Software engineers engaged in the AI programming revolution do not exhibit strong loyalty to any single coding tool, especially when superior new options emerge. According to a source familiar with the matter, among the hundreds of engineers at popular task management software developer Notion, the growth in usage of Anthropic's Claude Code and OpenAI's Codex is surpassing that of Cursor.
This shift within Notion reflects a broader industry trend: engineers are increasingly moving away from AI tools used within their Integrated Development Environments (IDEs) towards new alternatives. Essentially, they are transitioning from tools that assist with code completion to autonomous programming agents capable of handling entire tasks independently, which can even be utilized by non-technical personnel.
Although Cursor introduced its own programming agent functionality starting last summer, its brand identity remains closely tied to IDE-assisted tools. In contrast, Claude Code and Codex are recognized for their autonomous programming capabilities, excelling at building new features from scratch or refactoring codebases.
Of course, this is just a single company's example. Engineers typically use multiple programming tools for different tasks, so they likely still use Cursor in certain scenarios, and Notion's overall spending on the tool might even be increasing. However, this phenomenon points to the direction of the software development industry.
According to the informed source, at Notion, junior engineers or those working on simpler tasks more frequently use Claude Code. This tool is adept at anticipating user intent and performs well even with imprecise task descriptions, requiring minimal context or detail.
Notion's senior engineers, meanwhile, show a preference for Codex, believing it is more proficient at handling complex, time-consuming tasks. For instance, Notion engineers often set Codex agents to run tasks overnight, sometimes for continuous periods of up to eight hours. The source also noted, however, that engineers must provide precise and rigorous instructions to Codex, as it can easily misinterpret commands and deviate from the task objectives.
Cursor is not without opportunities to catch up. The company released its latest proprietary model, Composer 2, last week, claiming it outperforms Anthropic's top-tier model, Opus 4.6, on programming tasks and at a significantly lower cost. But Cursor needs to address its reputation. Some developers discovered that Composer 2 is based on the open-source model Kimi 2.5 from the Chinese company Moonshot, sparking substantial criticism on platform X. A Cursor employee confirmed this, although developing specialized models from open-source base models is a common industry practice.
More crucially, many engineers believe that Anthropic and OpenAI are best positioned to build the optimal supporting frameworks and programming agents around their own models, given their deep understanding of the underlying technology. However, if Cursor can train a proprietary model that is better tailored for programming contexts, this advantage held by the larger labs could diminish.
In this three-way competition, Claude Code currently leads in revenue, with annualized revenue exceeding $2.5 billion. Cursor's annualized revenue is over $2 billion, and OpenAI recently stated that Codex's annualized sales have surpassed $1 billion. Despite these substantial figures, the race has only just begun.