model routing, or multi-step reasoning.
In that world, the expensive and important part is not just the final answer. It is the whole process around the answer: reading the input, carrying context, choosing the right model, using tools, and managing the cost of repeated inference.
This is why the paper is useful for understanding the AI market. It treats usage as an operating pattern, not a vibes contest about which model is smartest this week.
What the Researchers Tested
The authors study OpenRouter metadata covering billions of prompt-completion pairs and more than 100 trillion tokens. They do not inspect raw user prompt or completion text. Instead, the analysis uses metadata such as timing, model and provider identifiers, token usage, streaming or cancellation state, latency, geography inferred from billing location, and whether tool-calling features were invoked.
OpenRouter says it supported more than 300 active models from more than 60 providers in 2025, with more than half of usage originating outside the United States. That breadth gives the paper a useful cross-market view, though still not a complete view of all AI usage.
For content categorization, the paper uses an opt-in sampling process. OpenRouter categorizes about 0.25% of prompts and responses through an internal GoogleTagClassifier module. The paper says users must explicitly opt in to this anonymized analytics sharing, and that users who prefer full privacy can keep their data inaccessible for analytics. This means category-level findings should be treated as sampled behavioral evidence, not a complete label for every request.
The analysis then looks at several usage dimensions:
- Open-weight versus closed model adoption.
- Reasoning model adoption and agentic inference.
- Tool-calling behavior.
- Prompt and completion token lengths.
- Task categories such as programming, roleplay, translation, science, health, and finance.
- Geography and language distribution.
- Retention patterns by model cohort.
- Cost versus usage dynamics across categories and models.
What They Found
Reasoning Models Moved From Niche to Default-Like
The paper’s cleanest macro finding is that reasoning-optimized models grew from a negligible share of usage in early 2025 to more than half of total tokens. The authors are careful about what this means: the measure is the share of all tokens routed through reasoning-optimized models, not the share of explicit reasoning tokens inside outputs.
Still, the implication is large. Users are not only asking models to produce fluent text. They are increasingly choosing models that can manage task state, follow multi-step logic, and support agent-style workflows.
The paper names model churn inside this category too. In the most recent data, xAI’s Grok Code Fast 1 had the largest share of reasoning traffic among the tracked models, ahead of Gemini 2.5 Pro and Gemini 2.5 Flash, after a period when Gemini 2.5 Pro, DeepSeek R1, and Qwen3 were prominent. The exact leaderboard will age quickly. The structural point matters more: users are actively routing work toward models that behave more like reasoning engines than autocomplete engines.
The Prompt Is Growing Faster Than the Answer
The most useful evidence for operators is in the prompt-completion shape. Average prompt tokens per request grew roughly fourfold, from around 1.5K to more than 6K. Average completion tokens nearly tripled, from about 150 to 400.
That asymmetry matters. The model is being asked to read and process much more context before answering. The paper describes this as a shift away from open-ended generation, like “write me an essay,” toward reasoning over codebases, documents, transcripts, long conversations, and other user-provided material.
Programming is the sharpest example. Requests involving code understanding, debugging, and code generation routinely exceed 20K input tokens, while most other categories remain comparatively flat and low-volume in prompt size. The paper argues that programming is the main driver behind prompt-token growth.
For anyone building AI products, this is a neon sign: context ingestion is becoming a product and infrastructure problem.
Agentic Inference Shows Up in Tool Use and Workload Shape
The paper treats agentic inference as the rise of multi-step, tool-assisted patterns where users employ models as components inside broader systems rather than as single-turn text generators.
Tool-calling matters because it is a proxy for that shift. The paper finds rising adoption of tool-calling and says the implication for operators is clear: models without reliable tool formats risk falling behind in enterprise adoption and orchestration environments.
This does not prove every tool call is a full agent. A function call can be simple. But at scale, rising tool use plus longer prompts plus reasoning model adoption points in the same direction: more AI traffic is becoming workflow traffic.
