The persevere-or-pivot review reads as a judgment on ideas, and its more demanding function is logistical. When a bet is killed, the people on it move to another task, and dead prototypes are sometimes merged into surviving ones: the Claude Chrome extension picked up work from ideas that did not stand on their own [15]. At two-week intervals that is roughly 26 decision points a year [18], a cadence that only holds if there is somewhere for freed engineers to land and no career cost in landing there [4].
Mann puts the hit rate at 20 to 30 percent [5], which means 70 to 80 percent of bets are ended or absorbed [17]. He also says the team has had more hits than he expected, and that the hits went far [16]. The record lacks a denominator: a rate quoted without a count of bets cannot be turned into a cost per hit or an expected value, so the honest reading is that the pattern produced at least three shipped things from about 20 people, not that its economics have been demonstrated [21].
The obvious comparison is to Area 120 with better models. Mann does not dispute the lineage: he cites Bell Labs as a model, worked on Area 120 at Google in 2018, and Google's X helped build Waymo and Google Brain [9]. What separates Labs is early sight of research, not the structure: Mann's developers hear what is going well inside Anthropic, and heard that new models were showing strength at agentic coding well before Claude Code was built in late 2024 [8]. Labs was created in 2024 [2] and the preview shipped in February 2025 [7], so the team's largest hit landed within roughly a year of the team existing [19]. That schedule owes more to knowing what the models could suddenly do than to any ideation ritual.
The board-deck version of this is a 20-person unit with two-week gates, and that deck hides the authority underneath it. Mann first had to convince Anthropic that it should release products at all rather than ride philanthropic funding toward advanced AI [10], and Labs now pushes other teams' researchers toward capabilities it wants, including better audio understanding of Amharic and the visual outputs behind Claude Design [20]. It also told a new engineer his proposed code analysis tool was not a big enough idea and gave him a larger brief instead [6]. Two-week gates cost nothing; the scarce resource is a sponsor who can redirect research and overrule a new hire's scope.
The near-term pressure comes from the listing. Anthropic is preparing to go public, with Labs expected to keep converting frontier models into products [11], and a write-off rate that internal sponsors treat as the price of the method looks different in a shareholder report. This quarter's decision for anyone borrowing the pattern is whether the sponsor can absorb that ratio in public view; the test over the next several years is whether Labs graduates a second product at Claude Code's scale [3].
