pandas at[] — read and write exactly one cell

What it does

df.at[row, col] reads or writes exactly one cell, addressed by label — df.at['Tokyo', 'revenue'] hands back a single scalar, not a Series. Its positional twin df.iat[2, 1] addresses the same cell by integer position. That's the entire API surface: neither accepts a list, a slice, or a boolean mask — anything beyond a scalar pair raises ('Invalid call for scalar access'). Both work on DataFrame and on Series — s.at['Vienna'] on the 'region' column is one label, one scalar — and both are accessors, not methods: no parentheses anywhere. The label fast path requires unique axes; with duplicate labels .at transparently falls back to .loc semantics (verified in example 6), a bugfix lineage that goes back to GH#33041 in pandas 1.1.0.

Why it matters

Single-cell access is everywhere in data glue code: a config value in a lookup table, one corrected sensor reading, one exchange rate in a conversion loop. Reaching for df.loc for those works, but pays for machinery you never use — masks, slices, the whole selection engine. Under pandas 3's Copy-on-Write, the honest one-cell write is df.at['row', 'col'] = x, and .iat[r, c] is the fastest way to touch a cell in a tight loop. There's one quiet perk too: assigning to a label that doesn't exist yet enlarges the frame (my go-to for 'insert one row into an object-dtype frame from a config dict'). Just remember that at is per-cell: about 9 µs per access on a 100k-row frame — a plain column lookup beats it, and vectorized beats everything.

Example

$ df.at['Tokyo', 'revenue']

np.int64(3990)

$ df.at['Berlin', 'units'] = 125   # morning stock recount
df.loc[['Berlin'], ['units']]

        units
Berlin    125

$ df['manager'] = None    # object dtype from the start
for city, m in managers.items():
    df.at[city, 'manager'] = m
df.loc[['Berlin', 'Vienna'], ['manager']]

        manager
Berlin   Müller
Vienna   Gruber

$ df.iat[2, 0] = 215   # Tokyo stock recount, by position
df.iat[2, 0]

np.int64(215)

$ df.at[['Berlin', 'Paris'], 'units']

InvalidIndexError: ['Berlin', 'Paris']

$ df.at['Prague', 'revenue'] = 770
df.tail(2)

        units  revenue region manager
Vienna   95.0   1615.0     EU  Gruber
Prague    NaN    770.0    NaN     NaN

$ d3 = df.copy()
d3.at['Paris', 'units'] = 'many'

TypeError: Invalid value 'many' for dtype 'float64'

Common flags

df.at['row', 'col']

label-based single-cell read or write — the task it exists for

df.iat[r, c]

positional single cell — pure ints, same scalar contract

df.at[row, col] = x

assignment form — passes through to the frame, no copy semantics in the way

df.at['new', 'col'] = x

unseen row label -> frame enlargement (column must already exist)

s.at['label']

one label on a Series returns that scalar — works the same

df.at[5, 'col']

label 5, NOT position 5 — on int-labeled indexes only .iat is unambiguous

df.at[mask, 'col']

ValueError 'Invalid call for scalar access' — .at is strictly scalar, masks go to .loc