OpenRouter data

Methodological note

Background information

Artificial intelligence (AI) models, and large language models (LLMs) in particular, can be accessed programmatically through application programming interfaces (APIs). An API is a standardised interface that allows one software application to communicate with another. In the case of an AI model, an API allows an application to send a request (for example, a prompt or other input) to an AI model and receive a response without the user interacting directly with the model through a web interface.

OpenRouter is a platform that provides a unified API gateway for accessing a large number of AI models from different developers and providers. Rather than requiring developers to establish and maintain separate integrations for each AI model, OpenRouter provides a common interface through which multiple models can be accessed. The service is widely used, with API requests coming from over 200 countries and territories.

Coverage

The dataset covers AI model requests made through OpenRouter. These data underpin the visualisations on OECD.AI. The analysis is based on the requests and usage information available through OpenRouter and therefore does not necessarily represent the full extent of global AI model usage.

OpenRouter data do not include usage through direct provider APIs, consumer-facing chatbots or private enterprise deployments. As a result, the dataset may under-represent general-purpose consumer use and other forms of AI adoption occurring outside the OpenRouter ecosystem.

Update frequency

OECD.AI visualisations are based on monthly observations and are updated quarterly.

Dimensions

The dataset contains several dimensions used to disaggregate the indicators.

  • Developer country/territory: The “developer” refers to the firm that pre-trains and fine-tunes an AI model. Country assignments are based on where a model developer’s lab or organisation is headquartered. 
  • Requester country/territory: The “requester” is the unique OpenRouter user and their location is based on where the API call originated, geographically.
  • Model weights: Models are classified as open or closed weights based on classifications provided by OpenRouter. Open weights indicate that the model’s parameters – the numerical values set during training that determine how the model generates its outputs – can be downloaded and run or fined-tuned by third parties, subject to the terms of the accompanying license.  
  • Requester type: Requests are classified as coming from Agentic, Human or Mixed sources. Agentic requests are initiated by an AI system rather than directly by a person,, often as one step in a longer automated workflow. Human requests are initiated directly by a human, typically through an application or interface. Mixed requests arise when OpenRouter API keys have instances where both agentic and human interactions were identified. Classification relies on signals observable in API traffic and does not capture user intent.
  • Month: The month in which the API request occurred on OpenRouter.

Usage and cost

Model usage is measured using tokens and costs recorded in the dataset. Tokens are units of information that AI models process. For text, a token can represent a word, part of a word, punctuation, or other pieces of text. Depending on the model and input, tokens can also represent other types of information, such as images or audio. In the visualisations, token volumes are expressed per million tokens.

Different types of tokens are represented in the dataset:

  • Prompt tokens: These are tokens sent to the model by the requester as input. Prompt tokens also include cached tokens, which are tokens that are reused from previous processing rather than being processed again from scratch. They generally require less inference and therefore incur lower costs.
  • Completion tokens: These are generated by the model in response to a prompt. They represent the output of the inference. Completion tokens also include reasoning tokens, which are tokens generated and used by the model as part of its reasoning or inference process, between receiving the prompt and producing the completion.
  • Total tokens: This is the sum of prompt and completion tokens for a specific grouping.

The average cost of requests to AI models is calculated from aggregated cost data and total token usage.

Average cost (per million tokens) = total cost / (total tokens / 1 000 000)

This expresses the average upstream usage cost per one million tokens. To ensure that the aggregate indicator is correctly weighted, total cost and total tokens are summed first at the relevant aggregation level, and the average cost is then calculated.

Industries

Industries in the source dataset come from self-declaration of OpenRouter users. As a result, the original dimension had over 10 000 industry labels. To reduce the volume, a multi-step classification procedure was performed.

  • Step 1: For consistency with existing OECD.AI databases and taxonomies, the industrial grouping from Preqin were manually adjusted to best capture the types of industries present in the OpenRouter source data.
  • Step 2: The list of industries was passed to GPT-5, with the following prompt:


Using the attached file, you are classifying an “Original Industry” label into exactly one of the 20 Consolidated Categories below. Read all 20 definitions before choosing. Where a name could fit more than one category, prefer the more specific category over a general/default one (e.g. Digital security beats Non-AI Software; AI, Data & Emerging Technology beats Non-AI Software; Mobility and autonomous vehicles beats Logistics, wholesale and retail).

