merge: integrate usage response models with service tier badges

Merge upstream PR #837, preserving its original commits and resolving the UsageModelDisplay layout conflict after #841. Cover response models alongside dynamic service-tier badges in detail tests.

apps/aether-gateway/src/handlers/public/support/user_me_usage.rs

@@ -567,6 +567,7 @@ fn build_users_me_usage_record_payload(
         "id": item.id,
         "model": item.model,
         "target_model": serde_json::Value::Null,
+        "response_model": item.provider_response_model(),
         "api_format": item.api_format,
         "endpoint_api_format": item.endpoint_api_format,
         "has_format_conversion": item.has_format_conversion,
@@ -681,6 +682,7 @@ fn build_users_me_usage_active_payload(item: &StoredRequestUsageAudit) -> serde_
         "client_ip": users_me_usage_metadata_string(item, "client_ip"),
         "user_agent": users_me_usage_metadata_string(item, "user_agent"),
         "target_model": item.target_model,
+        "response_model": item.provider_response_model(),
         "has_fallback": item.has_fallback(),
     });
     payload["end_to_end_time_ms"] = json!(users_me_usage_metadata_u64(item, "end_to_end_time_ms"));
@@ -1861,6 +1863,25 @@ mod tests {
         assert_eq!(payload["cache_creation_ephemeral_1h_input_tokens"], 6);
     }
 
+    #[test]
+    fn user_usage_payloads_expose_response_model_separately_from_mapping() {
+        let item = StoredRequestUsageAudit {
+            target_model: Some("provider-mapped-model".to_string()),
+            request_metadata: Some(json!({
+                "provider_response_model": "gpt-5.1"
+            })),
+            ..sample_usage("completed")
+        };
+
+        let record = build_users_me_usage_record_payload(&item, false, &BTreeMap::new(), false);
+        let active = build_users_me_usage_active_payload(&item);
+
+        for payload in [&record, &active] {
+            assert_eq!(payload["target_model"], "provider-mapped-model");
+            assert_eq!(payload["response_model"], "gpt-5.1");
+        }
+    }
+
     #[test]
     fn user_usage_payloads_project_end_to_end_timings_from_metadata() {
         let item = StoredRequestUsageAudit {

crates/aether-admin/src/observability/usage.rs

@@ -1349,6 +1349,7 @@ fn admin_usage_active_request_json(
     if let Some(target_model) = item.target_model.as_ref() {
         value["target_model"] = json!(target_model);
     }
+    value["response_model"] = json!(item.provider_response_model());
     if let Some(reasoning_effort) = item.provider_reasoning_effort() {
         value["reasoning_effort"] = json!(reasoning_effort);
     }
@@ -1445,6 +1446,11 @@ pub fn admin_usage_record_json(
     let object = payload
         .as_object_mut()
         .expect("admin usage record payload should be an object");
+    // 大型 json! 宏接近 Rust 的递归展开上限，响应模型在宏展开后补入即可避免编译失败。
+    object.insert(
+        "response_model".to_string(),
+        json!(item.provider_response_model()),
+    );
     object.insert(
         "end_to_end_time_ms".to_string(),
         json!(admin_usage_metadata_u64(item, "end_to_end_time_ms")),
@@ -2842,6 +2848,32 @@ mod tests {
         assert_eq!(record["client_is_stream"], false);
     }
 
+    #[test]
+    fn admin_usage_payloads_expose_response_model_separately_from_mapping() {
+        let item = StoredRequestUsageAudit {
+            target_model: Some("provider-mapped-model".to_string()),
+            request_metadata: Some(json!({
+                "provider_response_model": "gpt-5.1"
+            })),
+            ..sample_usage("completed", Some(200), None)
+        };
+
+        let record = admin_usage_record_json(
+            &item,
+            &BTreeMap::new(),
+            &BTreeMap::new(),
+            false,
+            false,
+            None,
+        );
+        let active = admin_usage_active_request_json(&item, None, None, None);
+
+        for payload in [&record, &active] {
+            assert_eq!(payload["target_model"], "provider-mapped-model");
+            assert_eq!(payload["response_model"], "gpt-5.1");
+        }
+    }
+
     #[test]
     fn admin_usage_payloads_project_end_to_end_timings_from_metadata() {
         let item = StoredRequestUsageAudit {

crates/aether-data/adapters/postgres/src/usage/queries/list_recent_usage_audits_prefix.sql

@@ -183,6 +183,7 @@ SELECT
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_reasoning_effort'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), '') IS NOT NULL
+      OR NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), '') IS NOT NULL
       OR ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')
       OR ("usage".request_metadata->>'upstream_is_stream') IN ('true', 'false')
       OR ("usage".request_metadata->>'websocket_mode') IN ('true', 'false')
@@ -208,6 +209,8 @@ SELECT
         NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), ''),
         'provider_actual_service_tier',
         NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), ''),
+        'provider_response_model',
+        NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), ''),
         'client_requested_stream',
         CASE
           WHEN ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')

crates/aether-data/adapters/postgres/src/usage/queries/list_usage_audits_prefix.sql

@@ -183,6 +183,7 @@ SELECT
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_reasoning_effort'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), '') IS NOT NULL
+      OR NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), '') IS NOT NULL
       OR ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')
       OR ("usage".request_metadata->>'upstream_is_stream') IN ('true', 'false')
       OR ("usage".request_metadata->>'websocket_mode') IN ('true', 'false')
@@ -208,6 +209,8 @@ SELECT
         NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), ''),
         'provider_actual_service_tier',
         NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), ''),
+        'provider_response_model',
+        NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), ''),
         'client_requested_stream',
         CASE
           WHEN ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')

crates/aether-data/adapters/postgres/src/usage/tests.rs

@@ -3611,6 +3611,8 @@ fn usage_sql_uses_json_null_placeholders_for_usage_payload_columns() {
         assert!(sql.contains("request_metadata->>'provider_reasoning_effort'"));
         assert!(sql.contains("request_metadata->>'provider_service_tier'"));
         assert!(sql.contains("request_metadata->>'provider_actual_service_tier'"));
+        assert!(sql.contains("request_metadata->>'provider_response_model'"));
+        assert!(sql.contains("'provider_response_model'"));
         assert!(sql.contains("request_metadata->>'websocket_mode'"));
         assert!(sql.contains("'websocket_mode'"));
         assert!(sql.contains("AS client_family"));

crates/aether-data/contracts/src/repository/usage/metadata_policy.rs

@@ -8,9 +8,10 @@ use serde_json::{Map, Value};
 use crate::repository::candidates::sanitize_request_candidate_skip_reason;
 
 use super::{
-    LIVE_SESSION_METADATA_KEY, PLAN_USAGE_RESERVATION_DEFERRED_METADATA_KEY,
-    PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY, PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
-    PROVIDER_REASONING_EFFORT_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
+    normalize_provider_response_model, LIVE_SESSION_METADATA_KEY,
+    PLAN_USAGE_RESERVATION_DEFERRED_METADATA_KEY, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY,
+    PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY, PROVIDER_REASONING_EFFORT_METADATA_KEY,
+    PROVIDER_RESPONSE_MODEL_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
     REALTIME_SESSION_METADATA_KEY, REQUESTED_REASONING_EFFORT_METADATA_KEY,
     ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY, ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY,
     USAGE_AVAILABLE_METADATA_KEY, USAGE_PRICING_AVAILABLE_METADATA_KEY,
@@ -94,6 +95,12 @@ pub fn sanitize_usage_request_metadata_object(source: &Map<String, Value>) -> Op
     ] {
         insert_known_string(source, &mut target, key, sanitize_service_tier);
     }
+    insert_known_string(
+        source,
+        &mut target,
+        PROVIDER_RESPONSE_MODEL_METADATA_KEY,
+        normalize_provider_response_model,
+    );
     insert_bounded_u64(
         source,
         &mut target,
@@ -1279,6 +1286,23 @@ mod tests {
         }
     }
 
