Verified Claude Builds Gallery

What qualifies as a build

A build is included when the public record identifies what was made, who made or deployed it, what Claude surface was involved, and what evidence supports the claim. That evidence can be an official customer story, a public repository, an official product document, or a builder post with enough detail to evaluate.

The gallery intentionally mixes enterprise systems and public projects. Spotify, Ramp, and Rakuten show how large organizations operationalize Claude Code. Zettair shows an individual expert steering an agent through a complex product. Public repositories show the tooling layer that has grown around the ecosystem.

• Included: case studies, public projects, templates, integrations, build systems, and evidence-backed demos.
• Excluded: unsourced social claims, prompt-only posts, speculative roadmaps, and duplicate tutorial content.
• Linked out: step-by-step implementation guidance belongs on sibling sites in the Claude Network.

The missing lens

The useful question is not whether Claude Code can produce code. The useful question is what kind of system the human built around it: a local feedback loop, a CI workflow, a fleet migration system, an MCP-connected operations loop, or a template that packages team norms.

Each project card names that pattern so readers can compare builds across different levels of maturity. A public demo and a production migration agent should not be judged by the same evidence bar, but both can teach something concrete.

Which model and surface built it, as of September 2026

Every card now carries a surface line, and that line is often an admission. The enterprise case studies on this site were published before the current lineup existed, and none of them name a model generation. Their numbers describe Claude Code and Claude Agent SDK as those products worked at the time, so they should not be read as Opus 5 or Fable 5.1 results.

The current generally available lineup is Claude Fable 5.1 for the most demanding long-horizon agentic work, Claude Opus 5 as the recommended starting point for most workloads, Claude Sonnet 5 for the speed and intelligence balance, and Claude Haiku 4.5 as the fastest and cheapest. Anthropic notes that Fable 5.1 defaults to High effort inside Claude Code and Medium on Claude.ai and in Claude Cowork, which is enough on its own to make two runs of the same prompt incomparable across surfaces.

The surface list has also grown past the terminal. A build can now come out of Claude Code, Claude Code on the web, the Claude Agent SDK, Claude in Chrome, or Claude Cowork, and each one leaves a different kind of evidence behind. For the full model comparison and pricing question, see claudecentral.com; this page only cares about attributing a build correctly.

• Record the surface and the model generation when the source states them, and say so plainly when it does not.
• Treat undated adoption metrics as claims about a product era, not about a specific model.
• Effort settings differ per surface, so a Claude Code result and a Cowork result are not the same experiment.