aiaimodel.cards
← All benchmarks

MGSM

We haven't written a description for this benchmark yet; see the card notes below for how each maker ran it.

Your workloadRAG over docs · 12K in · 600 out · 50K requests/mo · 30% cached
  1. 1Llama 4 Maverick · Metabest value
    92.3%
  2. 2Llama 3.3 70B · Metabest value
    91.1%
  3. 3Llama 4 Scout · Metabest value
    90.6%
  4. 4GPT-4.1 Mini · OpenAI
    88.2%
  5. 5GPT-4o Mini · OpenAI
    87.0%
  6. 6GPT-4.1 · OpenAI
    86.9%

“Best value”: no other model here reports a higher score at a lower cost for your workload. Scores are self-reported by each maker with different settings, so treat close calls as ties.

How each maker reported it

  • Llama 4 Maverick Benchmarks: Instruction tuned models

    Meta's own 0-shot results on bf16 weights. Llama 3.x models have no image support and a 128K context, so those cells are not reported.

  • Llama 3.3 70B Benchmarks: Instruction tuned models

    Meta's own results with its internal evaluation library; shots and metrics per Meta's table.

  • Llama 4 Scout Benchmarks: Instruction tuned models

    Meta's own 0-shot results on bf16 weights. Llama 3.x models have no image support and a 128K context, so those cells are not reported.

  • GPT-4.1 Mini openai/simple-evals README

    OpenAI's simple-evals results (zero-shot chain-of-thought, system prompt "You are a helpful assistant."). MATH is MATH-500 for models released on or after o1; OpenAI considers MGSM and DROP saturated for newer models. The repo stopped adding results in July 2025.

  • GPT-4o Mini openai/simple-evals README; GPT-4o System Card Tables 8–10, p.24–25

    OpenAI's simple-evals results (zero-shot chain-of-thought, system prompt "You are a helpful assistant."). MATH is MATH-500 for models released on or after o1; OpenAI considers MGSM and DROP saturated for newer models. The repo stopped adding results in July 2025. Language rows are 0-shot accuracy from the GPT-4o System Card (translated ARC-Easy, translated TruthfulQA, Uhura-Eval). GPT-4o values: simple-evals uses the gpt-4o-2024-05-13 snapshot; the system card does not name a snapshot.

  • GPT-4.1 openai/simple-evals README; Prompting Guide §1

    OpenAI's simple-evals results (zero-shot chain-of-thought, system prompt "You are a helpful assistant."). MATH is MATH-500 for models released on or after o1; OpenAI considers MGSM and DROP saturated for newer models. The repo stopped adding results in July 2025. SWE-bench Verified is OpenAI's own agentic harness result from the prompting guide.

Only each model's own maker's number is shown; scores a maker reports for rival models are left out.