Developer builds personal LLM coding benchmark across Python, C#, and Bash to cut through hype

A software developer grew frustrated with anecdotal LLM comparisons online and built a small personal benchmark to determine which AI model best suited their actual work. The test covered 14 coding problems across Python, C#, and Bash, running three models three times each inside sandboxed Docker containers and scoring on accuracy, cost, and latency. All three models achieved 100% pass rates on the problems they attempted, but one model was blocked by content filters on four of the 14 problems, reducing its effective coverage. Cost differences were modest at roughly 1.3x between cheapest and priciest, while latency varied far more dramatically, with a 6.6x gap between the fastest and slowest models. One model performed quickly on Python and C# but took up to 168 seconds per Bash problem, highlighting that a single overall speed ranking can be misleading without breaking results down by language.