Are 70% of projects destined to be canceled? Behind Claude Code's success, Anthropic has cultivated a team of "losers."

Behind every successful Claude Code project, there may be many more projects that have already been cancelled.

According to Business Insider, Anthropic has an internal "Labs" team of fewer than 20 people dedicated to this task: constantly developing new products, rapidly validating them, and then cutting most of the failed attempts. According to Ben Mann, the team leader and co-founder of Anthropic, the success rate of the Labs projects is only about 20% to 30%, meaning that approximately 70% to 80% of the "bets" ultimately fail.

But Anthropic doesn't view failure as a waste. On the contrary, this high attrition rate is becoming a crucial engine for the company to rapidly transform cutting-edge model capabilities into commercial products. Products such as Claude Code, Model Context Protocol (MCP), and Claude Design are all related to Labs. As Anthropic prepares for its IPO, the importance of this "failure" team is also rising.

In an interview, Mann stated, "We do see it as a true startup incubator and expect most of our bets to fail." In Anthropic's view, the key is not to make every project successful, but to use a sufficiently low cost of trial and error to achieve a few "hit" opportunities that are enough to change the business curve.

A two-week "survival or relocation" cycle: Most projects are destined to be canceled.

Labs operates in a way that is quite different from traditional product teams. The approximately 20 members work in rotation, continuously generating new product prototypes, which is internally referred to as "bets." Every two weeks or so, the team conducts a "persevere or pivot" review: poorly performing projects are either canceled or merged into other projects, and the members then move on to new attempts; a few validated projects "graduate" from Labs and form their own independent teams.

Claude Code was one of the most successful "bets." Mann initially assigned the task to Boris Cherny, who had recently joined the company. Cherny initially wanted to create a code analysis tool, but Mann felt that this direction was not bold enough and asked him to aim higher. Subsequently, the prototype developed by Cherny quickly became popular within Anthropic and was released to the public as a preview version in February 2025.

Subsequently, as the Claude model continued to upgrade, Claude Code gradually grew from an internal prototype into one of Anthropic's most influential products, and also became an important catalyst for the company's accelerated growth.

Labs' ability to conduct frequent trial and error has another key advantage: it is close enough to the forefront of research. Mann said that before the development of Claude Code started, the team learned from researchers about the potential of the new model in agent programming. Therefore, they were able to plan ahead before the model's capabilities matured and bring the research results to the product stage more quickly.

Product portfolio continues to expand: from programming tools to design software

Claude Code is not Labs' only achievement. The team also incubated the Model Context Protocol (MCP), used to connect AI agents with external data sources; Claude Design, launched this year, also comes from Labs. Instagram co-founder Mike Krieger is also on this team.

Mann defines Labs' goal as expanding the "action space" of AI, that is, continuously pushing the boundaries of tasks that AI models can truly accomplish. Therefore, Labs does more than just find the next blockbuster product; it also exposes the shortcomings of the model's capabilities in product practice and then pushes the research team to fill those gaps.

For example, Labs helped the Anthropic research team improve the audio model's understanding of Ethiopian Amharic; the development of Claude Design further stimulated the company's research interest in visual output capabilities. Product trial and error thus became not only a way to validate business needs, but also a path to drive the evolution of model capabilities.

This mechanism is now beginning to spread beyond Labs. Mann revealed that other product engineering teams at Anthropic have also established similar "betting" groups.

The challenging balance before and after an IPO: daring to fail while also facing competition head-on.

With Anthropic preparing for its IPO, Labs' problem is no longer just "can we create the next Claude Code", but rather how to find a balance between continuous innovation and commercialization.

Forrester analyst Mike Gualtieri points out that Anthropic is walking a tightrope: the company's constantly emerging new tools may directly penetrate the market of existing enterprise software. For example, Claude Design's product capabilities have already begun to compete with design software such as Adobe and Figma.

Gualtieri believes that as Anthropic gradually builds its own product platform, the company may create features that would allow users to reduce their reliance on other software. "This will be a delicate game."

At the same time, Labs itself must maintain a sufficiently high turnover rate. Projects typically "graduate" and become independent once the team size exceeds four people, with members constantly moving to other departments while new people continuously join. This means that the core of Labs is not a fixed team of 20 people, but a mechanism that continuously creates, filters, and amplifies product opportunities.

Looking further ahead, Mann hopes this mechanism can extend from software products to hard science fields such as biological research and clean energy storage, allowing AI to truly enter the broader real world. He believes that the ability to accelerate the development of hard sciences is "extremely exciting" and is also key to AI beginning to create greater real-world value.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.