ORM Benchmark: FastAPI-Startkit vs SQLModel vs SQLAlchemy (async)

An apples-to-apples performance benchmark of three async Python ORMs across SQLite, PostgreSQL 17, and MySQL 8.

Tested ORMs:

ORM	Type	PostgreSQL	MySQL	SQLite
FastAPI-Startkit	async	asyncpg	aiomysql	aiosqlite
SQLModel	async	asyncpg	aiomysql	aiosqlite
SQLAlchemy ORM (async)	async	asyncpg	aiomysql	aiosqlite

The FastAPI-Startkit ORM is MasoniteORM adapted for async, exposed through the public fastapi_startkit.masoniteorm API. SQLModel is a thin layer over SQLAlchemy, so it is expected to track closely with SQLAlchemy ORM (async).

Methodology

• Sequential, apples-to-apples: every ORM is run with CONCURRENTS=1, i.e. one row/query at a time with no asyncio.gather fan-out. MasoniteORM (FastAPI-Startkit) routes through a single shared connection manager and cannot safely use a concurrent model, so CONCURRENTS=1 is the only setting under which all three ORMs are directly comparable.
• Model: Test 1 — the simple model (4 fields: id, timestamp, level indexed, text indexed).
• Iterations: ITERATIONS=100 per operation.
• Environment: Apple Silicon macOS, Python 3.13.7, uvloop. PostgreSQL 17 and MySQL 8 run in Docker; SQLite is a local file (/tmp/db.sqlite3). Figures are single-run and subject to run-to-run variance at this iteration count.
• Numbers are Rows/sec (higher is better).

Operations

Code	Operation	Description
A	Insert: Single	Insert one row at a time
B	Insert: Batch	Insert many rows in a single transaction
C	Insert: Bulk	Use bulk insert operations
D	Filter: Large	Fetch a large result set
E	Filter: Small	Fetch limit 20 with random offset
F	Get	Fetch a single row by primary key
G	Filter: dict	Fetch large result set as dicts
H	Filter: tuple	Fetch large result set as tuples
I	Update: Whole	Update all fields of a row
J	Update: Partial	Update a single field
K	Delete	Delete a single row

GM = geometric mean across the available operations.

Results

SQLite

ORM	A	B	C	D	E	F	G	H	I	J	K	GM
FastAPI-Startkit	1,825	4,540	53,594	22,518	14,894	2,863	22,872	23,012	4,345	4,529	4,596	8,650
SQLModel	1,712	6,270	6,443	155,157	44,741	5,194	139,025	258,892	2,844	3,088	2,925	13,301
SQLAlchemy (async)	1,917	9,011	9,504	176,493	53,332	5,129	179,372	264,619	2,896	3,188	2,967	15,241

PostgreSQL 17

ORM	A	B	C	D	E	F	G	H	I	J	K	GM
FastAPI-Startkit	467	3,016	25,715	17,640	4,710	1,039	15,654	14,514	2,614	3,601	3,645	4,694
SQLModel	246	4,910	6,802	66,662	23,397	3,774	127,944	263,599	2,322	2,679	2,608	8,861
SQLAlchemy (async)	645	18,424	24,576	140,911	35,796	4,165	165,252	213,190	2,259	2,632	2,631	13,773

MySQL 8

ORM	A	B	C	D	E	F	G	H	I	J	K	GM
FastAPI-Startkit	247	3,221	18,362	18,049	6,894	1,002	15,358	15,492	N/A1	N/A1	2,764	4,7362
SQLModel	173	3,229	3,472	89,351	33,618	3,645	87,827	128,370	2,091	2,340	2,263	7,189
SQLAlchemy (async)	477	5,489	5,460	98,041	35,015	3,987	98,007	134,150	2,072	2,271	2,271	8,892

Analysis

• Reads (Filter: Large/dict/tuple, D/G/H) are where SQLModel and SQLAlchemy (async) pull far ahead — their thin result mapping streams large result sets an order of magnitude faster than FastAPI-Startkit on every database.
• Bulk insert (C) is FastAPI-Startkit's standout: its single multi-row INSERT reaches ~54k rows/sec on SQLite and ~26k on PostgreSQL, several times the SQLAlchemy-based ORMs, which fall back to batched inserts here.
• Single-row writes (A, F, I, J, K) are close across all three, dominated by per-statement round-trip cost under sequential execution.
• SQLModel vs SQLAlchemy (async): SQLModel tracks SQLAlchemy closely, as expected for a thin wrapper, and trails it modestly on the heaviest read and batch-insert paths (Session/validation overhead).
• On MySQL, FastAPI-Startkit's Update: Whole and Update: Partial are unavailable — see the note below.

Running the benchmark

# Clone and install
git clone https://github.com/fastapi-startkit/orm-benchmarks.git
cd orm-benchmarks
uv venv && source .venv/bin/activate
uv pip install -e .

# SQLite (default)
cd benchmarks && CONCURRENTS=1 sh bench.sh

# PostgreSQL
export DBTYPE=postgres PASSWORD=yourpassword PGPORT=5432
cd benchmarks && CONCURRENTS=1 sh bench.sh

# MySQL
export DBTYPE=mysql PASSWORD=yourpassword MYPORT=3306
cd benchmarks && CONCURRENTS=1 sh bench.sh

bench.sh runs all three ORMs across Test 1/2/3 and prints a combined table via present.py. bench.sh full and bench.sh extra raise the iteration count (1,000 and 10,000 respectively).

Footnotes

1. FastAPI-Startkit UPDATE operations are unavailable on MySQL due to an upstream MasoniteORM MySQL UPDATE-grammar bug that qualifies columns with a bogus users table (UPDATE `journal` SET `users`.`level` … WHERE `users`.`id` = … → Unknown column 'users.id' in 'where clause'). SQLite, PostgreSQL, and MySQL Delete are unaffected. ↩ ↩²

2. FastAPI-Startkit's MySQL geometric mean is computed over the 9 available operations (A–H, K), excluding the two N/A updates. ↩