Anything credits Claude with 800,000 user-built apps in new Anthropic case study

Anything says its users built more than 800,000 apps in five months using its Claude-powered coding agent. An Anthropic customer case study published October 30 describes a platform reaching 1.5 million users, including people without coding experience, and an agent taking on databases, payments and app-store submission rather than stopping at generated code.

The old pipeline stopped short of complete products

Anything’s earlier approach required substantial work just to keep code generation functioning. Models were prone to outages, the company says. Its team built multi-stage pipelines that called separate agents to produce code, then manually joined the results. It also maintained retrieval systems to find relevant code and software-package documentation.

That setup could produce simple, single-page applications. Building a complete product—with stored data, connected services and mobile deployment—was harder. For Anything’s nontechnical audience, generating plausible code was not enough. The agent needed to follow a long sequence of tasks, fix its own errors and preserve quality throughout a building session.

Choosing a model that could keep working

The company evaluated models on instruction-following, complex application building and repeated use of software tools. It also tested error correction over long conversations and accuracy with minimal invented information. Co-founder Marcus Lowe says Claude performed best in those evaluations; the case study does not provide comparative scores.

Anything’s architecture gives the agent visibility into both a user’s code and how the software behaves while running. That makes reliable tool use and reasoning across many steps central to the design. The team says it integrated Claude in one day using the Claude Agent SDK, a toolkit for building agents on top of Claude Code.

Reliability at scale was also important, and that allowed us to focus more on building the agent and solving the full user job rather than just making the AI function

Marcus Lowe, co-founder of Anything, in Anthropic’s case study

What users are shipping—and what success measures

The resulting agent handles backend logic—the software operating behind an app’s interface—as well as databases, mobile deployment, App Store submission and payments. Anything says users can now build multi-page products with external software connections, AI features and paywalls. The case study offers several examples of that broader scope:

• PipelinePro: A nontechnical founder built a recruiting and client-intelligence platform with candidate scoring, AI interview bots and a client portal. He is demonstrating and selling it to customers.
• Telefan: A French film director launched a streaming app for discovering hard-to-find French TV films, including discovery by voice.
• The Collective: A user without an engineering background published a social app for making friends and meeting people around Philadelphia.

Anything reports a 91–96% agent success rate for complex builds. It measures that using generation completion rates, agent self-reports and in-product user feedback. Those measures describe completion and reported outcomes, not an independent assessment of every app’s quality. The usage figures and performance results are company claims published by Anything’s AI supplier.

The next step remains a development goal. Anything is working toward agents that can tackle multiple parts of a product simultaneously and run for longer periods. Lowe expects user-built apps eventually to improve themselves as agents handle more complexity. The case study presents that as the team’s direction, not a demonstrated capability of the current system.