+    #[test]
+    fn persistence_projection_keeps_bounded_response_model_only_as_a_string() {
+        let metadata = sanitize_usage_request_metadata(Some(json!({
+            "provider_response_model": "  GPT-5.1  "
+        })))
+        .expect("response model should remain");
+        assert_eq!(metadata["provider_response_model"], "GPT-5.1");
+        assert!(sanitize_usage_request_metadata(Some(json!({
+            "provider_response_model": 42
+        })))
+        .is_none());
+        assert!(sanitize_usage_request_metadata(Some(json!({
+            "provider_response_model": "x".repeat(257)
+        })))
+        .is_none());
+    }
+
     #[test]
     fn persistence_projection_keeps_only_bounded_settlement_facts() {
         let metadata = sanitize_usage_request_metadata(Some(json!({

crates/aether-data/contracts/src/repository/usage/mod.rs

@@ -15,11 +15,12 @@ pub use policy::*;
 pub use types::{
     canonical_usage_body_ref_for, extract_provider_actual_service_tier_from_response,
     extract_provider_cache_ttl_minutes_from_metadata, extract_provider_reasoning_effort_from_body,
-    extract_provider_service_tier_from_body, normalize_provider_service_tier, parse_usage_body_ref,
+    extract_provider_response_model_from_bodies, extract_provider_service_tier_from_body,
+    normalize_provider_response_model, normalize_provider_service_tier, parse_usage_body_ref,
     resolve_provider_cache_ttl_minutes, resolve_provider_service_tier_from_request_capture,
-    usage_body_ref, usage_request_metadata_client_family, ApiKeyLastUsedDelta,
-    ManagementTokenCounterDelta, PendingUsageCleanupSummary, ProviderApiKeyWindowUsageRequest,
-    ProxyNodeCounterDelta, StoredProviderApiKeyUsageSummary,
+    usage_body_capture_is_authoritative, usage_body_ref, usage_request_metadata_client_family,
+    ApiKeyLastUsedDelta, ManagementTokenCounterDelta, PendingUsageCleanupSummary,
+    ProviderApiKeyWindowUsageRequest, ProxyNodeCounterDelta, StoredProviderApiKeyUsageSummary,
     StoredProviderApiKeyWindowUsageSummary, StoredProviderUsageSummary, StoredProviderUsageWindow,
     StoredRequestUsageAudit, StoredUsageAuditAggregation, StoredUsageAuditSummary,
     StoredUsageBodyPayload, StoredUsageBreakdownSummaryRow, StoredUsageCacheAffinityHitSummary,
@@ -46,9 +47,9 @@ pub use types::{
     UsageTimeSeriesQuery, UsageWriteRepository, LIVE_SESSION_METADATA_KEY,
     PLAN_USAGE_RESERVATION_DEFERRED_METADATA_KEY, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY,
     PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY, PROVIDER_REASONING_EFFORT_METADATA_KEY,
-    PROVIDER_SERVICE_TIER_METADATA_KEY, REALTIME_SESSION_METADATA_KEY,
-    REQUESTED_REASONING_EFFORT_METADATA_KEY, ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY,
-    ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY, USAGE_AVAILABLE_METADATA_KEY,
-    USAGE_PRICING_AVAILABLE_METADATA_KEY, WEBSOCKET_MODE_METADATA_KEY,
-    WEBSOCKET_TRANSPORT_METADATA_KEY,
+    PROVIDER_RESPONSE_MODEL_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
+    REALTIME_SESSION_METADATA_KEY, REQUESTED_REASONING_EFFORT_METADATA_KEY,
+    ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY, ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY,
+    USAGE_AVAILABLE_METADATA_KEY, USAGE_PRICING_AVAILABLE_METADATA_KEY,
+    WEBSOCKET_MODE_METADATA_KEY, WEBSOCKET_TRANSPORT_METADATA_KEY,
 };

crates/aether-data/contracts/src/repository/usage/types.rs

@@ -1,3 +1,4 @@
+use aether_ai_formats::normalize_api_format_alias;
 use async_trait::async_trait;
 use chrono::{DateTime, Utc};
 use serde_json::Value;
@@ -6,6 +7,7 @@ pub const PROVIDER_REASONING_EFFORT_METADATA_KEY: &str = "provider_reasoning_eff
 pub const REQUESTED_REASONING_EFFORT_METADATA_KEY: &str = "requested_reasoning_effort";
 pub const PROVIDER_SERVICE_TIER_METADATA_KEY: &str = "provider_service_tier";
 pub const PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY: &str = "provider_actual_service_tier";
+pub const PROVIDER_RESPONSE_MODEL_METADATA_KEY: &str = "provider_response_model";
 pub const PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY: &str = "provider_cache_ttl_minutes";
 pub const ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY: &str = "routing_candidate_skip_reason";
 pub const ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY: &str = "routing_failure_diagnostic";
@@ -119,6 +121,141 @@ pub fn normalize_provider_service_tier(value: &str) -> Option<String> {
     Some(value.to_ascii_lowercase())
 }
 
