Skip to content
Nilay Kabariya

Entry No.106·Tested··5 min read

Python 3.10 end of life: 14 of 47 packages quietly downgrade

Python 3.10 hit end of life on Oct 1. I installed 47 popular packages on 3.10: 14 silently got an older release, with no error. The list, and how to spot it.

by Nilay#python#pip#uvTESTED

Verdict

Nothing fails to install on Python 3.10 yet. pip resolves to the last release that still supports 3.10, with no warning: numpy 2.2.6 instead of 2.5.4, pandas 2.3.3 instead of 3.1.1, Django 5.2 instead of 6.1, SQLAlchemy 2.0 instead of 2.1, and 10 more of the 47 I tried. You only see an error when you pin the newest version.

Tested on

Python
3.10.22 and 3.11 (GitHub Actions setup-python)
pip / uv
pip 26.3 on the base Python 3.10 and 26.2.2 on the base 3.11 control (each fresh venv used its bundled pip); uv (latest on November 9, 2026)
Packages
47 popular packages, newest release on PyPI on November 9, 2026
OS
Ubuntu (GitHub Actions)

Python 3.10 reached end of life on October 1, 2026, and the libraries were ahead of it: many popular packages already require 3.11 or 3.12 in their newest release. What surprised me is what happens when you install them on 3.10 anyway. Nothing fails. pip looks for the newest release whose Requires-Python allows 3.10, installs that without any error. So a project can sit on year-old pandas or numpy without anyone deciding that.

I installed 47 widely used packages, one at a time into a fresh 3.10 virtual environment, on November 9, 2026, and compared what I got with the newest release on PyPI:

Terminal on GitHub Actions with Python 3.10.22: numpy installs 2.2.6 instead of 2.5.4, pandas 2.3.3 instead of 3.1.1, scipy 1.15.3 instead of 1.18.1, Pillow 12.3.0 instead of 13.0.0, Django 5.2.19 instead of 6.1.3, SQLAlchemy 2.0.57 instead of 2.1.8, google-genai 2.33.0 instead of 3.0.0, torch 2.14.1 instead of 2.15.0 and more; pinning pillow==13.0.0 fails with Ignored the following versions that require a different python version.
The 14 packages that installed an older release, and the error you get only when you pin the newest one. Full log.

What 3.10 gets you instead

Package Newest on Nov 9 Installed on 3.10 Newest needs
numpy 2.5.4 2.2.6 ≥ 3.12
scipy 1.18.1 1.15.3 ≥ 3.12
Django 6.1.3 5.2.19 ≥ 3.12
pandas 3.1.1 2.3.3 ≥ 3.11
matplotlib 3.11.2 3.10.9 ≥ 3.11
Pillow 13.0.0 12.3.0 ≥ 3.11
scikit-learn 1.9.1 1.7.2 ≥ 3.11
polars 2.1.0 2.0.0 ≥ 3.11
pyarrow 26.0.1 25.0.1 ≥ 3.11
SQLAlchemy 2.1.8 2.0.57 ≥ 3.11
google-genai 3.0.0 2.33.0 ≥ 3.11
transformers 5.22.2 5.20.2 ≥ 3.11
torch 2.15.0 2.14.1 ≥ 3.11
IPython 9.18.1 8.39.0 ≥ 3.11

The full table with all 47 is in the run log. On November 9 the other 33, including requests, FastAPI, pydantic, Flask, the OpenAI and Anthropic SDKs, LangChain, pytest, ruff and cryptography, still installed their newest release on 3.10. Since then, by PyPI metadata on November 16 (not a re-run), FastAPI from 0.162.6 (November 13) and LangChain 1.5.0 (November 11) also require Python 3.11, so a re-run today would show 16 of 47.

Moving to 3.11 isn’t always enough. In the same test on Python 3.11, numpy (2.4.6), SciPy (1.17.1) and Django (5.2.19) were still held back, because their newest releases need 3.12.

Some of these gaps are a whole major version: pandas 2 to 3, Django 5 to 6, google-genai 2 to 3. If you read release notes or answers written for the current version, they may describe behaviour your 3.10 environment doesn’t have, and the other way round.

