Cursor had AI rewrite SQLite: 15 times different model combination costs

Cursor's experiment with AI to rewrite SQLite shows that the cost differences for different models can be 15 times higher. In the experiment, AI delegate consulted only 835 pages of the SQLite manual, rewriting the compatible database with Rust. All four official configurations of the new version of the system have passed unpublished SQL tests. The tests included gpt-5.5 and glog 4.5 for planning and execution, while the other two groups were planned by opus 4.8 or fable 5 and executed by cheap Composer 2.5. The four components, which were $10,565, $1928, $1339 and $2234 respectively, were for the minimum programme planned for opus 4.8 and implemented by Composer 2.5. The implementation of agent consumed most of the tokens, and the cost of using cheap models decreased significantly. Cursor also tested the programme for the full scope of the work of the opus 4.8 and Fable 5 missions at a cost of $5153 and 2005.7 respectively. If only cost, the maximum option would be about 15 times the minimum formal option. The use of strong models for planning and the implementation of cheap models may be more cost-effective than using the strongest models throughout the process。