Open-Weight Models Are a Real Market, Not Just a Hobbyist Alternative
The paper finds substantial adoption of open-weight models. It also argues that the open model ecosystem is highly dynamic, with usage moving as new releases appear, capabilities improve, and pricing changes.
One interesting detail is the “medium is the new small” claim. The authors classify open models as small under 15B parameters, medium from 15B to 70B, and large at 70B or more. Their usage analysis suggests small models are losing favor while medium and large models capture more value. In practice, users seem to prefer models that are still deployable and efficient, but capable enough for serious work.
That is a useful corrective to two lazy narratives. One says open-weight models are just cheap substitutes. The other says only the largest frontier systems matter. OpenRouter’s data points to a more segmented market: model choice depends on workload, cost, context, and fit.
Roleplay Is Bigger Than the Productivity Story Admits
The paper repeatedly notes the outsized popularity of creative roleplay. In the cost-versus-usage section, it says roleplay nearly rivals programming in usage volume.
That is easy to dismiss if you only care about enterprise AI, but it would be a mistake. Roleplay is evidence that consumer demand is not only about productivity. People use LLMs for companionship, entertainment, identity play, narrative generation, and persistent interaction. Those use cases create very different retention, safety, model personality, and cost dynamics than enterprise coding assistants.
The broader lesson is that AI demand is not one market. It is a set of usage clusters with different willingness to pay, context needs, quality expectations, and failure modes.
Cost and Usage Split the Market Into Workload Types
The cost-versus-usage section is especially useful because it maps categories by both volume and effective cost.
The paper describes high-cost, high-usage professional workloads such as technology and science. It says technology is dramatically more expensive than other categories, possibly because complex system design or architecture tasks require stronger and more expensive inference.
It also identifies mass-market volume drivers: roleplay and programming. Programming has the highest usage volume while maintaining a more optimized median cost. Roleplay has immense volume too, nearly rivaling programming.
Then there are lower-volume, high-cost specialized domains such as finance, academia, health, and marketing, where users may pay more because accuracy, reliability, or domain knowledge matters. Lower-cost, lower-volume utility categories include translation, legal, and trivia.
This matters because “AI cost” is not one number. Different workloads have different token shapes, quality bars, model choices, and willingness to pay.
AI Usage Is Global, But Still English-Heavy
The paper says North America remains the largest region by spend, but it accounts for less than half of total spend for much of the observed period. Europe stays relatively stable, while Asia grows sharply. Asia’s share of global spend rises from roughly 13% early in the dataset to about 31% in the most recent period.
Language distribution is still heavily English. The paper reports English at 82.87% of token volume, Simplified Chinese at 4.95%, Russian at 2.47%, Spanish at 1.43%, Thai at 1.03%, and all other languages combined at 7.25%.
For model builders, this means global demand is real but not evenly expressed. The market is still heavily shaped by English-language models, developer-heavy usage, and the availability of strong Chinese open-weight models such as DeepSeek and Qwen.
Why It Happens
The paper points to a few reinforcing mechanisms.
First, reasoning models changed what users expect. Once models can manage longer chains of logic, users start giving them harder work. That work requires more context and more state.
Second, coding and technical reasoning naturally pull in large inputs. A model asked to debug a codebase cannot reason from a tiny prompt. It needs files, traces, previous attempts, errors, dependencies, and architecture context. The same is true for document analysis, transcript synthesis, and long-running agent workflows.
Third, model routing platforms make experimentation easier. OpenRouter lowers the friction of trying different models across providers. That creates a market where users can quickly move toward models that better fit a task, whether the reason is price, latency, context window, coding ability, roleplay quality, or tool support.
Fourth, tool use turns a model call into part of a system. Once a model can call tools, the product surface changes. The model is no longer only an answer generator. It becomes a decision point inside a workflow, and that workflow has infrastructure requirements.
