diff --git a/guide/src/SUMMARY.md b/guide/src/SUMMARY.md
index 2198c3792e9..e75095f4bef 100644
--- a/guide/src/SUMMARY.md
+++ b/guide/src/SUMMARY.md
@@ -24,6 +24,7 @@
 - [Debugging](debugging.md)
 - [Features reference](features.md)
 - [Memory management](memory.md)
+- [Performance](performance.md)
 - [Advanced topics](advanced.md)
 - [Building and distribution](building_and_distribution.md)
   - [Supporting multiple Python versions](building_and_distribution/multiple_python_versions.md)
diff --git a/guide/src/performance.md b/guide/src/performance.md
new file mode 100644
index 00000000000..23fb59c4e90
--- /dev/null
+++ b/guide/src/performance.md
@@ -0,0 +1,94 @@
+# Performance
+
+To achieve the best possible performance, it is useful to be aware of several tricks and sharp edges concerning PyO3's API.
+
+## `extract` versus `downcast`
+
+Pythonic API implemented using PyO3 are often polymorphic, i.e. they will accept `&PyAny` and try to turn this into multiple more concrete types to which the requested operation is applied. This often leads to chains of calls to `extract`, e.g.
+
+```rust
+# #![allow(dead_code)]
+# use pyo3::prelude::*;
+# use pyo3::{exceptions::PyTypeError, types::PyList};
+
+fn frobnicate_list(list: &PyList) -> PyResult<&PyAny> {
+    todo!()
+}
+
+fn frobnicate_vec(vec: Vec<&PyAny>) -> PyResult<&PyAny> {
+    todo!()
+}
+
+#[pyfunction]
+fn frobnicate(value: &PyAny) -> PyResult<&PyAny> {
+    if let Ok(list) = value.extract::<&PyList>() {
+        frobnicate_list(list)
+    } else if let Ok(vec) = value.extract::<Vec<&PyAny>>() {
+        frobnicate_vec(vec)
+    } else {
+        Err(PyTypeError::new_err("Cannot frobnicate that type."))
+    }
+}
+```
+
+This suboptimal as the `FromPyObject<T>` trait requires `extract` to have a `Result<T, PyErr>` return type. For native types like `PyList`, it faster to use `downcast` (which `extract` calls internally) when the error value is ignored. This avoids the costly conversion of a `PyDowncastError` to a `PyErr` required to fulfil the `FromPyObject` contract, i.e.
+
+```rust
+# #![allow(dead_code)]
+# use pyo3::prelude::*;
+# use pyo3::{exceptions::PyTypeError, types::PyList};
+# fn frobnicate_list(list: &PyList) -> PyResult<&PyAny> { todo!() }
+# fn frobnicate_vec(vec: Vec<&PyAny>) -> PyResult<&PyAny> { todo!() }
+#
+#[pyfunction]
+fn frobnicate(value: &PyAny) -> PyResult<&PyAny> {
+    // Use `downcast` instead of `extract` as turning `PyDowncastError` into `PyErr` is quite costly.
+    if let Ok(list) = value.downcast::<PyList>() {
+        frobnicate_list(list)
+    } else if let Ok(vec) = value.extract::<Vec<&PyAny>>() {
+        frobnicate_vec(vec)
+    } else {
+        Err(PyTypeError::new_err("Cannot frobnicate that type."))
+    }
+}
+```
+
+## Access to GIL-bound reference implies access to GIL token
+
+Calling `Python::with_gil` is effectively a no-op when the GIL is already held, but checking that this is the case still has a cost. If an existing GIL token can not be accessed, for example when implementing a pre-existing trait, but a GIL-bound reference is available, this cost can be avoided by exploiting that access to GIL-bound reference gives zero-cost access to a GIL token via `PyAny::py`.
+
+For example, instead of writing
+
+```rust
+# #![allow(dead_code)]
+# use pyo3::prelude::*;
+# use pyo3::types::PyList;
+
+struct Foo(Py<PyList>);
+
+struct FooRef<'a>(&'a PyList);
+
+impl PartialEq<Foo> for FooRef<'_> {
+    fn eq(&self, other: &Foo) -> bool {
+        Python::with_gil(|py| self.0.len() == other.0.as_ref(py).len())
+    }
+}
+```
+
+use more efficient
+
+```rust
+# #![allow(dead_code)]
+# use pyo3::prelude::*;
+# use pyo3::types::PyList;
+# struct Foo(Py<PyList>);
+# struct FooRef<'a>(&'a PyList);
+#
+impl PartialEq<Foo> for FooRef<'_> {
+    fn eq(&self, other: &Foo) -> bool {
+        // Access to `&'a PyAny` implies access to `Python<'a>`.
+        let py = self.0.py();
+        self.0.len() == other.0.as_ref(py).len()
+    }
+}
+```
diff --git a/newsfragments/3304.added.md b/newsfragments/3304.added.md
new file mode 100644
index 00000000000..f3b62eba58e
--- /dev/null
+++ b/newsfragments/3304.added.md
@@ -0,0 +1 @@
+Added a "performance" section to the guide collecting performance-related tricks and problems.
diff --git a/src/lib.rs b/src/lib.rs
index 1209347ac7a..5987c666d71 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -496,6 +496,7 @@ pub mod doc_test {
         "guide/src/migration.md" => guide_migration_md,
         "guide/src/module.md" => guide_module_md,
         "guide/src/parallelism.md" => guide_parallelism_md,
+        "guide/src/performance.md" => guide_performance_md,
         "guide/src/python_from_rust.md" => guide_python_from_rust_md,
         "guide/src/python_typing_hints.md" => guide_python_typing_hints_md,
         "guide/src/rust_cpython.md" => guide_rust_cpython_md,