Artificial intelligence is increasingly becoming a customer of other artificial intelligence systems. Data from OpenRouter shows that AI agent token usage has surged roughly 14-fold since February 2026, rising from about 0.51 trillion tokens to 7.3 trillion tokens on the platform. During the same period, human-driven usage increased only 2.8 times, highlighting how quickly AI-to-AI interactions are becoming a major source of inference demand.
The shift marks an important change in the AI market. Instead of people directly sending prompts to models, increasingly capable AI agents are generating their own model requests as they plan, reason, call tools, write code and complete multi-step tasks. OpenRouter’s data suggests that agents have consumed more tokens than humans on the platform since February 6, 2026, although raw token growth does not translate directly into equivalent increases in AI spending because a large share of agent traffic uses cheaper cached prompts.
OpenRouter AI Usage At A Glance
| Metric | February 2026 | August 2026 | Change |
|---|---|---|---|
| Agentic token usage | ~0.51T | 7.3T | ~14x |
| Human token usage | — | 1.4T | ~2.8x since February |
| Agent usage vs human usage | Below human | >5x human | Major shift |
| Agent-human crossover | — | February 6, 2026 | Agents overtook humans |
| Agent token caching | — | Nearly 70% | Lower effective cost |
| OpenRouter model coverage | 200+ models historically | Broad multi-model platform | Multi-provider |
The figures are based on OpenRouter data cited by The Decoder and analysis from OpenRouter’s Peter Walker.
AI Agents Are Now Consuming More Tokens Than Humans
The most significant finding is that AI agents have moved ahead of humans in token consumption on OpenRouter.
According to the data, February 6, 2026 may have been the last day when humans consumed more tokens than AI agents on the platform. Since then, agentic usage has expanded dramatically, while human usage has grown much more slowly.
February 2026
Human AI Usage
████████████████████
Agentic AI Usage
██████
August 2026
Human AI Usage
████████
Agentic AI Usage
████████████████████████████████████By August, agentic systems were consuming more than five times the human-driven token volume reported in the dataset.
Agentic Token Usage Surges 14X
OpenRouter data shows agentic token consumption increasing from approximately 510 billion tokens in February to 7.3 trillion tokens by August.
That represents roughly a 14-fold increase in only about six months.
Agentic Usage Growth
| Period | Agentic Tokens |
|---|---|
| February 2026 | 0.51 trillion |
| August 2026 | 7.3 trillion |
| Increase | ~6.79 trillion |
| Growth multiple | ~14x |
By comparison, human usage grew only around 2.8x during the same period.
This creates an unusual situation in which the fastest-growing consumers of AI inference are no longer necessarily people.
What Is Agentic AI?
Agentic AI refers to systems that can perform tasks through multiple steps rather than simply responding once to a user’s prompt.
An agent may decide what it needs to do, call another AI model, access a tool, inspect the result, modify its approach and continue until the task is complete.
User Request
↓
AI Agent
↓
Plans Task
↓
Calls AI Model
↓
Uses Tools
↓
Analyzes Result
↓
Calls AI Again
↓
Validates Output
↓
Completes TaskEvery additional model call can generate more input and output tokens.
This means a single user instruction can create a much larger amount of AI inference than a conventional chatbot conversation.
One Human Prompt Can Trigger Many AI Calls
Traditional chatbot usage generally follows a simple pattern:
User → AI → Answer
Agentic systems can instead create chains of interactions:
User → Agent → Model → Tool → Model → Tool → Model → Result
That difference is central to understanding why token consumption is increasing so rapidly.
Traditional AI Vs Agentic AI
| Feature | Traditional Chatbot | Agentic System |
|---|---|---|
| User interaction | Direct | Often indirect |
| Model calls | Usually limited | Multiple |
| Planning | Limited | Extensive |
| Tool use | Limited | Common |
| Iteration | Low | High |
| Context size | Smaller | Often larger |
| Token consumption | Lower | Much higher |
| Autonomy | Low | Higher |
The growth of agentic systems is therefore changing the economics of AI inference.
Agents Can Use 13-15 Times More Tokens Per Request
Recent OpenRouter data cited in reporting indicates that agentic workloads can consume approximately 13 to 15 times more tokens per request than typical human interactions.
This helps explain how AI agents can overtake humans in total token usage even though humans remain the ultimate source of many tasks.
Typical Human Interaction
████
Agentic Interaction
████████████████████████████████████████████████An agent may consume more tokens because it has to include tool definitions, context, reasoning, intermediate results and repeated model calls.
Coding Is A Major Agentic Use Case
Software development is particularly well suited to agentic AI.
A coding agent can inspect a repository, identify a problem, modify files, run tests, analyze errors and attempt another solution.
Each step can trigger additional model inference.
Coding Task
↓
Inspect Codebase
↓
Identify Problem
↓
Generate Code
↓
Run Tests
↓
Analyze Error
↓
Modify Code
↓
Run Tests Again
↓
Final SolutionA human developer might issue a handful of prompts during this process. An agent can generate many more internal requests while performing the same task.