When each one dropped 3.10

The end-of-life date isn’t when the trouble started. Most of these packages dropped 3.10 months before it, so a 3.10 project has been drifting behind since 2025. From each package’s release history on PyPI:

Package First release without 3.10 Released
IPython 9.0.0 February 2025
numpy 2.3.0 June 2025
SciPy 1.16.0 June 2025
Django 6.0 (needs 3.12) December 2025
scikit-learn 1.8.0 December 2025
pandas 3.0.0 January 2026
matplotlib 3.11.0 June 2026
SQLAlchemy 2.1.0 September 2026
pyarrow 26.0.0 October 2026
Pillow 13.0.0 October 2026
polars 2.1.0 October 2026
transformers 5.21.0 October 2026
torch 2.15.0 October 2026
google-genai 2.34.0 October 2026

Two of those are easy to miss. google-genai dropped 3.10 in a minor release, 2.34.0, so a google-genai<3 pin doesn’t keep you on the newest 2.x; on 3.10 it stops at 2.33.0. And six of them dropped 3.10 in October alone, around the end-of-life date (pyarrow, Pillow, polars, transformers, torch and google-genai), so a 3.10 environment built at the start of October and one built today differ in six packages, with no change to your requirements.

The error you only see when you pin

pip only complains when you ask for a version 3.10 can’t have. Pinning the newest Pillow:

ERROR: Ignored the following versions that require a different python version: 13.0.0 Requires-Python >=3.11
ERROR: Could not find a version that satisfies the requirement pillow==13.0.0
ERROR: No matching distribution found for pillow==13.0.0

uv says it more directly, for the same pin:

error: No solution found when resolving dependencies
  cause: Because the current Python version (3.10.22) does not satisfy Python>=3.11 and pillow==13.0.0 depends on Python>=3.11, ...

So a lockfile or requirements.txt written on 3.11 or newer will fail on 3.10 with these messages, while an unpinned install on 3.10 succeeds with older versions. That difference is the usual reason “it installs on my machine” and “CI can’t resolve it” happen in the same project.

To see whether your environment is being held back, compare what you have with the newest release on PyPI: python -m pip show <package> gives your version, and https://pypi.org/pypi/<package>/json lists the newest release and its requires_python. If the newest release needs a newer Python than you run, and pip install -U doesn’t move it, your Python version is what holds it back, not your requirements. For a shared project, also check which Python CI and Docker images use: a python:3.10 base image keeps installing these older releases long after the developers’ own machines have moved on, and the two will report different versions for the same requirements file. The fix is moving the project to 3.11 or newer (3.12 if you need the newest numpy, SciPy or Django).

The run

GitHub Actions ubuntu-latest, November 9, 2026, with Python 3.10.22 from actions/setup-python (the runner’s base pip was 26.3, and 26.2.2 on the 3.11 control; each package went into a fresh venv, which uses the pip bundled with that Python, and the log does not print that pip). Each package was installed alone into a fresh venv with --only-binary :all:, and the installed version was compared with the newest release on PyPI and its Requires-Python. The same script ran on Python 3.11 as a control. The release dates come from each package’s release history on PyPI, read the same day. The script and full output are public.

— N.K., end of entry No.106

Useful? Pass it on:Post on XFollow @EmotionalMatter

Related entries

  1. No.094

    Python CERTIFICATE_VERIFY_FAILED, tested on 3.12 to 3.15

    certificate verify failed: unable to get local issuer certificate, or Basic Constraints of CA cert not marked critical on Python 3.13+. Each cause tested.

    > requests.exceptions.SSLError: HTTPSConnectionPool(host='pypi.org', port=443): Max retries exceeded with url: /simple/

    FIXED3 minWindows
  2. No.064

    Fix: Microsoft Visual C++ 14.0 or greater is required

    pip install fails on Windows asking for Microsoft C++ Build Tools. Four causes reproduced, Failed building wheel for pyyaml fixed, and when Build Tools help.

    > error: Microsoft Visual C++ 14.0 or greater is required. Get it with "Microsoft C++ Build Tools": https://visualstudio.microsoft.com/visual-cpp-build-tools/

    FIXED8 minWindows
  3. No.063

    Python 3.15 on Windows: which packages install (tested)

    Most of the stack is ready: numpy, pandas, scipy, scikit-learn, polars, pydantic, FastAPI, Django and the OpenAI and Anthropic SDKs all install on 3.15, and so do PyTorch (since torch 2.15.0) and sentencepiece (since 0.2.3). What fails on Windows: TensorFlow (no 3.15 builds at all), PyYAML and everything that needs it, such as transformers (no 3.15 wheel, so pip tries to compile it), jupyterlab, and langchain (C dependencies with no 3.15 wheels).

    TESTED11 minWindows

tip: paste the exact error text