+/// 清洗模型名称，保留大小写，只去除首尾空白。
+pub fn normalize_provider_response_model(value: &str) -> Option<String> {
+    let value = value.trim();
+    if value.is_empty() || value.len() > 256 {
+        return None;
+    }
+    Some(value.to_string())
+}
+
+fn extract_model_at_paths(value: &Value, paths: &[&[&str]]) -> Option<String> {
+    paths.iter().find_map(|path| {
+        let value = path
+            .iter()
+            .try_fold(value, |current, key| current.as_object()?.get(*key))?;
+        value.as_str().and_then(normalize_provider_response_model)
+    })
+}
+
+fn response_model_paths(provider_api_format: Option<&str>) -> &'static [&'static [&'static str]] {
+    match normalize_api_format_alias(provider_api_format.unwrap_or_default()).as_str() {
+        "gemini:generate_content" => {
+            // Gemini 原生响应使用 modelVersion；部分网关会改写为 model。
+            &[&["modelVersion"], &["model_version"], &["model"]]
+        }
+        "gemini:embedding" => {
+            // Gemini Embedding 可能返回 model、modelVersion 或 Vertex 的 deployedModelId。
+            &[
+                &["model"],
+                &["modelVersion"],
+                &["model_version"],
+                &["deployedModelId"],
+                &["deployed_model_id"],
+            ]
+        }
+        "gemini:interactions" => {
+            // Interactions 请求既可能叫 model，也可能叫 agent；响应优先读取 model。
+            &[
+                &["model"],
+                &["modelVersion"],
+                &["model_version"],
+                &["agent"],
+            ]
+        }
+        _ => &[&["model"]],
+    }
+}
+
+fn extract_model_from_known_response_wrappers(
+    response_body: &Value,
+    paths: &[&[&str]],
+) -> Option<String> {
+    // 只展开协议中已知的 response/chunks 包装，避免在候选内容、工具参数等任意嵌套
+    // 对象中搜索同名字段，误把 role="model" 一类内容当成响应模型。
+    extract_model_at_paths(response_body, paths)
+        .or_else(|| {
+            response_body
+                .get("response")
+                .and_then(|response| extract_model_at_paths(response, paths))
+        })
+        .or_else(|| {
+            response_body
+                .get("chunks")
+                .and_then(Value::as_array)
+                .and_then(|chunks| {
+                    chunks.iter().rev().find_map(|chunk| {
+                        extract_model_at_paths(chunk, paths).or_else(|| {
+                            chunk
+                                .get("response")
+                                .and_then(|response| extract_model_at_paths(response, paths))
+                        })
+                    })
+                })
+        })
+        .or_else(|| {
+            response_body
+                .get("response")
+                .and_then(|response| response.get("chunks"))
+                .and_then(Value::as_array)
+                .and_then(|chunks| {
+                    chunks.iter().rev().find_map(|chunk| {
+                        extract_model_at_paths(chunk, paths).or_else(|| {
+                            chunk
+                                .get("response")
+                                .and_then(|response| extract_model_at_paths(response, paths))
+                        })
+                    })
+                })
+        })
+}
+
+fn extract_provider_model_from_response_body(
+    response_body: &Value,
+    provider_api_format: Option<&str>,
+) -> Option<String> {
+    extract_model_from_known_response_wrappers(
+        response_body,
+        response_model_paths(provider_api_format),
+    )
+}
+
+fn extract_provider_model_from_request_body(
+    request_body: &Value,
+    request_api_format: Option<&str>,
+) -> Option<String> {
+    let paths: &[&[&str]] =
+        match normalize_api_format_alias(request_api_format.unwrap_or_default()).as_str() {
+            "gemini:interactions" => &[&["model"], &["agent"]],
+            _ => &[&["model"]],
+        };
+    extract_model_at_paths(request_body, paths)
+}
+
+/// 只有请求体和响应体都可作为完整事实时，才计算响应模型，避免用截断内容猜测。
+pub fn extract_provider_response_model_from_bodies(
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+    request_api_format: Option<&str>,
+    response_body: Option<&Value>,
+    response_body_state: Option<UsageBodyCaptureState>,
+    provider_api_format: Option<&str>,
+) -> Option<String> {
+    if !usage_body_capture_is_authoritative(request_body, request_body_state)
+        || !usage_body_capture_is_authoritative(response_body, response_body_state)
+    {
+        return None;
+    }
+
+    let request_model =
+        extract_provider_model_from_request_body(request_body?, request_api_format)?;
+    let response_model =
+        extract_provider_model_from_response_body(response_body?, provider_api_format)?;
+
+    (request_model != response_model).then_some(response_model)
+}
+
 /// Resolves a provider processing tier exclusively from the final upstream request.
 ///
 /// A complete captured body is authoritative, including when it contains no tier. The metadata
@@ -129,7 +266,7 @@ pub fn resolve_provider_service_tier_from_request_capture(
     provider_request_body_state: Option<UsageBodyCaptureState>,
     request_metadata: Option<&Value>,
 ) -> Option<String> {
-    if request_body_capture_is_authoritative(provider_request_body, provider_request_body_state) {
+    if usage_body_capture_is_authoritative(provider_request_body, provider_request_body_state) {
         return extract_provider_service_tier_from_body(provider_request_body);
     }
 
@@ -153,7 +290,7 @@ pub fn resolve_provider_service_tier_from_request_capture(
         .and_then(normalize_provider_service_tier)
 }
 
-fn request_body_capture_is_authoritative(
+pub fn usage_body_capture_is_authoritative(
     request_body: Option<&Value>,
     request_body_state: Option<UsageBodyCaptureState>,
 ) -> bool {
@@ -184,7 +321,7 @@ fn resolve_reasoning_effort_from_request_capture(
     request_metadata: Option<&Value>,
     metadata_key: &str,
 ) -> Option<String> {
-    if request_body_capture_is_authoritative(request_body, request_body_state) {
+    if usage_body_capture_is_authoritative(request_body, request_body_state) {
         return extract_provider_reasoning_effort_from_body(request_body);
     }
 
@@ -701,6 +838,11 @@ impl StoredRequestUsageAudit {
             })
     }
 
+    pub fn provider_response_model(&self) -> Option<String> {
+        self.request_metadata_string(PROVIDER_RESPONSE_MODEL_METADATA_KEY)
+            .and_then(normalize_provider_response_model)
+    }
+
     pub fn provider_cache_ttl_minutes(&self) -> Option<i64> {
         resolve_provider_cache_ttl_minutes(
             self.endpoint_api_format
@@ -2646,7 +2788,8 @@ fn parse_timestamp(value: i64, field_name: &str) -> Result<u64, crate::DataLayer
 mod tests {
     use super::{
         canonical_usage_body_ref_for, extract_provider_actual_service_tier_from_response,
-        extract_provider_service_tier_from_body, normalize_provider_reasoning_effort,
+        extract_provider_response_model_from_bodies, extract_provider_service_tier_from_body,
+        normalize_provider_reasoning_effort, normalize_provider_response_model,
         normalize_provider_service_tier, resolve_provider_cache_ttl_minutes, usage_body_ref,
         StoredRequestUsageAudit, UpsertUsageRecord, UsageBodyCaptureState, UsageBodyCaptureStorage,
         UsageBodyField, UsageProviderPerformanceQuery, REALTIME_SESSION_METADATA_KEY,
@@ -2672,6 +2815,122 @@ mod tests {
         }
     }
 
+    #[test]
+    fn response_model_requires_authoritative_different_top_level_models() {
+        let request = json!({"model": " gpt-5 "});
+        let response = json!({"model": " gpt-5.1 "});
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&request),
+                Some(UsageBodyCaptureState::Inline),
+                Some("openai:chat"),
+                Some(&response),
+                Some(UsageBodyCaptureState::Inline),
+                Some("openai:chat"),
+            ),
+            Some("gpt-5.1".to_string())
+        );
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&request),
+                Some(UsageBodyCaptureState::Inline),
+                Some("openai:responses"),
+                Some(&json!({"model": "gpt-5"})),
+                Some(UsageBodyCaptureState::Inline),
+                Some("openai:responses"),
+            ),
+            None
+        );
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&request),
+                Some(UsageBodyCaptureState::Truncated),
+                Some("openai:chat"),
+                Some(&response),
+                Some(UsageBodyCaptureState::Inline),
+                Some("openai:chat"),
+            ),
+            None
+        );
+    }
+
+    #[test]
+    fn response_model_keeps_case_and_rejects_invalid_values() {
+        assert_eq!(
+            normalize_provider_response_model("  GPT-5.1  "),
+            Some("GPT-5.1".to_string())
+        );
+        assert_eq!(normalize_provider_response_model("  "), None);
+        assert_eq!(normalize_provider_response_model(&"x".repeat(257)), None);
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&json!({"model": "gpt-5"})),
+                None,
+                Some("openai:chat"),
+                Some(&json!({"model": 42})),
+                None,
+                Some("openai:chat"),
+            ),
+            None
+        );
+    }
+
+    #[test]
+    fn response_model_uses_provider_format_specific_nested_paths() {
+        let request = json!({"model": "gemini-2.5-flash"});
+        let response = json!({
+            "response": {
+                "modelVersion": "gemini-2.5-flash-001",
+                "candidates": [{"content": {"role": "model"}}]
+            }
+        });
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&request),
+                Some(UsageBodyCaptureState::Inline),
+                Some("gemini:generate_content"),
+                Some(&response),
+                Some(UsageBodyCaptureState::Inline),
+                Some("gemini:generate_content"),
+            ),
+            Some("gemini-2.5-flash-001".to_string())
+        );
+
+        let wrapped_chunks = json!({
+            "chunks": [
+                {"response": {"modelVersion": "gemini-old"}},
+                {"response": {"modelVersion": "gemini-final"}}
+            ]
+        });
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&request),
+                Some(UsageBodyCaptureState::Inline),
+                Some("gemini:generate_content"),
+                Some(&wrapped_chunks),
+                Some(UsageBodyCaptureState::Inline),
+                Some("gemini:generate_content"),
+            ),
+            Some("gemini-final".to_string())
+        );
+
+        let ambiguous = json!({
+            "metadata": {"model": "do-not-use"},
+            "candidates": [{"content": {"role": "model"}}]
+        });
+        assert_eq!(
+            extract_provider_response_model_from_bodies(
+                Some(&request),
+                Some(UsageBodyCaptureState::Inline),
+                Some("gemini:generate_content"),
+                Some(&ambiguous),
+                Some(UsageBodyCaptureState::Inline),
+                Some("gemini:generate_content"),
+            ),
+            None
+        );
+    }
+
     #[test]
     fn provider_fact_normalization_preserves_non_ascii_case_and_byte_count() {
         for normalize in [

