Anthropic's business model is fundamentally that of an AI foundation model company — a business that trains large language models and generates revenue by providing access to those models through APIs, cloud partnerships, and consumer applications, while simultaneously pursuing safety research that is the company's primary stated purpose and its most important long-term differentiation. The API business is the largest and most strategically important revenue stream. Developers, enterprises, and researchers access Claude models through Anthropic's API at pricing that varies by model capability and token volume — Claude Haiku being the fastest and cheapest, Claude Sonnet balancing capability and cost, and Claude Opus being the most capable and most expensive. This tiered pricing structure serves different customer segments simultaneously: cost-sensitive high-volume applications use Haiku, mainstream enterprise applications use Sonnet, and premium applications requiring maximum reasoning capability use Opus. Revenue is consumption-based — customers pay per token of input and output processed — which aligns Anthropic's commercial incentives with customer usage growth. The Claude.ai consumer application — a web and mobile interface that allows anyone to interact with Claude directly, similar to ChatGPT's consumer interface — serves both as a direct consumer revenue source (through Claude Pro subscription at 20 USD per month) and as a brand-building and talent-attracting platform. Consumer adoption generates revenue at relatively low marginal cost (the infrastructure required to serve API customers also serves claude.ai users) and creates public awareness of Claude's capabilities that influences enterprise purchase decisions. The free tier of claude.ai provides a customer acquisition pathway that converts free users to paid subscribers and demonstrates Claude's quality to potential enterprise customers. The cloud platform partnerships with AWS and Google Cloud are the most commercially leveraged revenue channel. When AWS makes Claude available through Amazon Bedrock, Anthropic earns revenue proportional to usage without needing to establish individual commercial relationships with each Bedrock customer. The cloud platforms' large enterprise customer bases and existing sales relationships dramatically expand the distribution of Claude beyond what Anthropic's direct sales force can reach. These partnerships also provide infrastructure support — AWS and Google Cloud provide computing resources that are essential for running inference at scale — that reduces the capital intensity of serving growing customer demand. Enterprise direct contracts represent a third revenue channel, where Anthropic establishes direct commercial relationships with large enterprise customers seeking customized Claude deployments, priority support, higher rate limits, and compliance capabilities (HIPAA, SOC2) that are not available in the standard API. These enterprise contracts generate higher revenue per customer and provide strategic relationships with organizations that are making significant AI infrastructure investment decisions. Enterprise customers in financial services, healthcare, legal, and technology sectors are Anthropic's most commercially valuable relationships and the primary target for its enterprise sales investment. The research publication model — in which Anthropic publishes safety research, model cards, and technical papers that advance the field — is not directly revenue-generating but is commercially important in several ways. Publications establish Anthropic's credibility as a genuine safety research organization rather than merely a safety-marketing commercial company, attract the technical talent that requires working at an intellectually serious research organization, influence regulatory discussions in ways that may favor companies with demonstrated safety commitments, and build the brand reputation that supports enterprise sales to buyers who prioritize responsible AI vendors. The cost structure of an AI foundation model company is dominated by two categories: compute (the training and inference computing required to develop and deploy frontier models) and talent (the elite researchers, engineers, and operational staff required to build and run these systems). Training a frontier model requires compute investments measured in tens to hundreds of millions of dollars per training run, and the continuous advancement of model capability requires successive training runs as the company develops better architectures, training procedures, and data curation methods. Inference costs — serving existing models to API customers — scale with usage and are in principle covered by API revenue, but the thin margins of inference at scale require efficient infrastructure and optimization to remain commercially sustainable.