AI Is Becoming A Customer Of AI
The growth of agentic usage changes the traditional relationship between AI providers and customers.
Historically, AI companies primarily sold inference to humans and businesses.
Now, one AI system can become a customer of another AI model.
Human
↓
AI Agent
↓
AI Model A
↓
AI Model B
↓
AI Model C
↓
Tools / APIs
↓
Final ResultThis creates a new layer of machine-to-machine AI consumption.
The user may see only one interaction, while several AI systems could be working behind the scenes.
AI Agents Could Become The Largest Source Of Inference Demand
If the current trend continues, the majority of future AI inference could increasingly come from automated systems rather than direct human prompts.
That would fundamentally change how AI infrastructure companies plan capacity.
Potential Shift In AI Consumption
| Earlier AI Model | Emerging AI Model |
|---|---|
| Humans generate most requests | Agents generate many requests |
| Short conversations | Long-running workflows |
| One model call | Multiple model calls |
| Lower context requirements | Larger context requirements |
| Human-controlled interaction | Partly autonomous execution |
| Predictable workloads | Dynamic workloads |
This does not mean humans become irrelevant. Humans remain the source of goals and decisions in many systems. But AI can increasingly handle the intermediate work.
Token Growth Does Not Equal Equivalent Cost Growth
There is an important caveat to the 14x figure.
The Decoder reports that nearly 70% of agent token usage comes from cached prompts. Cached tokens are billed at much lower rates than fresh input tokens.
Therefore, the increase from 0.51 trillion to 7.3 trillion tokens does not mean inference costs increased 14-fold.
Token Volume Vs Effective Cost
Raw Token Usage
0.51T
↓
7.3T
↓
~14x Growth
But...
~70% Agent Tokens
Are Cached
↓
Lower Pricing
↓
Actual Cost Growth
Much Smaller Than
Token GrowthThis distinction is critical for understanding the economics of agentic AI.
Why Cached Prompts Matter
Agents often repeatedly send similar context to models.
For example, a coding agent may repeatedly provide repository information, tool definitions or instructions while working through a task.
Instead of processing the entire context from scratch each time, providers can use prompt caching.
That reduces both cost and latency.
Agentic Token Economics
| Token Type | Role | Relative Cost |
|---|---|---|
| Fresh input | New context | Higher |
| Cached input | Reused context | Much lower |
| Output | Generated response | Generally higher |
| Reasoning | Intermediate computation | Can increase usage |
The growing use of caching therefore makes agentic systems more economically viable.
OpenRouter Provides A Unique View Of AI Usage
OpenRouter operates as a multi-model AI inference platform, allowing developers to access models from multiple providers.
Its position gives the company visibility into how developers use different models and how workloads are changing.
OpenRouter’s broader research has analyzed more than 100 trillion tokens of real-world LLM interactions, providing a large dataset for studying model usage patterns.
However, OpenRouter’s data is not a complete measurement of the entire global AI market.
The platform has its own user mix and has historically had a strong presence among open-weight model users.
Open-Weight Models Could Influence The Numbers
The Decoder notes that OpenRouter’s traffic is particularly exposed to open-weight models, which can be less token-efficient than some models from OpenAI or Anthropic.
This means the precise numbers should not automatically be extrapolated to the entire AI industry.
Still, the underlying direction is important: AI agents are generating increasingly large amounts of machine-to-machine inference.
OpenRouter Data Limitations
| Factor | Implication |
|---|---|
| OpenRouter user base | Not identical to global AI users |
| Open-weight model exposure | May influence token efficiency |
| Developer-heavy platform | Could overrepresent technical workloads |
| Multi-model routing | Usage patterns differ from single-provider APIs |
| Cached tokens | Raw volume overstates effective cost |
The data is best viewed as an important indicator rather than a complete census of global AI usage.
Reasoning Models Helped Create Token Inflation
The growth of agentic AI follows another important change in AI models: reasoning models.
Traditional models generally generated an answer directly.
Reasoning models can spend additional inference computation working through a problem before producing the final response.
This increases token consumption even when the user sends only one request.
Traditional Model
Prompt
↓
Answer
Reasoning Model
Prompt
↓
Internal Reasoning
↓
More Computation
↓
Answer
Agentic System
Prompt
↓
Planning
↓
Reasoning
↓
Tool Calls
↓
More Reasoning
↓
Validation
↓
Final AnswerThe combination of reasoning and agentic workflows can therefore create substantial increases in inference demand.
AI Workloads Are Becoming More Complex
The increase in token usage is not necessarily a sign that AI systems are inefficient.
In many cases, more tokens are being consumed because the tasks themselves are becoming more complicated.
An agent that spends thousands of tokens solving a difficult engineering problem may deliver substantially more value than a chatbot that generates a short answer.
The important metric is therefore not simply tokens consumed, but value generated per unit of inference.