crates/aether-usage/runtime/src/event_wire.rs

@@ -15,9 +15,9 @@ use super::{BorrowedUsageEventEnvelope, UsageEvent, UsageEventData, USAGE_EVENT_
 use crate::body_capture::mark_usage_event_capture_truncated;
 use crate::request_metadata::{
     attach_client_request_body_metadata, attach_provider_request_body_metadata,
-    attach_provider_response_body_metadata, clear_client_request_body_metadata,
-    clear_provider_request_body_metadata, request_body_derived_facts_action,
-    RequestBodyDerivedFactsAction,
+    attach_provider_response_body_metadata, attach_provider_response_model_metadata,
+    clear_client_request_body_metadata, clear_provider_request_body_metadata,
+    request_body_derived_facts_action, RequestBodyDerivedFactsAction,
 };
 
 const DIAGNOSTIC_FIELDS: [&str; 8] = [
@@ -238,6 +238,15 @@ impl WireOverrides {
             RequestBodyDerivedFactsAction::Clear | RequestBodyDerivedFactsAction::Preserve => {}
         }
         metadata = attach_provider_response_body_metadata(metadata, data.response_body.as_ref());
+        metadata = attach_provider_response_model_metadata(
+            metadata,
+            request_body,
+            data.request_body_state,
+            data.api_format.as_deref(),
+            data.response_body.as_ref(),
+            data.response_body_state,
+            data.endpoint_api_format.as_deref(),
+        );
         // Billing reads raw-body TTL before metadata regardless of capture state.
         // Preserve that precedence independently of reasoning and tier authority.
         if let Some(cache_ttl) = body_cache_ttl {

crates/aether-usage/runtime/src/record.rs

@@ -5,9 +5,9 @@ use aether_data_contracts::DataLayerError;
 
 use crate::request_metadata::{
     attach_client_request_body_metadata, attach_provider_request_body_metadata,
-    clear_client_request_body_metadata, clear_provider_request_body_metadata,
-    request_body_derived_facts_action, sanitize_usage_request_metadata,
-    RequestBodyDerivedFactsAction,
+    attach_provider_response_model_metadata, clear_client_request_body_metadata,
+    clear_provider_request_body_metadata, request_body_derived_facts_action,
+    sanitize_usage_request_metadata, RequestBodyDerivedFactsAction,
 };
 use crate::{UsageEvent, UsageEventType};
 
@@ -85,6 +85,16 @@ pub fn build_upsert_usage_record_from_event(
         }
         RequestBodyDerivedFactsAction::Preserve => {}
     }
+    // 响应模型必须在 body 被裁剪前从客户端请求体和上游响应体共同派生；缺少权威 body 时保留已派生事实。
+    data.request_metadata = attach_provider_response_model_metadata(
+        data.request_metadata,
+        data.request_body.as_ref(),
+        data.request_body_state,
+        data.api_format.as_deref(),
+        data.response_body.as_ref(),
+        data.response_body_state,
+        data.endpoint_api_format.as_deref(),
+    );
     let now_unix_secs = event.timestamp_ms / 1_000;
 
     Ok(UpsertUsageRecord {

crates/aether-usage/runtime/src/request_metadata.rs

@@ -1,13 +1,15 @@
 use aether_contracts::ExecutionPlan;
 use aether_data_contracts::repository::usage::{
     extract_provider_actual_service_tier_from_response,
-    extract_provider_reasoning_effort_from_body, extract_provider_service_tier_from_body,
-    normalize_provider_service_tier, resolve_provider_cache_ttl_minutes,
+    extract_provider_reasoning_effort_from_body, extract_provider_response_model_from_bodies,
+    extract_provider_service_tier_from_body, normalize_provider_service_tier,
+    resolve_provider_cache_ttl_minutes,
     sanitize_usage_request_metadata as project_usage_request_metadata,
     sanitize_usage_request_metadata_object as project_usage_request_metadata_object,
     sanitize_usage_request_metadata_ref as project_usage_request_metadata_ref,
-    UsageBodyCaptureState, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY,
-    PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY, PROVIDER_REASONING_EFFORT_METADATA_KEY,
+    usage_body_capture_is_authoritative, UsageBodyCaptureState,
+    PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY, PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
+    PROVIDER_REASONING_EFFORT_METADATA_KEY, PROVIDER_RESPONSE_MODEL_METADATA_KEY,
     PROVIDER_SERVICE_TIER_METADATA_KEY, REQUESTED_REASONING_EFFORT_METADATA_KEY,
 };
 use serde_json::{Map, Value};
@@ -249,6 +251,83 @@ pub(crate) fn attach_provider_response_body_metadata(
     attach_provider_actual_service_tier_metadata(metadata, actual_service_tier.as_deref())
 }
 
