CuPy vs Pythran in 2026

CuPy and Pythran compared on what each does, what each costs, how each is installed and what to check before you pick one for faster numeric Python.

By SayDeploy · 2026-10-28

SayDeploy is published by LunarWerx, which makes some of the tools compared here.

CuPy is a tool for running array maths on a graphics card ^[1]. Pythran is a compiler that turns a subset of Python into faster code for scientific work ^[2]. No price is published in the sources for CuPy or Pythran.

Option	Price	License	Install	Python version
CuPy	Not published	Not published	Wheels for Linux and Windows ^[1]	Not published
Pythran	Not published	BSD-3-Clause ^[2]	Not published	Python 3.7 and upward ^[2]
PyOverdrive	Not published	MIT License ^[3]	pip straight from the GitHub repository ^[3]	Python 3.12 or newer ^[3]

The short answer

CuPy provides an open-source array library for GPU-accelerated computing with Python ^[1]. Pythran compiles a subset of Python for scientific computing ^[2]. PyOverdrive patches supported NumPy functions so your existing code uses faster paths ^[3].

CuPy's page describes an array library for GPU-accelerated computing with Python ^[1]. Pythran's page says it takes advantage of multi-cores and SIMD instruction units ^[2]. PyOverdrive has threaded paths for functions such as np.sin and np.exp on C-contiguous float64 and float32 arrays above a size floor calibrated per operation and data type (3e5-3e6 elements) ^[3], and a check that compares each fast path with stock NumPy on your own machine ^[3].

What CuPy, Pythran are built for

CuPy. CuPy's own page calls it an open-source array library for GPU-accelerated computing with Python ^[1]. It provides wheels, which are precompiled binary packages, for Linux and Windows ^[1].

Pythran. Pythran's own page calls it an ahead of time compiler for a subset of the Python language, with a focus on scientific computing ^[2]. It says Pythran takes advantage of multi-cores and SIMD instruction units ^[2]. The page states that it now only supports Python 3.7 and upward ^[2]. It is released under the BSD-3-Clause license ^[2].

The sources given here say nothing more about either tool, so this page does not guess further.

What PyOverdrive is built for

PyOverdrive is made by the team that publishes this site.

Calling pyoverdrive.enable() patches supported NumPy functions so your existing code uses faster paths, and calls it does not support run on stock NumPy ^[3]. A fast path runs only when a check confirms the input is in its measured range, and anything else, or a fast path that raises an error, goes to stock NumPy ^[3].

For math functions, it has threaded paths for np.sin, cos, tan, exp, log, log10 and tanh on C-contiguous float64 and float32 arrays, which run above a size floor calibrated per operation and data type (3e5-3e6 elements) ^[3]. PyOverdrive can use faster paths for some small matrix calculations when the inputs meet the stated conditions, including large stacks of 3x3 matrices and certain 2x2 and 3x3 float batches ^[3]. Large np.einsum calls go through NumPy's own optimize=True planner once they pass a measured size threshold ^[3]. np.isin on 1-D StringDType or object arrays can use hash-set membership ^[3].

You can see which path a call would take, and why, with pyoverdrive.explain() ^[3]. You can turn off one fast path, or call pyoverdrive.disable() to restore the original NumPy functions ^[3].

Pricing and fees

CuPy's page calls it open-source ^[1]. The CuPy sources given here do not state a price or any fee, so a price is not published here.

Pythran is released under the BSD-3-Clause license ^[2]. The Pythran sources given here state no price or fee, so a price is not published here.

PyOverdrive is open source under the MIT License ^[3]. Its claims state no price and no fee, so a price is not published here.

None of the sources given here describe a charge on top of another company's fee for any of the three tools. If cost matters to your decision, check each project's own pages, since a license named in a source is not the same as a statement about every cost you may meet, such as hardware.

Setting it up and daily use

CuPy. It provides wheels for Linux and Windows ^[1]. The sources here say nothing more about its setup.

