Measuring the Performance of Our Models on Real-World Tasks

Source: https://openai.com/research/measuring-model-performance-on-real-world-tasks Author: OpenAI Date: 2025-09-26

Summary

OpenAI’s framework for evaluating models on real-world tasks rather than academic benchmarks. Motivated by the growing gap between benchmark performance (improving) and real-world utility (harder to measure).

Key Claims

  • Benchmark saturation: standard benchmarks (MMLU, HumanEval) are saturating — frontier models score near-ceiling, making differentiation difficult.
  • Real-world task eval: OpenAI developed 50 tasks drawn from actual user requests — coding projects, business analysis, research questions.
  • Key finding: benchmark rankings and real-world task rankings don’t always align. Models that excel on benchmarks don’t always excel on real tasks.
  • The difficulty gap: real tasks are harder than benchmarks because they’re ambiguous, multi-step, and require judgment.
  • Takeaway: model selection should be based on real-world task performance, not academic benchmarks.

Entities

Concepts

  • Coding Agents — coding tasks are central to their real-world eval
  • Reward Hacking — benchmark saturation = models gaming narrow evals