+pub(crate) fn attach_provider_response_model_metadata(
+    metadata: Option<Value>,
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+    request_api_format: Option<&str>,
+    response_body: Option<&Value>,
+    response_body_state: Option<UsageBodyCaptureState>,
+    provider_api_format: Option<&str>,
+) -> Option<Value> {
+    let both_bodies_are_authoritative =
+        usage_body_capture_is_authoritative(request_body, request_body_state)
+            && usage_body_capture_is_authoritative(response_body, response_body_state);
+    let response_model = extract_provider_response_model_from_bodies(
+        request_body,
+        request_body_state,
+        request_api_format,
+        response_body,
+        response_body_state,
+        provider_api_format,
+    );
+    if !both_bodies_are_authoritative && response_model.is_none() {
+        return metadata;
+    }
+
+    let mut object = match metadata {
+        Some(Value::Object(object)) => object,
+        _ => Map::new(),
+    };
+    // 完整终态 body 是最终候选的权威事实；相同、无效或缺失模型都要清除旧候选值。
+    if both_bodies_are_authoritative {
+        object.remove(PROVIDER_RESPONSE_MODEL_METADATA_KEY);
+    }
+    if let Some(response_model) = response_model {
+        object.insert(
+            PROVIDER_RESPONSE_MODEL_METADATA_KEY.to_string(),
+            Value::String(response_model),
+        );
+    }
+    (!object.is_empty()).then_some(Value::Object(object))
+}
+
+/// 终态候选无法完成比较时，显式清除旧响应模型，避免重试/故障转移残留。
+pub(crate) fn refresh_provider_response_model_metadata(
+    metadata: Option<Value>,
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+    request_api_format: Option<&str>,
+    response_body: Option<&Value>,
+    response_body_state: Option<UsageBodyCaptureState>,
+    provider_api_format: Option<&str>,
+) -> Option<Value> {
+    let mut object = match metadata {
+        Some(Value::Object(object)) => object,
+        _ => Map::new(),
+    };
+    object.remove(PROVIDER_RESPONSE_MODEL_METADATA_KEY);
+
+    if usage_body_capture_is_authoritative(request_body, request_body_state)
+        && usage_body_capture_is_authoritative(response_body, response_body_state)
+    {
+        if let Some(response_model) = extract_provider_response_model_from_bodies(
+            request_body,
+            request_body_state,
+            request_api_format,
+            response_body,
+            response_body_state,
+            provider_api_format,
+        ) {
+            object.insert(
+                PROVIDER_RESPONSE_MODEL_METADATA_KEY.to_string(),
+                Value::String(response_model),
+            );
+        }
+    }
+    (!object.is_empty()).then_some(Value::Object(object))
+}
+
 /// Refreshes the response-derived tier for a terminal snapshot. Complete response objects are
 /// authoritative even when they contain no tier (which clears a stale candidate value). Capture
 /// placeholders/absent bodies are not authoritative, so a terminal summary already present in
@@ -312,6 +391,7 @@ pub(crate) fn attach_provider_actual_service_tier_metadata(
 #[cfg(test)]
 mod tests {
     use aether_contracts::{ExecutionPlan, RequestBody};
+    use aether_data_contracts::repository::usage::UsageBodyCaptureState;
     use serde_json::{json, Value};
     use std::collections::BTreeMap;
 
@@ -323,8 +403,9 @@ mod tests {
     use super::{
         attach_client_request_body_metadata, attach_provider_actual_service_tier_metadata,
         attach_provider_request_body_metadata, attach_provider_response_body_metadata,
-        build_usage_request_metadata_seed, merge_usage_request_metadata,
-        merge_usage_request_metadata_owned, refresh_provider_response_body_metadata,
+        attach_provider_response_model_metadata, build_usage_request_metadata_seed,
+        merge_usage_request_metadata, merge_usage_request_metadata_owned,
+        refresh_provider_response_body_metadata, refresh_provider_response_model_metadata,
         retain_first_byte_request_metadata, sanitize_usage_request_metadata,
         sanitize_usage_request_metadata_ref,
     };
@@ -802,6 +883,44 @@ mod tests {
         );
     }
 
+    #[test]
+    fn response_model_metadata_is_independent_from_mapping_and_clears_stale_values() {
+        let metadata = attach_provider_response_model_metadata(
+            Some(json!({"provider_response_model": "old-model", "trace_id": "trace-1"})),
+            Some(&json!({"model": "gpt-5"})),
+            Some(UsageBodyCaptureState::Inline),
+            Some("openai:chat"),
+            Some(&json!({"model": "gpt-5.1"})),
+            Some(UsageBodyCaptureState::Inline),
+            Some("openai:chat"),
+        )
+        .expect("response model should be attached");
+        assert_eq!(metadata["provider_response_model"], "gpt-5.1");
+        assert_eq!(metadata["trace_id"], "trace-1");
+
+        let metadata = refresh_provider_response_model_metadata(
+            Some(json!({"provider_response_model": "gpt-5.1"})),
+            Some(&json!({"model": "gpt-5"})),
+            Some(UsageBodyCaptureState::Inline),
+            Some("openai:chat"),
+            Some(&json!({"model": "gpt-5"})),
+            Some(UsageBodyCaptureState::Inline),
+            Some("openai:chat"),
+        );
+        assert!(metadata.is_none());
+
+        let metadata = refresh_provider_response_model_metadata(
+            Some(json!({"provider_response_model": "gpt-5.1"})),
+            None,
+            Some(UsageBodyCaptureState::Disabled),
+            Some("openai:chat"),
+            Some(&json!({"model": "gpt-5.2"})),
+            Some(UsageBodyCaptureState::Inline),
+            Some("openai:chat"),
+        );
+        assert!(metadata.is_none());
+    }
+
     #[test]
     fn terminal_response_refresh_replaces_stale_actual_tier() {
         let metadata = refresh_provider_response_body_metadata(

crates/aether-usage/runtime/src/runtime.rs

@@ -21,9 +21,10 @@ use crate::executor::spawn_on_usage_background_runtime;
 use crate::queue::is_permanent_enqueue_error;
 use crate::request_metadata::{
     attach_client_request_body_metadata, attach_provider_request_body_metadata,
-    attach_provider_response_body_metadata, clear_client_request_body_metadata,
-    clear_provider_request_body_metadata, request_body_derived_facts_action,
-    retain_first_byte_request_metadata, RequestBodyDerivedFactsAction,
+    attach_provider_response_body_metadata, attach_provider_response_model_metadata,
+    clear_client_request_body_metadata, clear_provider_request_body_metadata,
+    request_body_derived_facts_action, retain_first_byte_request_metadata,
+    RequestBodyDerivedFactsAction,
 };
 use crate::settlement::{
     reconcile_usage_policy_cost_for_event_with_result, settle_usage_with_reconciled_cost,
@@ -5297,8 +5298,17 @@ fn preserve_request_facts_with_legacy_missing(
 
 fn preserve_provider_response_facts(event: &mut UsageEvent) {
     let metadata = event.data.request_metadata.take();
-    event.data.request_metadata =
+    let metadata =
         attach_provider_response_body_metadata(metadata, event.data.response_body.as_ref());
+    event.data.request_metadata = attach_provider_response_model_metadata(
+        metadata,
+        event.data.request_body.as_ref(),
+        event.data.request_body_state,
+        event.data.api_format.as_deref(),
+        event.data.response_body.as_ref(),
+        event.data.response_body_state,
+        event.data.endpoint_api_format.as_deref(),
+    );
 }
 
 impl UsageQueueHealthSnapshot {

crates/aether-usage/runtime/src/write.rs

@@ -17,9 +17,10 @@ use crate::body_capture::{
 };
 use crate::request_metadata::{
     attach_client_request_body_metadata, attach_provider_actual_service_tier_metadata,
-    attach_provider_request_body_metadata, build_usage_request_metadata_seed,
-    merge_usage_request_metadata, merge_usage_request_metadata_owned,
-    refresh_provider_response_body_metadata, sanitize_usage_request_metadata,
+    attach_provider_request_body_metadata, attach_provider_response_model_metadata,
+    build_usage_request_metadata_seed, merge_usage_request_metadata,
+    merge_usage_request_metadata_owned, refresh_provider_response_body_metadata,
+    refresh_provider_response_model_metadata, sanitize_usage_request_metadata,
     sanitize_usage_request_metadata_ref,
 };
 use crate::{
@@ -724,6 +725,15 @@ fn build_terminal_usage_event_from_seed_impl(
         Some(model.as_str()),
         provider_request.as_ref(),
     );
+    let request_metadata = attach_provider_response_model_metadata(
+        request_metadata,
+        request_body.as_ref(),
+        body_states.request_body_state,
+        Some(client_contract.as_str()),
+        provider_response.as_ref(),
+        body_states.response_body_state,
+        Some(provider_contract.as_str()),
+    );
 
     let mut data = UsageEventData {
         user_id,
@@ -1047,6 +1057,15 @@ pub fn build_sync_terminal_usage_seed(
         context_seed.request_metadata,
         provider_response_full.as_ref(),
     );
+    let request_metadata = refresh_provider_response_model_metadata(
+        request_metadata,
+        context_seed.request_body.as_ref(),
+        context_seed.body_states.request_body_state,
+        Some(context_seed.client_contract.as_str()),
+        provider_response_full.as_ref(),
+        provider_response_body_state,
+        Some(context_seed.provider_contract.as_str()),
+    );
 