  1. IT infrastructure and hardware — Companies that build or operate the physical and network backbone of computing, such as data centers, hosting, cloud infrastructure, networking equipment, servers, and computer/IT hardware manufacturing (not consumer electronics, which go to Consumer products and manufacturing).
  2. Mobility and autonomous vehicles — Companies focused on vehicles, transportation technology, or autonomous/self-driving systems as their core product, including automotive manufacturing, EVs, and AV software/hardware (routine vehicle sales, rental, or repair go to Consumer services or Logistics, wholesale and retail instead).
  3. Other — Use only when the industry name is legitimate and specific but does not reasonably fit any other category on this list.
  4. Business processes and support services — Companies that provide back-office, administrative, staffing, HR, outsourcing, facilities, or general B2B consulting services to other businesses (excluding legal, financial, or IT-specific consulting, which have their own categories).
  5. Healthcare, drugs and biotechnology — Companies delivering medical care or wellness services, or developing/manufacturing pharmaceuticals, medical devices, or biotechnology products.
  6. Financial, legal and insurance services — Companies providing banking, investment, accounting, insurance, or legal/law-firm services (law enforcement and courts go to Government, Nonprofit and defence instead).
  7. Media, social platforms and marketing — Companies producing media, entertainment, creative content, journalism, advertising, marketing, or social/content platforms as their core offering.
  8. Digital security — Companies whose core offering is cybersecurity, information security, or digital/data protection (e.g. cyber, infosec, encryption, firewall, penetration testing, threat intelligence, identity & access management); physical guard/alarm/surveillance services go to Business processes and support services instead.
  9. Logistics, wholesale and retail — Companies moving, storing, distributing, or selling goods to other businesses or consumers, including shipping, freight, warehousing, wholesale trade, and general retail.
  10. Education and training — Companies whose core offering is teaching, instruction, or training, at any level, in any subject (including subject-specific training like financial, medical, or industry education) as long as delivering that instruction is the actual product being sold.
  11. Energy, raw materials and utilities — Companies producing, distributing, or trading energy, fuel, utilities, or extracted raw materials (mining, oil & gas, power, water).
  12. Government, Nonprofit and defence — Government agencies, public-sector bodies, nonprofit/NGO organizations, national defense, law enforcement, judiciary/courts, and homeland/national security programs.
  13. Consumer products and manufacturing — Companies that manufacture or sell physical consumer goods not covered by a more specific category (electronics, apparel, household goods, general industrial manufacturing).
  14. Consumer services — Companies providing personal or local services directly to individual consumers (repair, personal care, non-lodging leisure services) that aren’t travel/hospitality or business-facing.
  15. Travel, leisure and hospitality — Companies in lodging, tourism, travel booking, dining out, entertainment venues, or recreational leisure activities.
  16. Agriculture, food and beverages — Companies engaged in farming, food or beverage production, processing, or manufacturing.
  17. Construction and real estate services — Companies that build, develop, manage, or broker physical buildings and property, including architecture and building-trade services.
  18. Non-AI Software & Digital Technology — Companies whose core offering is software, apps, IT services, or digital platforms, and that do not fit a more specific technology category (Digital security or AI, Data & Emerging Technology).
  19. AI, Data & Emerging Technology — Companies whose core offering is artificial intelligence, machine learning, data science/analytics, or other frontier/emerging technologies (blockchain, quantum, robotics R&D).
  20. Unspecified — The input is blank, “N/A”, “Unspecified”, or otherwise contains no usable information about the industry.

Output format: Return a csv, with the original industry and the category name exactly as written above, with no explanation.

  • Step 3: A second prompt is passed to Sonnet 5:


You are auditing a dataset of industry classifications. You will receive rows with two columns: Original Industry (a free-text industry label) and Consolidated Industry (the category it was assigned to).

Judge each row using your own independent, standard industry classification logic; do not assume the existing assignment is correct, and do not try to reverse-engineer any rules the original classifier may have used.

The only valid categories are these 20 (use these exact names in your output, nothing else): IT infrastructure and hardware; Mobility and autonomous vehicles; Other; Business processes and support services; Healthcare, drugs and biotechnology; Financial, legal and insurance services; Media, social platforms and marketing; Digital security; Logistics, wholesale and retail; Education and training; Energy, raw materials and utilities; Government, Nonprofit and defence; Consumer products and manufacturing; Consumer services; Travel, leisure and hospitality; Agriculture, food and beverages; Construction and real estate services; Non-AI Software & Digital Technology; AI, Data & Emerging Technology; Unspecified

Two known, deliberate exceptions. Do not flag these as errors, since they reflect an intentional choice, not an oversight:

  • Subject-specific training (e.g. “Financial Education,” “Real Estate Education”) is intentionally classified as Education and training, not by its subject matter.
  • Any industry naming a cybersecurity-specific term (cyber, infosec, encryption, penetration testing, threat intelligence, etc.) is intentionally classified as Digital security, even if “software” or “IT” also appears in the name.

For every other row, determine whether the classification is Correct, Misclassified, or Uncertain.

Only include rows that are Misclassified or Uncertain in your output; omit rows you judge Correct.

Output as CSV with these exact columns: Original Industry, Consolidated Industry, Suggested Category, Reason. Reproduce the first two columns exactly as given. Suggested Category must be one of the 20 category names above. Reason is one sentence explaining the logic for the change. Wrap any field containing a comma in double quotes so the CSV parses correctly.

Process every row in the input (do not skip or truncate). If the input is large, work through it systematically from top to bottom and do not stop until every row has been evaluated.

  • Step 4: Human judgement was applied to the audit, either accepting and applying or rejecting the suggested changes.

For updates and new industries not already captured, a similar procedure will be conducted, using the data from the existing categorisation to inform the LLM, or direct human classification.