What This Means for Builders
Builders should treat context management as core infrastructure, not a prompt-engineering afterthought. If prompt sizes are rising faster than completions, the product is increasingly judged by how well it gathers, compresses, routes, caches, and explains context.
That creates several product implications.
Model routing becomes a first-class feature. Different workloads need different cost-quality-latency tradeoffs, and the model leaderboard changes too quickly for hardcoded assumptions.
Tool reliability becomes a distribution advantage. If tool-calling is becoming a sign of high-value workflows, then reliable tool schemas, execution traces, permissions, retries, and rollback matter as much as raw model quality.
Prompt caching and context reuse become economic infrastructure. The more a product repeatedly reads the same codebase, document corpus, or account history, the more value there is in avoiding naive full-context replay.
Usage analytics need to move beyond request counts. Builders should track prompt length, completion length, tool-use rate, model mix, retries, cancellations, latency, and cost by workflow type. Without that telemetry, AI product teams will not know whether they are improving the system or merely shifting cost into invisible context ingestion.
What This Means for Buyers and Operators
For operators, the paper is a warning against buying AI as if all usage were interchangeable. A seat-license mental model is too crude when one user might ask short writing questions while another sends 20K-token coding contexts through a reasoning model with tools.
AI procurement should ask what kind of workload is being bought:
- Is it simple generation, or reasoning over existing context?
- Does it require tools?
- Does it need a premium model, or can a cheaper open-weight model handle the task?
- Does the workflow repeat the same context often enough to benefit from caching?
- Is the value in final text, or in the intermediate reasoning and actions?
The paper also suggests why some AI products feel expensive without obviously producing more output. The visible output may be small, but the hidden input work may be large. A model that returns a concise answer after reading a codebase or document set is doing most of its work before the answer appears.
For finance and operations teams, this pushes AI cost governance toward workload-level accounting. The useful unit is not “number of prompts.” It is cost per task, cost per workflow, cost per retained user, or cost per successfully completed action.
What to Watch Next
The field should watch whether prompt growth continues, especially in programming and agentic workflows. If input tokens keep rising faster than outputs, context management will become one of the central battlegrounds in AI infrastructure.
Model builders should watch whether reasoning-model share remains above half of usage or whether cheaper non-reasoning models reclaim volume as routing improves. A stable high reasoning share would suggest that users prefer deliberate models even when they cost more.
Infrastructure providers should watch tool-calling quality, not just model availability. As tool use becomes more common, the platform that gives developers reliable execution, permissioning, traces, and cost observability may capture more value than the platform with the largest model menu.
Buyers should watch category-level economics. Programming, roleplay, technology, finance, health, and translation are not the same AI market. Their cost structures and quality requirements differ too much to manage under one generic AI budget.
Limitations and Caveats
The biggest caveat is that this is OpenRouter data. OpenRouter is large and diverse, but it is still one platform. Enterprise usage inside closed systems, local deployments, direct vendor API usage, and internal tools may follow different patterns.
The study is observational. OpenRouter usage is shaped by the models it offers, their prices, user preferences, launch effects, and routing behavior. The paper can show strong behavioral patterns, but it cannot prove that those patterns represent the entire AI ecosystem.
Several important variables are proxies. The paper infers geography from billing location. It identifies agentic inference through patterns such as multi-step or tool-invocation calls. Those are useful signals, but they are not perfect measures of the underlying behavior.
The category analysis relies on sampled, opt-in classification of about 0.25% of prompts and responses. That is still a large sample given the platform scale, but it is not the same as labeling every request.
Finally, the paper was written in a fast-moving model market. Specific model rankings will age quickly. The more durable contribution is the workload shape: longer prompts, heavier context, more reasoning-model usage, rising tool use, and segmented economics across task categories.
Source
Malika Aubakirova, Alex Atallah, Chris Clark, Justin Summerville, Anjney Midha. (2026). State of AI: An Empirical 100 Trillion Token Study with OpenRouter. arXiv preprint arXiv:2601.10088. Available at: https://arxiv.org/abs/2601.10088