     TerminalUsageSeed {
         terminal_state,
@@ -1225,6 +1244,15 @@ pub fn build_stream_terminal_usage_seed(
         context_seed.request_metadata,
         provider_response_full.as_ref(),
     );
+    let request_metadata = refresh_provider_response_model_metadata(
+        request_metadata,
+        context_seed.request_body.as_ref(),
+        context_seed.body_states.request_body_state,
+        Some(context_seed.client_contract.as_str()),
+        provider_response_full.as_ref(),
+        provider_response_body_state,
+        Some(context_seed.provider_contract.as_str()),
+    );
     // The parser's terminal summary is authoritative when a response body is truncated or the
     // body and summary disagree; attach it after the body refresh so it wins.
     let request_metadata = attach_provider_actual_service_tier_metadata(

frontend/src/api/dashboard.ts

@@ -219,6 +219,7 @@ export interface RequestDetail {
   has_format_conversion?: boolean | null
   model: string
   target_model?: string | null  // 映射后的目标模型名
+  response_model?: string | null  // 上游响应体实际返回的模型名
   requested_reasoning_effort?: string | null
   reasoning_effort?: string | null
   service_tier?: string | null

frontend/src/api/me.ts

@@ -60,6 +60,7 @@ export interface UsageRecordDetail {
   reasoning_effort?: string | null
   service_tier?: string | null
   actual_service_tier?: string | null
+  response_model?: string | null
   input_tokens: number
   effective_input_tokens?: number
   output_tokens: number
@@ -392,6 +393,7 @@ export const meApi = {
       has_format_conversion?: boolean | null
       has_fallback?: boolean | null
       target_model?: string | null
+      response_model?: string | null
       request_type?: string | null
       requested_reasoning_effort?: string | null
       reasoning_effort?: string | null
@@ -438,6 +440,7 @@ export const meApi = {
       has_format_conversion?: boolean | null
       has_fallback?: boolean | null
       target_model?: string | null
+      response_model?: string | null
       request_type?: string | null
       requested_reasoning_effort?: string | null
       reasoning_effort?: string | null

frontend/src/api/usage.ts

@@ -20,6 +20,7 @@ export interface UsageRecord {
   reasoning_effort?: string | null
   service_tier?: string | null
   actual_service_tier?: string | null
+  response_model?: string | null
   input_tokens: number
   effective_input_tokens?: number
   output_tokens: number
@@ -618,6 +619,7 @@ export const usageApi = {
       has_format_conversion?: boolean | null
       has_fallback?: boolean | null
       target_model?: string | null
+      response_model?: string | null
       request_type?: string | null
       requested_reasoning_effort?: string | null
       reasoning_effort?: string | null
@@ -685,6 +687,7 @@ export const usageApi = {
       has_format_conversion?: boolean | null
       has_fallback?: boolean | null
       target_model?: string | null
+      response_model?: string | null
       request_type?: string | null
       requested_reasoning_effort?: string | null
       reasoning_effort?: string | null

frontend/src/api/usageRecords.ts

@@ -19,6 +19,7 @@ export interface UsageRecord {
   model: string
   target_model?: string | null  // 映射后的目标模型名（若无映射则为空）
   model_version?: string | null  // Provider 返回的实际模型版本（列表轻量字段）
+  response_model?: string | null  // 上游响应体实际返回的模型名
   request_type?: string | null  // 由请求语义识别出的操作类型
   requested_reasoning_effort?: string | null  // 用户请求侧 reasoning 级别，用于展示转换关系
   reasoning_effort?: string | null  // 从发送给 Provider 的请求体提取的 reasoning 级别

frontend/src/features/usage/components/RequestDetailDrawer.vue

@@ -1016,6 +1016,7 @@ const emit = defineEmits<{
     reasoningEffort?: string | null
     serviceTier?: string | null
     actualServiceTier?: string | null
+    responseModel?: string | null
     imageProgress?: ImageProgress | null
     errorMessage?: string | null
     updatedAt?: string | null
@@ -1158,6 +1159,7 @@ type HeaderModelTextField =
   | 'model'
   | 'target_model'
   | 'model_version'
+  | 'response_model'
   | 'request_type'
   | 'requested_reasoning_effort'
   | 'reasoning_effort'
@@ -1166,6 +1168,7 @@ type HeaderModelTextField =
 
 const FINAL_PROVIDER_HEADER_FIELDS = new Set<HeaderModelTextField>([
   'target_model',
+  'response_model',
   'reasoning_effort',
   'service_tier',
   'actual_service_tier',
@@ -1252,6 +1255,7 @@ watch(
     props.summaryRecord?.model,
     props.summaryRecord?.target_model,
     props.summaryRecord?.model_version,
+    props.summaryRecord?.response_model,
     props.summaryRecord?.request_type,
     props.summaryRecord?.requested_reasoning_effort,
     props.summaryRecord?.reasoning_effort,
@@ -1298,6 +1302,7 @@ function emitDetailRequestState(nextDetail: RequestDetail) {
   const reasoningEffort = resolveHeaderModelTextField('reasoning_effort', nextDetail)
   const serviceTier = resolveHeaderModelTextField('service_tier', nextDetail)
   const actualServiceTier = resolveHeaderModelTextField('actual_service_tier', nextDetail)
+  const responseModel = resolveHeaderModelTextField('response_model', nextDetail)
 
   emit('requestState', {
     id,
@@ -1334,6 +1339,7 @@ function emitDetailRequestState(nextDetail: RequestDetail) {
     ...(reasoningEffort ? { reasoningEffort } : {}),
     ...(serviceTier ? { serviceTier } : {}),
     ...(actualServiceTier ? { actualServiceTier } : {}),
+    ...(responseModel ? { responseModel } : {}),
     errorMessage: nextDetail.error_message ?? undefined,
     updatedAt: nextDetail.updated_at ?? undefined,
   })
@@ -1610,7 +1616,7 @@ const headerModelRecord = computed(() => {
   return {
     model: resolveHeaderModelTextField('model', currentDetail) ?? '-',
     target_model: resolveHeaderModelTextField('target_model', currentDetail),
-    model_version: resolveHeaderModelTextField('model_version', currentDetail),
+    response_model: resolveHeaderModelTextField('response_model', currentDetail),
     request_type: resolveHeaderModelTextField('request_type', currentDetail),
     requested_reasoning_effort: resolveHeaderModelTextField(
       'requested_reasoning_effort',

