Skip to content

LLM Evaluation Report

Model Date Total Response Time (s) Tests Passed Mean CodeBLEU (0-1) Mean Usefulness Score (0-4) Mean Functional Correctness Score (0-4)
gpt-5.4 2026-03-18 371.302 151 0.30016 3.85976 3.88415
gpt-5 2026-03-18 3307.15 160 0.312013 3.71951 3.82927
gpt-5-mini 2026-03-18 2223.12 161 0.305418 3.79268 3.93293
claude-opus-4-6 2026-03-18 630.643 164 0.38829 3.87195 3.90854
claude-sonnet-4-6 2026-03-18 604.89 161 0.379059 3.85366 3.90854
claude-opus-4-1 2026-03-18 635.166 157 0.349491 3.85366 3.92683
claude-sonnet-4-5 2026-03-18 546.74 162 0.331766 3.89024 3.95732
claude-haiku-4-5 2026-03-18 280.497 154 0.317284 3.84756 3.92073
gemini-3.1-pro-preview 2026-03-18 3339.78 162 0.395161 3.73171 3.82317
gemini-3.1-flash-lite-preview 2026-03-18 176.493 148 0.370935 3.77439 3.87805
gemini-3-flash-preview 2026-03-18 2146.97 142 0.395257 3.59146 3.60366
gemini-2.5-pro 2026-03-18 2788.94 118 0.373488 3.2561 3.38415
gemini-2.5-flash 2026-03-18 952.543 148 0.338621 3.7439 3.83537

Total Response Time (s): The total time taken by the model to generate all the outputs.

Tests passed: The number of unit tests that the model has passed during evaluation, out of a total of 164 tests.

Mean CodeBLEU: Average CodeBLEU score, a metric for evaluating code generation quality based on both syntactic and semantic correctness.

Mean Usefulness Score: Average rating of the model's output usefulness as rated by a LLM model.

  • 0: Snippet is not at all helpful, it is irrelevant to the problem.
  • 1: Snippet is slightly helpful, it contains information relevant to the problem, but it is easier to write the solution from scratch.
  • 2: Snippet is somewhat helpful, it requires significant changes (compared to the size of the snippet), but is still useful.
  • 3: Snippet is helpful, but needs to be slightly changed to solve the problem.
  • 4: Snippet is very helpful, it solves the problem.

Mean Functional Correctness Score: Average score of the functional correctness of the model's outputs, assessing how well the outputs meet the functional requirements, rated by a LLM model.

  • 0 (failing all possible tests): The code snippet is totally incorrect and meaningless.
  • 4 (passing all possible tests): The code snippet is totally correct and can handle all cases.