According to Insight Beating monitoring, Cursor enabled a group of AI Agents to only read 835 pages of the SQLite manual and rewrite a compatible database in Rust. They did not have access to the SQLite source code, program, test suite, or the internet. The new system's four formal configurations all passed a set of undisclosed SQL tests.
The formal tests consisted of four combinations. GPT-5.5 and Grok 4.5 handled planning and execution, respectively. The other two groups had Opus 4.8 or Fable 5 in charge of planning, followed by the cost-effective Composer 2.5 for execution.
The costs for the four configurations were $10,565, $1,928, $1,339, and $2,234, respectively. The lowest-cost solution involved Opus 4.8 for planning and Composer 2.5 for execution. The execution Agents consumed most of the Tokens, and switching to a cheaper model significantly reduced the costs.
Cursor also retroactively ran scenarios where Opus 4.8 and Fable 5 handled all tasks, costing $5,153 and $20,057, respectively. These two configurations only received informal scores and were not included in the quality comparison. Looking only at the expenses, the highest cost was approximately 15 times the lowest formal solution.
Assigning the intensive model to planning and using a cheaper model for execution may be more cost-effective than using the most powerful model throughout.