frontend/src/features/usage/components/UsageModelDisplay.vue

@@ -16,11 +16,22 @@
         :class="modelClass"
         data-usage-model-source
       >{{ record.model }}</span>
-      <template v-if="actualModel">
-        <span
-          class="order-last basis-full min-w-0 break-all whitespace-normal text-muted-foreground"
-          data-usage-model-target
-        ><span class="mr-1">-&gt;</span>{{ actualModel }}</span>
+      <template v-if="hasModelFacts">
+        <div
+          class="order-last basis-full flex min-w-0 flex-wrap items-center gap-x-2 gap-y-0.5 text-muted-foreground"
+          data-usage-model-facts
+        >
+          <span
+            v-if="mappingModel"
+            class="min-w-0 break-all whitespace-normal"
+            data-usage-model-mapping
+          ><span class="mr-1 text-[10px] text-muted-foreground/70">映射模型</span>{{ mappingModel }}</span>
+          <span
+            v-if="responseModel"
+            class="min-w-0 break-all whitespace-normal"
+            data-usage-model-response
+          ><span class="mr-1 text-[10px] text-muted-foreground/70">响应模型</span>{{ responseModel }}</span>
+        </div>
       </template>
       <template v-if="!shouldStackBadges">
         <Badge
@@ -83,7 +94,7 @@ interface ModelBadgePresentation {
 interface UsageModelDisplayRecord {
   model: string
   target_model?: string | null
-  model_version?: string | null
+  response_model?: string | null
   request_type?: string | null
   requested_reasoning_effort?: string | null
   reasoning_effort?: string | null
@@ -113,15 +124,20 @@ const props = withDefaults(defineProps<{
   showReasoningBadge: true,
 })
 
-const actualModel = computed(() => {
+const mappingModel = computed(() => {
   const targetModel = normalizeText(props.record.target_model)
-  if (targetModel && targetModel !== props.record.model) return targetModel
-
-  const modelVersion = normalizeText(props.record.model_version)
-  if (modelVersion && modelVersion !== props.record.model) return modelVersion
+  if (targetModel && targetModel !== normalizeText(props.record.model)) return targetModel
   return null
 })
 
+const responseModel = computed(() => {
+  const response = normalizeText(props.record.response_model)
+  if (response && response !== normalizeText(props.record.model)) return response
+  return null
+})
+
+const hasModelFacts = computed(() => mappingModel.value !== null || responseModel.value !== null)
+
 const reasoningLabel = computed(() => {
   const requested = normalizeText(props.record.requested_reasoning_effort)
   const actual = normalizeText(props.record.reasoning_effort)
@@ -191,7 +207,7 @@ const modelBadges = computed<ModelBadgePresentation[]>(() => {
 })
 
 const shouldStackBadges = computed(() => (
-  actualModel.value === null && modelBadges.value.length >= 3
+  !hasModelFacts.value && modelBadges.value.length >= 3
 ))
 
 function normalizeText(value: string | null | undefined): string | null {

frontend/src/features/usage/components/UsageRecordsTable.vue

@@ -1656,20 +1656,6 @@ function getApiFormatTooltip(record: UsageRecord): string {
   return displayFormat
 }
 
-// 获取实际使用的模型（优先 target_model，其次列表接口下发的 model_version）
-// 只有当实际模型与请求模型不同时才返回，用于显示映射箭头
-function getActualModel(record: UsageRecord): string | null {
-  // 优先显示模型映射
-  if (record.target_model && record.target_model !== record.model) {
-    return record.target_model
-  }
-  // 其次显示 Provider 返回的实际版本（如 Gemini 的 modelVersion）
-  if (record.model_version && record.model_version !== record.model) {
-    return record.model_version
-  }
-  return null
-}
-
 function getReasoningEffort(record: UsageRecord): string | null {
   const requested = record.requested_reasoning_effort?.trim()
   const actual = record.reasoning_effort?.trim()
@@ -1744,15 +1730,18 @@ function getServiceTierTitle(record: UsageRecord): string {
 
 // 获取模型列的 tooltip
 function getModelTooltip(record: UsageRecord): string {
-  const actualModel = getActualModel(record)
   const reasoningEffort = getReasoningEffort(record)
   const serviceTierTitle = getServiceTierTitle(record)
   const tierSuffix = serviceTierTitle ? `\n${serviceTierTitle}` : ''
   const cyberSuffix = hasCyberPolicyError(record) ? '\nCyber Policy: blocked' : ''
   const suffix = `${reasoningEffort ? `\nReasoning: ${reasoningEffort}` : ''}${tierSuffix}${cyberSuffix}`
-  if (actualModel) {
-    return `${record.model} -> ${actualModel}${suffix}`
-  }
-  return `${record.model}${suffix}`
+  const requestModel = record.model.trim()
+  const mappingModel = record.target_model?.trim()
+  const responseModel = record.response_model?.trim()
+  return [
+    requestModel,
+    mappingModel && mappingModel !== requestModel ? `映射模型: ${mappingModel}` : null,
+    responseModel && responseModel !== requestModel ? `响应模型: ${responseModel}` : null,
+  ].filter((line): line is string => Boolean(line)).join('\n') + suffix
 }
 </script>

frontend/src/features/usage/components/__tests__/RequestDetailDrawer.pricing.spec.ts

@@ -313,13 +313,14 @@ describe('RequestDetailDrawer settlement pricing', () => {
     { tier: 'priority', label: 'Fast', badgeKey: 'fast' },
     { tier: 'ultrafast', label: 'ultrafast', badgeKey: 'service-tier' },
     { tier: 'future-tier', label: 'future-tier', badgeKey: 'service-tier' },
-  ])('shows mapping, reasoning, $tier, and Cyber together in the model header', async ({ tier, label, badgeKey }) => {
+  ])('shows mapping, response model, reasoning, $tier, and Cyber together in the model header', async ({ tier, label, badgeKey }) => {
     apiMocks.getRequestDetail.mockResolvedValue({
       ...buildEmbeddingDetail(),
       id: 'usage-cyber-risk-demo',
       request_id: 'req_usage-cyber-risk-demo',
       model: 'gpt-5',
       target_model: 'gpt-5.1',
+      response_model: 'gpt-5.2',
       requested_reasoning_effort: 'xhigh',
       reasoning_effort: 'max',
       request_body: {
@@ -382,10 +383,11 @@ describe('RequestDetailDrawer settlement pricing', () => {
       )
       const modelRow = modelLayout?.firstElementChild
       expect(modelRow?.textContent).toContain('gpt-5')
-      expect(modelRow?.textContent).toContain('->')
-      expect(modelRow?.textContent).toContain('gpt-5.1')
-      expect(modelRow?.querySelector('[data-usage-model-target]')?.classList.contains('basis-full'))
-        .toBe(true)
+      expect(modelRow?.classList.contains('flex-wrap')).toBe(true)
+      expect(modelRow?.querySelector('[data-usage-model-mapping]')?.textContent)
+        .toContain('gpt-5.1')
+      expect(modelRow?.querySelector('[data-usage-model-response]')?.textContent)
+        .toContain('gpt-5.2')
       expect(modelRow?.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent)
         .toContain('xhigh -> max')
       expect(modelRow?.querySelector(`[data-request-detail-model-badge="${badgeKey}"]`)?.textContent)
@@ -462,7 +464,7 @@ describe('RequestDetailDrawer settlement pricing', () => {
     await vi.waitFor(() => {
       expect(document.body.querySelector('[data-request-detail-model-badge="cyber"]')?.textContent)
         .toContain('Cyber')
-      expect(document.body.querySelector('[data-usage-model-target]')?.textContent)
+      expect(document.body.querySelector('[data-usage-model-mapping]')?.textContent)
         .toContain('gpt-5.1')
       expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent)
         .toContain('xhigh -> max')
@@ -573,8 +575,8 @@ describe('RequestDetailDrawer settlement pricing', () => {
 
     await vi.waitFor(() => {
       expect(apiMocks.getRequestDetail).toHaveBeenCalledTimes(1)
-      expect(document.body.querySelector('[data-usage-model-target]')?.textContent?.trim())
-        .toBe('->gpt-5.1')
+      expect(document.body.querySelector('[data-usage-model-mapping]')?.textContent?.trim())
+        .toBe('映射模型gpt-5.1')
       expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent?.trim())
         .toBe('xhigh -> max')
       expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')).toBeNull()
@@ -585,7 +587,7 @@ describe('RequestDetailDrawer settlement pricing', () => {
     })
   })
 
