diff --git a/test/e2e/test-hermes-inference-switch.sh b/test/e2e/test-hermes-inference-switch.sh
index 002bb687ea0..f8dc33fd675 100755
--- a/test/e2e/test-hermes-inference-switch.sh
+++ b/test/e2e/test-hermes-inference-switch.sh
@@ -469,7 +469,11 @@ pid_before="$(hermes_gateway_pid)"
 ENV_HASH_BEFORE=$(openshell sandbox exec --name "$SANDBOX_NAME" -- sha256sum /sandbox/.hermes/.env 2>/dev/null | awk '{print $1}') || true
 
 info "Switching Hermes to ${SWITCH_PROVIDER} / ${SWITCH_MODEL} with nemohermes inference set..."
-switch_output=$(nemohermes inference set --provider "$SWITCH_PROVIDER" --model "$SWITCH_MODEL" 2>&1)
+# --no-verify skips the upfront endpoint verification (60s timeout) that is
+# redundant here: Phase 4 (check_inference_local, check_hermes_api_chat) already
+# validates live inference after the switch.  Removing the verify step avoids
+# flaky failures when the upstream model API is temporarily slow.  (#4101-nightly)
+switch_output=$(nemohermes inference set --provider "$SWITCH_PROVIDER" --model "$SWITCH_MODEL" --no-verify 2>&1)
 switch_rc=$?
 if [ "$switch_rc" -eq 0 ]; then
   pass "nemohermes inference set completed without --sandbox"