What's new in pandas 3.1: DataFrame.select, filter with conditions, and a much faster read_csv

Nine months after the big 3.0 release, pandas 3.1 arrived on 22 October 2026. It is a smaller release, but three changes affect almost everyone who uses pandas: a new DataFrame.select method, boolean conditions in filter, and a CSV reader that is much faster and now reads large files in parallel. It also starts the removal of the inplace argument. We installed it and tested every example on this page. If you are still on pandas 2, read our pandas 3 upgrade guide first.

DataFrame.select: choose and create columns in one step

Until now, picking columns used df[["a", "b"]] and creating columns used assign. The new select does both, and works with pd.col, the expression helper added in pandas 3.0. Later columns can use ones created earlier in the same call.

import pandas as pd

df = pd.DataFrame({
    "first_name": ["John", "Alice", "Bob"],
    "last_name": ["Smith", "Cooper", "Marley"],
    "age": [61, 22, 35],
})

df.select(
    "first_name",
    age_months=pd.col("age") * 12,
    age_days=pd.col("age_months") * 30,
)
#   first_name  age_months  age_days
# 0       John         732     21960
# 1      Alice         264      7920
# 2        Bob         420     12600

filter with conditions

DataFrame.filter used to select columns or rows by label only. It now takes a cond argument: a boolean mask, a function, or a pd.col expression. Missing values in the mask count as False unless you say otherwise.

df.filter(pd.col("age") > 30)              # expression
df.filter(cond=df["age"] > 30)             # boolean mask
df.filter(lambda d: d["age"] < 40)         # function

mask = pd.array([True, None, False], dtype="boolean")
df.filter(cond=mask)            # missing value treated as False: 1 row
df.filter(cond=mask, na=True)   # missing value kept: 2 rows

One change to watch: passing a plain list of True and False values to filter now gives a warning. Use items= to select labels, or cond= for a mask.

A much faster read_csv

The default C engine was rewritten for speed, and large, uncompressed local files of 5 MB or more are now read in parallel. The result is identical to a single-threaded read. We timed a 90 MB file with 2 million rows on a small 2-core cloud machine, best of five runs:

Even on one thread, 3.1 was 3.6 times faster than 3.0, and with parallel reading it matched Polars on this file. The new mode.max_threads option limits the threads, which is useful on shared servers:

with pd.option_context("mode.max_threads", 1):
    df = pd.read_csv("large.csv")

Goodbye, inplace

The inplace argument is deprecated on methods that never really worked in place, including drop, rename, reset_index, set_index, dropna, sort_values, query and eval. In our test, pandas warned even when inplace=False was passed, so remove the argument entirely. fillna and replace keep it.

# Before
df.drop(columns="age", inplace=True)

# After
df = df.drop(columns="age")

Breaking changes to check

Smaller additions worth knowing

• pd.concat gains union_categories=True to combine categorical columns with different categories.
• Periods format in f-strings: f"{period:%Y-%m}" gives 2026-10.
• Timestamp.round, floor and ceil accept a Timedelta.
• Time strings with a space before AM or PM, such as "3:25:00 PM", now parse.
• Index.replace works like Series.replace.

How to upgrade

More on Python data tools: the pandas 3 upgrade guide, what's new in Polars 2.1, pandas vs Polars in 2026 and the pandas cheat sheet.

Sources: pandas 3.1.0 release notes, "What's new in 3.1.0" (22 October 2026), and our own tests with pandas 3.0.6 and 3.1.0, Polars 2.1.0 and PyArrow on Python 3.13. Timings are from one machine and are indicative only.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 28 October 2026. About me