AI Infrastructure Demand Could Rise Rapidly
If agentic workloads continue expanding, AI infrastructure providers could face significantly greater demand for inference capacity.
The industry has historically focused heavily on training large AI models.
But as those models become widely deployed, inference can become a massive and recurring source of compute demand.
AI Model Training
↓
Model Deployment
↓
Human Usage
↓
AI Agents
↓
More Model Calls
↓
More Inference
↓
More Compute DemandThe rise of agents could therefore extend the AI infrastructure boom beyond model training.
AI Agents Could Create A New AI Consumption Loop
One particularly important development is the possibility of AI systems creating demand for other AI systems.
An agent may use one model for planning, another for coding, another for image generation and another for verification.
That creates an ecosystem in which AI providers can effectively become customers of one another.
Multi-Model Agent Architecture
| AI Component | Possible Function |
|---|---|
| Planning model | Breaks down task |
| Reasoning model | Solves complex problems |
| Coding model | Writes software |
| Vision model | Analyzes images |
| Search model | Retrieves information |
| Verification model | Checks results |
| Final model | Produces response |
This type of orchestration could become increasingly common as developers optimize AI systems for cost and performance.
Multi-Model AI Could Become The Standard
OpenRouter’s platform itself reflects this trend because developers can route workloads across different models.
Instead of choosing one AI provider for everything, an application can select a model based on the task.
For example, a cheaper model might handle simple classification while a more capable reasoning model handles complex decisions.
This approach can reduce costs while improving performance.
Example AI Model Routing
Incoming Task
↓
Task Classification
│
┌───┼────┐
↓ ↓ ↓
Easy Medium Complex
↓ ↓ ↓
Cheap Mid-tier Reasoning
Model Model Model
│ │ │
└──────┴─────────┘
↓
ResultThe rise of agentic workloads could accelerate this multi-model architecture.
The AI Economy Could Become More Recursive
Traditional software has humans using software.
AI introduces another possibility:
AI using AI to produce software, information and actions.
This creates a more recursive technology ecosystem.
Human Goal
↓
AI Agent
↓
AI Models
↓
Tools
↓
More AI Models
↓
Generated Work
↓
Human / Business OutcomeThe more capable these systems become, the less direct human interaction may be required for individual tasks.
Businesses Will Need To Track AI Unit Economics
The 14x increase in agentic token consumption also creates a warning for companies deploying AI agents.
A business may initially estimate AI costs based on the number of employees using a chatbot.
That model becomes inaccurate when one employee’s request triggers dozens or hundreds of model calls.
Traditional AI Cost Model
Users × Prompts × Average Tokens
Agentic AI Cost Model
Users × Tasks × Model Calls × Context × Reasoning × Tool Use
This is a much more complicated cost structure.
Companies will therefore need better monitoring of AI workloads and inference spending.
Agentic AI Could Change Software Business Models
The rise of machine-to-machine AI usage could also affect software companies.
Traditional SaaS products generally charge per user or seat.
Agentic systems could instead generate demand based on the number of tasks completed, actions performed or compute consumed.
That could encourage new pricing models around AI execution.
Possible AI Pricing Models
| Model | Description |
|---|---|
| Per user | Traditional SaaS approach |
| Per token | AI inference pricing |
| Per task | Charge for completed workflow |
| Per action | Charge for agent execution |
| Usage tiers | Pricing based on AI consumption |
| Outcome-based | Pricing linked to business result |
As agents become more autonomous, companies may increasingly care about the cost of completing a task rather than simply the number of AI prompts.
The Bigger Picture
The 14-fold increase in agentic token consumption on OpenRouter shows that AI is beginning to consume AI at an unprecedented scale. Agentic usage rose from about 0.51 trillion tokens in February to 7.3 trillion by August 2026, while human usage grew only 2.8 times. AI agents have consequently become the dominant token consumers on OpenRouter, with their usage exceeding human consumption by more than five times.
However, raw token growth needs to be interpreted carefully. Nearly 70% of agent token usage comes from cached prompts, meaning the effective cost increase is substantially smaller than the 14x increase in token volume. OpenRouter’s open-weight-heavy user base also means the data should not be treated as a perfect representation of the entire AI industry.
The more important development is the structural change in AI consumption. AI models are increasingly being used not only by humans but by other AI systems that plan, reason, call tools, write code and execute complex workflows. That could make agentic inference one of the most important drivers of future AI infrastructure demand.
Looking Ahead
The next phase of the AI industry may therefore be defined by AI systems using AI systems. As agents become more capable, a single human request could trigger dozens of model calls, tool interactions and reasoning cycles. This could dramatically increase the amount of inference required to complete useful work, even if the number of human prompts remains relatively stable.
For AI infrastructure companies, the opportunity is enormous but so is the challenge of controlling costs. Prompt caching, model routing and task-specific model selection will become increasingly important as businesses deploy agents at scale. OpenRouter’s latest data suggests that the transition is already underway: humans may initiate many AI tasks, but increasingly, AI itself is becoming the biggest customer of AI
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