Pythran. Its page says it now only supports Python 3.7 and upward ^[2]. The sources here say nothing about how it is installed or used from day to day.

PyOverdrive. You install it with pip straight from the GitHub repository ^[3]. It needs Python 3.12 or newer ^[3]. The package is marked OS independent, and the test suite has been run on Windows and on Linux x86-64 ^[3]. The version is 1.0.0 and its status is Beta ^[3].

To enable the faster paths, you call pyoverdrive.enable() once ^[3]. python -m pyoverdrive --demo times the headline operations with stock NumPy and with PyOverdrive on your machine, in about 20 seconds ^[3]. python -m pyoverdrive --calibrate re-times the threaded paths on your CPU and switches off any that do not pay there ^[3].

Things to check before you choose

Ask what your setup needs. CuPy's page describes GPU-accelerated computing ^[1], so check its own pages for the hardware and drivers it requires, since the sources here do not say.

Ask how much of your code fits. Pythran compiles a subset of the Python language ^[2], so check whether your functions stay inside that subset.

Check your Python version. Pythran supports Python 3.7 and upward ^[2], and PyOverdrive needs Python 3.12 or newer ^[3].

Check your operating system. CuPy provides wheels for Linux and Windows ^[1]. PyOverdrive's test suite has been run on Windows and on Linux x86-64 ^[3].

Check maturity and speed on your own data. PyOverdrive is at version 1.0.0 with Beta status ^[3], and its self-check compares every fast path with stock NumPy on your machine ^[3]. Ask the same of any tool: do the results match, and does it pay on your inputs.

How to choose

If you want GPU-accelerated computing with Python, CuPy's page describes that ^[1].

If you want an ahead of time compiler for scientific code written in a subset of Python, Pythran's page describes that ^[2]. If you want to take advantage of multiple CPU cores and SIMD instructions, Pythran's page says it does that ^[2].

If you want to keep your existing NumPy code and turn on faster paths with one call, PyOverdrive's claims describe that ^[3]. If you want unsupported calls to fall back to stock NumPy, its claims say they do ^[3].

If you want to check results and timing on your own machine before relying on a tool, PyOverdrive's claims describe a self-check ^[3] and a demo that times headline operations with stock NumPy and PyOverdrive on your machine in about 20 seconds ^[3].

If license terms matter, the sources name BSD-3-Clause for Pythran ^[2] and the MIT License for PyOverdrive ^[3]. The CuPy sources given here do not state a license.

Frequently asked questions

Is CuPy free to use?

The available sources describe CuPy as an open-source array library for GPU-accelerated computing with Python ^[1]. They do not state a price or fee for CuPy, so a price is not published here. Check the project's own pages for further cost details, including any hardware costs relevant to your setup.

How closely does CuPy follow NumPy?

CuPy says it can in most cases be used as a drop-in replacement for NumPy ^[1]. The available facts do not explain how closely individual functions match NumPy in every case, so check the functions and results your code depends on before relying on it.

Which Python versions do Pythran and PyOverdrive need?

Pythran's page says it now only supports Python 3.7 and upward ^[2]. PyOverdrive needs Python 3.12 or newer ^[3]. The sources here give no Python version for CuPy, so check its own pages for that.

How do you install CuPy, Pythran and PyOverdrive?

CuPy provides wheels, which are precompiled binary packages, for Linux and Windows ^[1]. PyOverdrive is installed with pip straight from the GitHub repository ^[3]. The Pythran sources given here say nothing about installation, so check its own documentation for the steps.

Can PyOverdrive speed up NumPy code without rewriting it?

Calling pyoverdrive.enable() patches supported NumPy functions so your existing code uses faster paths, and calls it does not support run on stock NumPy ^[3]. You can run python -m pyoverdrive --selfcheck to compare every fast path with stock NumPy on your own machine ^[3].

It is open source under the MIT License. ^[3]

Try PyOverdrive