diff --git a/README.md b/README.md
index 20e0e2e..f270cd6 100644
--- a/README.md
+++ b/README.md
@@ -126,6 +126,15 @@ studio --path "$SITE_PATH" wp option update \
 	connectors_ai_anthropic_api_key "$ANTHROPIC_API_KEY"
 ```
 
+> **Cost note.** The bundled demo agents use `model: auto`, so the active
+> provider picks the model — which on a cloud provider is its *default*, often
+> the priciest (e.g. Opus). To control spend, set each agent's
+> `default_config['model']` to a specific id the provider exposes (e.g. a dated
+> Anthropic Haiku id). If a configured model id doesn't resolve, wp-ai-client
+> silently falls back to the provider default; openclaWP logs that substitution
+> as `[openclawp] model_substitution …` so you can catch it. Pair with the
+> per-site budget caps (`budget_daily_usd`, `budget_daily_turns`).
+
 ### Path B — an existing WordPress site
 
 If a site already exists (Pressable, a VPS, [Local](https://localwp.com/), your own Docker, …), drop the plugins in. Anthropic shown to keep the block short — the Ollama recipe in Path A applies verbatim, just call `wp` instead of `studio --path … wp`.
diff --git a/includes/class-openclawp-agent-registrar.php b/includes/class-openclawp-agent-registrar.php
index e085487..d77ca1c 100644
--- a/includes/class-openclawp-agent-registrar.php
+++ b/includes/class-openclawp-agent-registrar.php
@@ -53,7 +53,7 @@ public static function maybe_register_site_introspection_agent(): void {
 				'owner_resolver' => static fn(): int => get_current_user_id(),
 				'default_config' => array(
 					'provider'  => 'auto',
-					'model'     => 'claude-haiku-4-5',
+					'model'     => 'auto',
 					'tools'     => array(
 						'openclawp/get-recent-posts',
 						'openclawp/count-comments',
@@ -93,7 +93,7 @@ public static function maybe_register_loop_demo_agent(): void {
 				'owner_resolver' => static fn(): int => get_current_user_id(),
 				'default_config' => array(
 					'provider'  => 'auto',
-					'model'     => 'claude-haiku-4-5',
+					'model'     => 'auto',
 					'tools'     => array( 'openclawp/get-time' ),
 					'max_turns' => 5,
 				),
@@ -131,7 +131,7 @@ public static function maybe_register_example_agent(): void {
 				'owner_resolver' => static fn(): int => get_current_user_id(),
 				'default_config' => array(
 					'provider' => 'auto',
-					'model'    => 'claude-haiku-4-5',
+					'model'    => 'auto',
 				),
 				'meta'           => array(
 					'source_plugin'  => 'openclawp/openclawp.php',
@@ -177,7 +177,7 @@ public static function maybe_register_workflow_drafter_agent(): void {
 				'owner_resolver' => static fn(): int => get_current_user_id(),
 				'default_config' => array(
 					'provider' => 'auto',
-					'model'    => 'claude-haiku-4-5',
+					'model'    => 'auto',
 				),
 				'meta'           => array(
 					'source_plugin'  => 'openclawp/openclawp.php',
@@ -305,7 +305,7 @@ public static function maybe_register_coordinator_demo_agent(): void {
 				'owner_resolver' => static fn(): int => get_current_user_id(),
 				'default_config' => array(
 					'provider'  => 'auto',
-					'model'     => 'claude-haiku-4-5',
+					'model'     => 'auto',
 					'max_turns' => 6,
 				),
 				'subagents'      => array(
diff --git a/includes/class-openclawp-runner.php b/includes/class-openclawp-runner.php
index de6a3db..405e6d8 100644
--- a/includes/class-openclawp-runner.php
+++ b/includes/class-openclawp-runner.php
@@ -368,6 +368,17 @@ private static function build_turn_runner( WP_Agent $agent, array $session, stri
 				if ( null !== $preference ) {
 					$builder = $builder->using_model_preference( $preference );
 				}
+				// The model id the agent asked for (empty when 'auto'/unset). Used
+				// below to flag silent provider substitution — wp-ai-client falls
+				// back to the provider's default model (often the priciest) when the
+				// requested id isn't one the active provider exposes, rather than
+				// erroring, which has surprised installs with unexpected billing.
+				$requested_model_id = '';
+				if ( is_array( $preference ) ) {
+					$requested_model_id = (string) ( $preference[1] ?? '' );
+				} elseif ( is_string( $preference ) ) {
+					$requested_model_id = $preference;
+				}
 
 				$start_us  = microtime( true );
 				$generated = $builder->generate_text_result();
@@ -390,6 +401,25 @@ private static function build_turn_runner( WP_Agent $agent, array $session, stri
 				$telemetry['model']       = self::extract_model_id( $generated );
 				$telemetry['token_usage'] = self::extract_token_usage( $generated );
 
+				// Surface silent model substitution: the agent asked for one model
+				// but the provider answered with another (its default fallback). This
+				// is otherwise invisible and can mean paying for a premium model
+				// (e.g. an unrecognised id resolving to Opus). Record it on the turn
+				// telemetry and warn once per turn.
