Performance Comparison of AI Agent, Data Analysis, and Machine Learning Workloads Across Python 3.9–3.15 Interpreter Versions

Abstract

This study installs CPython 3.9, 3.11, 3.13, 3.14, and 3.15 with the same build chain and measures the execution time of 13 workloads emulating AI agents, data analysis, and machine learning on a single Apple M3 Pro computer. For the 8 workloads that use only the standard library, the geometric-mean speedup relative to 3.9 was 1.24 in 3.11 and 1.22 in 3.15. That is, no further improvement was observed after 3.11. Between 3.14 and 3.15 using the same library versions, the differences in the NumPy, pandas, scikit-learn, and Pydantic workloads were within 2%. Explicit lazy imports in 3.15 reduced the startup time of a CLI that imports heavy modules it does not use from 222.98 ms to 20.85 ms. The experimental JIT was 1.24 times faster on pure-Python k-means but slower on startup time and CSV aggregation. Values re-measured in a separate session differed by up to 12.4%, so differences of less than 3% should be interpreted as indistinguishable.

Key Contributions

• We designed 13 workloads emulating three domains, divided into a standard-library layer and a C-extension layer.
• We reported the execution times and speedups of five versions, together with bootstrap intervals computed at the worker-process level.
• We separately measured the effects of lazy imports and the experimental JIT in 3.15, and identified major packages that cannot be installed on 3.15.

Conclusion

Standard-library workloads became 1.24 times faster in 3.11 relative to 3.9, and 3.15, at 1.22 times, was no faster than that. C-extension workloads differed by within 2% across versions when using the same library versions. The largest difference in 3.15 was lazy imports, which reduced the startup time of a CLI that does not use heavy modules by a factor of 10.69. Therefore, the rationale for upgrading to 3.15 lies in features such as lazy imports rather than in computational speed. Future work requires measurements on Linux servers, with varying input sizes, on GIL-free builds, and including real LLM calls.