diff --git a/tests/benchmark-models/test_petab_benchmark.py b/tests/benchmark-models/test_petab_benchmark.py
index 109839f0f3..4892100877 100644
--- a/tests/benchmark-models/test_petab_benchmark.py
+++ b/tests/benchmark-models/test_petab_benchmark.py
@@ -53,6 +53,10 @@ class GradientCheckSettings:
     # Absolute and relative tolerances for finite difference gradient checks.
     atol_check: float = 1e-3
     rtol_check: float = 1e-2
+    # Absolute and relative tolerances for fiddy consistency check between
+    # forward/backward/central differences.
+    atol_consistency: float = 1e-5
+    rtol_consistency: float = 1e-1
     # Step sizes for finite difference gradient checks.
     step_sizes = [
         1e-1,
@@ -91,6 +95,7 @@ class GradientCheckSettings:
     atol_sim=1e-14,
     rtol_sim=1e-14,
     noise_level=0.01,
+    atol_consistency=1e-3,
 )
 settings["Okuonghae_ChaosSolitonsFractals2020"].step_sizes.extend([0.2, 0.005])
 settings["Oliveira_NatCommun2021"] = GradientCheckSettings(
@@ -234,7 +239,10 @@ def test_benchmark_gradient(model, scale, sensitivity_method, request):
         sizes=cur_settings.step_sizes,
         direction_ids=parameter_ids,
         method_ids=[MethodId.CENTRAL, MethodId.FORWARD, MethodId.BACKWARD],
-        success_checker=Consistency(atol=1e-5, rtol=1e-1),
+        success_checker=Consistency(
+            rtol=cur_settings.rtol_consistency,
+            atol=cur_settings.atol_consistency,
+        ),
         expected_result=expected_derivative,
         relative_sizes=not scale,
     )
@@ -287,6 +295,8 @@ def assert_gradient_check_success(
     df = check_result.df
     df["abs_diff"] = np.abs(df["expectation"] - df["test"])
     df["rel_diff"] = df["abs_diff"] / np.abs(df["expectation"])
+    df["atol_success"] = df["abs_diff"] <= atol
+    df["rtol_success"] = df["rel_diff"] <= rtol
     max_adiff = df["abs_diff"].max()
     max_rdiff = df["rel_diff"].max()
     with pd.option_context("display.max_columns", None, "display.width", None):