Open Access

Table 3

Average total cost per run in US dollars by LLM model and query generation method.

LLM (query generation method) Total Schema Difficulty Decomp SQL Self-correction Ratio
Claude Opus 4.6 (Step-by-step) 2.87 0.23 0.43 1.07 1.05 0.09 2.41
Claude Opus 4.6 (Direct) 1.55 0.23 - - 1.19 0.13 1.30
GPT-5.2-Codex (Step-by-step) 1.20 0.15 0.15 0.43 0.38 0.10 1.01
Gemini 2.5 Pro (Step-by-step) 1.19 0.43 0.11 0.29 0.30 0.06 1.00
Gemini 2.5 Pro (Direct) 0.85 0.43 - - 0.34 0.08 0.72
GPT-5.2-Codex (Direct) 0.82 0.15 - - 0.56 0.11 0.69
Claude Sonnet 4.5 (Step-by-step) 0.71 0.14 0.06 0.26 0.14 0.11 0.60
GPT-5.3-Codex (Step-by-step) 0.61 0.07 0.07 0.19 0.21 0.06 0.51
Claude Sonnet 4.5 (Direct) 0.39 0.14 - - 0.16 0.08 0.33
GPT-5.3-Codex (Direct) 0.34 0.07 - - 0.20 0.07 0.28
Gemini 3 Flash (Direct) 0.30 0.08 - - 0.20 0.02 0.25
Gemini 3 Flash (Step-by-step) 0.27 0.08 0.04 0.06 0.08 0.01 0.23

Notes. Values are aggregated over all queries from the test set and averaged over 10 runs. The cost is split into each stage of the query. The ratio compares the total cost against Gemini 2.5 Pro (sbs).

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.