-  it('uses detail model_version when the lightweight summary has null model facts', async () => {
+  it('does not use legacy model_version as a response-model fallback', async () => {
     apiMocks.getRequestDetail.mockResolvedValue({
       ...buildEmbeddingDetail(),
       id: 'usage-version-summary',
@@ -628,8 +630,8 @@ describe('RequestDetailDrawer settlement pricing', () => {
     await nextTick()
 
     await vi.waitFor(() => {
-      expect(document.body.querySelector('[data-usage-model-target]')?.textContent?.trim())
-        .toBe('->gpt-5.1-2026-07-17')
+      expect(document.body.querySelector('[data-usage-model-response]')).toBeNull()
+      expect(document.body.querySelector('[data-usage-model-mapping]')).toBeNull()
     })
   })
 
@@ -689,7 +691,7 @@ describe('RequestDetailDrawer settlement pricing', () => {
     }
     await nextTick()
 
-    expect(document.body.querySelector('[data-usage-model-target]')).toBeNull()
+    expect(document.body.querySelector('[data-usage-model-mapping]')).toBeNull()
     expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent?.trim())
       .toBe('xhigh')
     expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')).toBeNull()

frontend/src/features/usage/components/__tests__/UsageRecordsTable.spec.ts

@@ -461,10 +461,11 @@ describe('UsageRecordsTable', () => {
       .toBe('会话压缩')
   })
 
-  it('shows mapping, reasoning, Fast, and Cyber in the model area', () => {
+  it('shows mapping and response models as separately labelled facts', () => {
     const root = mountUsageRecordsTable([buildRecord({
       model: 'gpt-5',
       target_model: 'gpt-5.1',
+      response_model: 'gpt-5.2',
       requested_reasoning_effort: 'xhigh',
       reasoning_effort: 'max',
       service_tier: 'priority',
@@ -477,6 +478,8 @@ describe('UsageRecordsTable', () => {
 
     expect(root.textContent).toContain('gpt-5')
     expect(root.textContent).toContain('gpt-5.1')
+    expect(root.textContent).toContain('映射模型')
+    expect(root.textContent).toContain('响应模型')
     expect(root.textContent).toContain('xhigh -> max')
     expect(root.textContent).toContain('Fast')
     const reasoningBadge = root.querySelector<HTMLElement>('[data-usage-model-badge="reasoning"]')
@@ -510,15 +513,35 @@ describe('UsageRecordsTable', () => {
     expect(inlineLayout).not.toBeNull()
     const modelRow = inlineLayout?.firstElementChild
     expect(modelRow?.textContent).toContain('gpt-5')
-    expect(modelRow?.textContent).toContain('->')
-    expect(modelRow?.textContent).toContain('gpt-5.1')
-    expect(modelRow?.querySelector('[data-usage-model-target]')?.classList.contains('basis-full')).toBe(true)
-    expect(modelRow?.querySelector('[data-usage-model-target]')?.classList.contains('order-last')).toBe(true)
+    expect(modelRow?.querySelector('[data-usage-model-mapping]')?.textContent).toContain('映射模型')
+    expect(modelRow?.querySelector('[data-usage-model-response]')?.textContent).toContain('响应模型')
     expect(modelRow?.querySelector('[data-usage-model-badge="reasoning"]')?.textContent).toContain('xhigh -> max')
     expect(modelRow?.querySelector('[data-usage-model-badge="fast"]')?.textContent).toContain('Fast')
     expect(modelRow?.querySelector('[data-usage-model-badge="cyber"]')?.textContent).toContain('Cyber')
   })
 
+  it('shows a response model without inventing a mapping arrow or using model_version', () => {
+    const root = mountUsageRecordsTable([buildRecord({
+      response_model: 'gpt-5.1',
+      model_version: 'legacy-version',
+      target_model: null,
+    })])
+
+    expect(root.querySelector('[data-usage-model-mapping]')).toBeNull()
+    expect(root.querySelector('[data-usage-model-response]')?.textContent).toContain('响应模型')
+    expect(root.textContent).not.toContain('legacy-version')
+    expect(root.textContent).not.toContain('->')
+  })
+
+  it('hides a response model when it matches the request model', () => {
+    const root = mountUsageRecordsTable([buildRecord({
+      response_model: 'gpt-5',
+      target_model: null,
+    })])
+
+    expect(root.querySelector('[data-usage-model-response]')).toBeNull()
+  })
+
   it('stacks three model badges even without a model mapping', () => {
     const root = mountUsageRecordsTable([buildRecord({
       model: 'gpt-5',

frontend/src/features/usage/composables/useUsageData.ts

@@ -716,7 +716,13 @@ export function useUsageData(options: UseUsageDataOptions) {
           ? (typeof record.actual_service_tier === 'string' && record.actual_service_tier.trim()
               ? record.actual_service_tier
               : null)
-          : existing.actual_service_tier
+          : existing.actual_service_tier,
+        // 终态列表快照是最终候选的权威事实；空值必须清除旧候选的响应模型。
+        response_model: statusProgressed
+          ? (typeof record.response_model === 'string' && record.response_model.trim()
+              ? record.response_model
+              : null)
+          : existing.response_model
       }
     })
   }

frontend/src/views/shared/Usage.vue

@@ -651,6 +651,10 @@ async function pollActiveRequests() {
         record.actual_service_tier = typeof update.actual_service_tier === 'string' && update.actual_service_tier.trim()
           ? update.actual_service_tier
           : null
+        // 活跃接口返回的是当前最终候选快照，服务端空值要清除旧响应模型。
+        record.response_model = typeof update.response_model === 'string' && update.response_model.trim()
+          ? update.response_model
+          : null
         // 管理员接口返回额外字段
         // 只有当返回的 provider 不是 pending/unknown/unknow 时才更新，避免覆盖已有的正确值
         if ('provider' in update && typeof update.provider === 'string') {
@@ -1150,6 +1154,7 @@ function handleDetailRequestState(update: {
   reasoningEffort?: string | null
   serviceTier?: string | null
   actualServiceTier?: string | null
+  responseModel?: string | null
   imageProgress?: ImageProgress | null
   errorMessage?: string | null
   updatedAt?: string | null
@@ -1292,6 +1297,11 @@ function handleDetailRequestState(update: {
       ? update.actualServiceTier
       : null
   }
+  if ('responseModel' in update) {
+    record.response_model = typeof update.responseModel === 'string'
+      ? update.responseModel
+      : null
+  }
   if ('imageProgress' in update) {
     const nextProgress = update.imageProgress ?? null
     if (!sameImageProgress(record.image_progress, nextProgress)) {