Testing 12 LLMs on Bug Fixing: The Cheapest Tokens Lead to the Highest Real Cost

deusaquilus · reddit · 2026-08-05

A developer tasked 12 AI models with fixing the same slow Postgres query, providing them with an MCP server to test and verify solutions at a cost of $0.20 per tool call.

The experiment revealed that expensive models typically found the correct answer quickly with minimal calls, while cheaper models repeatedly tested non-working code. Although the cheapest tokens cost 1/135th of the priciest, factoring in tool calls brought the total cost for all models within a tight range ($1.82 to $3.51). The author highlighted Kimi K3 as uniquely cost-effective and intelligent, concluding that counting tool calls is more important than just counting tokens.

Original post →

More from coding & agent

coding & agent channel →