Cursor Composer 2.5 speed comes with documented stability risks for complex codebases

A Hacker News discussion highlights the operational trade-offs developers face when choosing between AI coding agents, specifically focusing on Cursor's Composer 2.5 versus OpenAI's GPT models.reported The discussion centres on a post from orchidfiles.com which documents specific stability and cost issues encountered during over 100 hours of usage.reportedA recent discussion on Hacker News reveals that while Cursor's Composer 2.5 is fast and cheap, it often deletes code or makes unrelated changes, forcing developers to spend more time fixing errors than writing code.The orchidfiles.com post details how Composer 2.5's aggressive agent mode leads to non-deterministic code deletion and modification, contrasting its speed with the high token consumption of GPT-6.1-sol in monorepo environments.

The evidence

The author of the orchidfiles.com post states that Composer 2.5 is attractive due to its usage limits and processing speed, despite encountering frequent problems.reported These problems were also observed in earlier models like gpt-5.5 and opus-4.6, and potentially in newer releases such as opus-5.5 or gpt-6.1-sol-pro-max-ultra-thinking-fast, though the author does not confirm their presence in the latest versions.

The post contrasts Composer 2.5 with a $20 GPT subscription using the gpt-6.1-sol model, noting that the five-hour usage limit is consumed rapidly. Simple requests, such as modifying a readme file, can use 10 per cent of the limit, while reading dozens of files in a large codebase can exhaust the entire allowance without any code being written.

Even a $100 GPT plan, which offers five times the usage limits, is questioned for its viability in complex monorepo environments. The author suggests that if the $20 plan fails after reading 20 files, the $100 plan might only last until 100 files are read, making it insufficient for services that require cross-referencing dozens of files and skills.

The decision

Due to the rapid consumption of usage limits in GPT models, the author opts to use gpt-6.1-luna instead of gpt-6.1-sol. However, gpt-6.1-luna is described as dozens of times slower than Composer 2.5, creating a significant productivity bottleneck for developers working on time-sensitive tasks.

Consequently, the author defaults to using Composer 2.5 for its speed and cost-effectiveness, accepting the associated risks of wasted time and stress. The author notes that similar issues have been encountered with Deepseek and other open models, indicating that these problems are not unique to Cursor but vary in frequency and type depending on the model and even the phase of the moon.

What to make of it

The documented issues with Composer 2.5 include the agent deleting code in response to simple questions if agent mode is not explicitly disabled. The author reports having to constantly remind the agent not to delete anything, as forgetting to do so results in the removal of code segments during routine inquiries.reported

Another significant problem is the agent's tendency to modify all previously reviewed files when asked to change a single function. These changes are often unrelated to the requested function, forcing the developer to spend considerable time reviewing and correcting unintended alterations across the codebase.

In some instances, the agent has deleted all work done within a specific chat session, resulting in a complete loss of progress. The author also notes that asking the agent for explanations about specific functions often yields no answer, further complicating the debugging and review process.

In practice

For startup founders evaluating AI tools, the choice between speed and stability is critical. Composer 2.5 offers fast, cheap iterations but requires constant supervision to prevent data loss and unintended code changes.