+				$used_model = (string) $telemetry['model'];
+				if ( '' !== $requested_model_id && '' !== $used_model && $requested_model_id !== $used_model ) {
+					$telemetry['model_requested'] = $requested_model_id;
+					// phpcs:ignore WordPress.PHP.DevelopmentFunctions.error_log_error_log -- Matches openclaWP's telemetry channel (grep error_log for [openclawp]).
+					error_log(
+						sprintf(
+							'[openclawp] model_substitution agent=%s requested=%s used=%s — the configured model id did not resolve on the active provider and silently fell back. Pin a model id the provider exposes to control cost.',
+							(string) $agent_obj->get_slug(),
+							$requested_model_id,
+							$used_model
+						)
+					);
+				}
+
 				$tool_calls = self::extract_tool_calls( $generated );
 				$telemetry['tool_call_count'] = count( $tool_calls );
 				// `toText()` throws "No text content found in first candidate"
diff --git a/tests/integration/prompt-assembly/__snapshots__/coordinator--chat.json b/tests/integration/prompt-assembly/__snapshots__/coordinator--chat.json
index 3d3c7b3..d90709a 100644
--- a/tests/integration/prompt-assembly/__snapshots__/coordinator--chat.json
+++ b/tests/integration/prompt-assembly/__snapshots__/coordinator--chat.json
@@ -12,9 +12,7 @@
 			"role": "user"
 		}
 	],
-	"model_preference": {
-		"model": "claude-haiku-4-5"
-	},
+	"model_preference": null,
 	"runtime_context": {
 		"client_context": {
 			"client_name": "chat-block",
diff --git a/tests/integration/prompt-assembly/__snapshots__/example--whatsapp.json b/tests/integration/prompt-assembly/__snapshots__/example--whatsapp.json
index 6d0b69c..92140b3 100644
--- a/tests/integration/prompt-assembly/__snapshots__/example--whatsapp.json
+++ b/tests/integration/prompt-assembly/__snapshots__/example--whatsapp.json
@@ -9,9 +9,7 @@
 			"role": "user"
 		}
 	],
-	"model_preference": {
-		"model": "claude-haiku-4-5"
-	},
+	"model_preference": null,
 	"runtime_context": {
 		"client_context": {
 			"client_name": "whatsapp",
diff --git a/tests/integration/prompt-assembly/__snapshots__/loop-demo--chat.json b/tests/integration/prompt-assembly/__snapshots__/loop-demo--chat.json
index 5e6386e..6e4ca67 100644
--- a/tests/integration/prompt-assembly/__snapshots__/loop-demo--chat.json
+++ b/tests/integration/prompt-assembly/__snapshots__/loop-demo--chat.json
@@ -9,9 +9,7 @@
 			"role": "user"
 		}
 	],
-	"model_preference": {
-		"model": "claude-haiku-4-5"
-	},
+	"model_preference": null,
 	"runtime_context": {
 		"client_context": {
 			"client_name": "chat-block",
diff --git a/tests/integration/prompt-assembly/__snapshots__/site-introspection--whatsapp.json b/tests/integration/prompt-assembly/__snapshots__/site-introspection--whatsapp.json
index 238598b..3b21a13 100644
--- a/tests/integration/prompt-assembly/__snapshots__/site-introspection--whatsapp.json
+++ b/tests/integration/prompt-assembly/__snapshots__/site-introspection--whatsapp.json
@@ -9,9 +9,7 @@
 			"role": "user"
 		}
 	],
-	"model_preference": {
-		"model": "claude-haiku-4-5"
-	},
+	"model_preference": null,
 	"runtime_context": {
 		"client_context": {
 			"client_name": "whatsapp",
diff --git a/tests/integration/prompt-assembly/__snapshots__/workflow-drafter--chat.json b/tests/integration/prompt-assembly/__snapshots__/workflow-drafter--chat.json
index 1669017..b6198fb 100644
--- a/tests/integration/prompt-assembly/__snapshots__/workflow-drafter--chat.json
+++ b/tests/integration/prompt-assembly/__snapshots__/workflow-drafter--chat.json
@@ -9,9 +9,7 @@
 			"role": "user"
 		}
 	],
-	"model_preference": {
-		"model": "claude-haiku-4-5"
-	},
+	"model_preference": null,
 	"runtime_context": {
 		"client_context": {
 			"client_name": "chat-block",