Anthropic's Labs Team: 70% of Projects Fail, but Drive Key Innovations

Within Anthropic, a Labs team of fewer than 20 members is dedicated to rapidly prototyping and validating new product ideas. According to team lead Ben Mann, only about 20% to 30% of projects succeed, with 70% to 80% ultimately being discontinued. This team, regarded as a startup incubator, has spawned products such as Claude Code, the Model Context Protocol (MCP), and Claude Design—among them, Claude Code has become one of Anthropic’s most impactful products. As the company prepares for its IPO, the importance of this “failure-prone” team is growing, and its methodology is being adopted by other product teams, with potential future expansion into hard science fields such as biological research and clean energy storage.

Article author, source: Business Insider

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

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

But Anthropic does not view failure as a waste. On the contrary, this high-failure-rate mechanism is becoming a key engine for rapidly transforming cutting-edge model capabilities into commercial products. Products such as Claude Code, the Model Context Protocol (MCP), and Claude Design are all connected to Labs. As Anthropic prepares for its IPO, the importance of this "failure" team is growing.

Mann said in an interview: "We truly view it as a startup incubator and expect most bets to fail." From Anthropic’s perspective, the key isn’t for every project to succeed, but rather to achieve a few transformative "hits" at a sufficiently low cost of experimentation.

Labs operates very differently from traditional product teams. Around 20 members work on a rotating basis, continuously developing new product prototypes, internally referred to as "bets." Approximately every two weeks, the team conducts a "persevere or pivot" review: underperforming projects are either shut down or merged into other initiatives, and members immediately move on to new attempts; a small number of validated projects "graduate" from Labs and form 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 originally intended to build a code analysis tool, but Mann felt this direction wasn’t bold enough and urged him to aim higher. The prototype Cherny developed quickly gained popularity within Anthropic and was released publicly in preview form in February 2025.

Since then, as the Claude models have continued to evolve, Claude Code has grown from an internal prototype into one of Anthropic’s most impactful products and a key catalyst for the company’s accelerated growth.

Labs has another key advantage: it is close enough to the frontiers of research. Mann said that before the Claude Code project began, the team learned from researchers about the potential of new models in agent programming, allowing them to strategically position themselves ahead of time and accelerate the translation of research findings into products, even before the model capabilities were fully mature.

Claude Code is not the only achievement from Labs. The team has also developed the Model Context Protocol (MCP) to connect AI agents with external data sources; Claude Design, launched this year, also originated from Labs. Mike Krieger, co-founder of Instagram, is also part of this team.

Mann defined Labs' goal as expanding the "action space" of AI—continuously pushing the boundaries of what AI models can actually accomplish. Therefore, Labs does more than just seek the next breakout product; it also identifies gaps in model capabilities through product implementation and drives the research team to address them.

For example, Labs previously motivated Anthropic’s research team to improve the audio model’s understanding of Amharic, a language spoken in Ethiopia; the development of Claude Design further sparked the company’s interest in enhancing its visual output capabilities. As a result, product experimentation is no longer merely about validating commercial needs—it has also become a pathway for reverse-engineering 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 "bets" groups.

As Anthropic prepares for its IPO, the question Labs must address is no longer just “Can we build the next Claude Code?”, but how to balance ongoing innovation with commercialization.

Forrester analyst Mike Gualtieri noted that Anthropic is walking a "high wire": the company’s newly released tools are beginning to directly compete with existing software in enterprise customers’ stacks. For example, Claude Design’s capabilities are now rivaling design tools like Adobe and Figma.

Gualtieri believes that as Anthropic gradually builds its own product platform, the company may develop features sufficient to reduce users' reliance on other software. "This will be a delicate balancing act."

Meanwhile, Labs itself must maintain a high level of personnel turnover. Projects typically "graduate" and spin off independently once their teams grow beyond four people, with members continuously moving to other departments while new talent continually joins. This means that Labs’ core is not a fixed team of 20 people, but rather a system designed to continuously generate, filter, and scale 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, enabling AI to truly enter a broader range of real-world applications. He believes that the ability to accelerate progress in hard sciences is "extremely exciting" and key to AI beginning to generate greater real-world value.