CursorがSQLiteを再実装、費用の3分の2は計画役のモデルが使ったCursor reimplemented SQLite from scratch as an AI coding experiment, revealing…

Cursor, the AI-assisted code editor developed by Anysphere, has published results from an internal experiment in which it used its own tooling to reimplement SQLite from scratch. The most notable finding, according to the report, is that roughly two-thirds of the total cost went to the model responsible for planning and design rather than the model that actually wrote the code. For teams weighing the economics of large-scale AI code generation, the result is a useful data point about where money is actually spent.

SQLite is an unusually demanding target for such an exercise. It is one of the most widely deployed pieces of software in the world, embedded in browsers, mobile operating systems, and countless applications, and it is known for its extensive test suite and careful handling of edge cases. Reimplementing it is not a matter of stitching together boilerplate; it requires reasoning about storage layout, query parsing, transaction semantics, and B-tree structures. That makes it a reasonable proxy for the kind of complex, interdependent codebase where AI assistance is often assumed to struggle.

The experiment appears to have separated the work into two roles: a planning model that decomposed the problem, made architectural decisions, and sequenced tasks, and a code-generation model that produced the actual implementation. This split reflects a broader pattern in agentic coding, where a higher-level reasoning step directs a series of lower-level execution steps. The finding that planning consumed the majority of the budget suggests that architectural reasoning, not raw code output, is the dominant cost driver in projects of this scale.

There are several plausible reasons for that imbalance. Planning-oriented or reasoning models typically generate large numbers of intermediate tokens as they explore options, and reasoning tokens are often billed at premium rates. Planning also tends to be revisited repeatedly as the project encounters obstacles, whereas individual code-generation calls, though numerous, may each be comparatively cheap. The report frames the two-thirds figure as evidence that the thinking portion of an AI workflow can outweigh the typing portion, a reversal of the intuition that generating many lines of code should be the expensive part.

The result lands amid intense competition among agentic coding tools. Cursor itself is built on a fork of Visual Studio Code and has grown quickly by layering autocomplete, chat, and multi-file editing on top of frontier models from providers such as OpenAI and Anthropic. Rivals and adjacent products include GitHub Copilot and its workspace planning features, Anthropic's Claude Code, and autonomous agents such as Devin, all of which increasingly frame coding as a plan-then-execute loop rather than line-by-line completion.

Understanding cost structure matters because these tools generally pass through, or absorb, the token costs of the underlying models. If planning dominates, then optimizations such as caching plans, reusing context, choosing cheaper reasoning models, or constraining how often the design phase is re-run could have a larger effect on total spend than optimizing code generation. It also complicates simple per-line or per-feature pricing, since the expensive work is the part users do not see.

Some caveats are worth noting. The experiment is a single case study run by the maker of the tool, so the specific ratio is likely sensitive to the models chosen, the prompting strategy, and how the two phases were defined and measured. A two-thirds share for planning may not generalize to smaller tasks, to different problem domains, or to future models whose pricing and reasoning behavior differ. The exercise also does not by itself establish how correct or complete the resulting reimplementation was, which is a separate and important question for any AI-generated system software.

Even with those qualifications, the report adds concrete numbers to a discussion that is often conducted in the abstract. As AI coding moves from single-file assistance toward whole-project generation, the balance between reasoning and output, and its cost implications, is likely to become a central concern for both tool builders and the teams that adopt them. Measuring where the budget goes, rather than assuming it tracks lines of code, may prove to be the more practical lesson from the experiment.