diff --git a/backend/crates/gateway-protocol/src/openai/events.rs b/backend/crates/gateway-protocol/src/openai/events.rs
index 4a5a55009..1fd6dae46 100644
--- a/backend/crates/gateway-protocol/src/openai/events.rs
+++ b/backend/crates/gateway-protocol/src/openai/events.rs
@@ -65,7 +65,8 @@ pub fn extract_usage(body: &Value) -> Option<TokenUsage> {
     let image_output_tokens =
         nested_number_field(body, &["tool_usage", "image_gen", "output_tokens"])
             .unwrap_or_default();
-    let total_tokens = number_field(usage, "total_tokens").unwrap_or(input_tokens + output_tokens);
+    let total_tokens = number_field(usage, "total_tokens")
+        .unwrap_or_else(|| input_tokens.saturating_add(output_tokens));
 
     let has_usage = [
         "input_tokens",
@@ -98,7 +99,7 @@ pub fn extract_usage(body: &Value) -> Option<TokenUsage> {
     })
 }
 
-/// 校验计价所需的输入、输出和缓存读取字段，拒绝缺失、非法或不一致的数值。
+/// 校验计价所需的 token 字段，拒绝缺失、非法、溢出或不一致的数值。
 ///
 /// 用量提取允许缺失字段补零；计价必须额外确认这些零值有明确的上游事实。
 #[must_use]
@@ -122,11 +123,28 @@ pub fn billable_usage_is_complete(response: &Value, usage: TokenUsage) -> bool {
         Some(value) => value.as_u64(),
         None => Some(0),
     };
+    let written = raw
+        .pointer("/input_tokens_details/cache_write_tokens")
+        .or_else(|| raw.pointer("/prompt_tokens_details/cache_write_tokens"))
+        .or_else(|| raw.get("cache_write_tokens"));
+    let written = match written {
+        Some(value) => value.as_u64(),
+        None => Some(0),
+    };
+    let total = usage.input_tokens.checked_add(usage.output_tokens);
 
     input == Some(usage.input_tokens)
         && output == Some(usage.output_tokens)
         && cached == Some(usage.cached_tokens)
-        && usage.cached_tokens <= usage.input_tokens
+        && written == Some(usage.cache_write_tokens)
+        && usage
+            .cached_tokens
+            .checked_add(usage.cache_write_tokens)
+            .is_some_and(|cached| cached <= usage.input_tokens)
+        && total == Some(usage.total_tokens)
+        && raw
+            .get("total_tokens")
+            .is_none_or(|value| value.as_u64() == total)
 }
 
 /// 从完整 SSE 文本中提取最终可见用量。
diff --git a/backend/crates/gateway-protocol/tests/openai/events.rs b/backend/crates/gateway-protocol/tests/openai/events.rs
index e66ac3108..306cc9fbd 100644
--- a/backend/crates/gateway-protocol/tests/openai/events.rs
+++ b/backend/crates/gateway-protocol/tests/openai/events.rs
@@ -5,6 +5,57 @@ use gateway_protocol::openai::events::{
 };
 use serde_json::json;
 
+#[test]
+fn billable_usage_should_validate_cache_writes_totals_and_overflow() {
+    for (raw, expected) in [
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "input_tokens_details": {"cached_tokens": 20, "cache_write_tokens": 80}, "total_tokens": 110}),
+            true,
+        ),
+        (
+            json!({"prompt_tokens": 100, "completion_tokens": 10, "prompt_tokens_details": {"cached_tokens": 20, "cache_write_tokens": 80}}),
+            true,
+        ),
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "cached_tokens": 20, "cache_write_tokens": 90}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "cache_write_tokens": -1}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "input_tokens_details": {"cache_write_tokens": "1"}, "cache_write_tokens": 1}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "cache_write_tokens": 1.5}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "total_tokens": 111}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 100, "output_tokens": 10, "total_tokens": "110"}),
+            false,
+        ),
+        (json!({"input_tokens": u64::MAX, "output_tokens": 1}), false),
+        (
+            json!({"input_tokens": u64::MAX, "output_tokens": 0, "cached_tokens": u64::MAX, "cache_write_tokens": 1}),
+            false,
+        ),
+    ] {
+        let response = json!({"usage": raw});
+        let usage = extract_usage(&response).expect("提取用量不得因溢出而崩溃");
+        assert_eq!(
+            billable_usage_is_complete(&response, usage),
+            expected,
+            "{response}"
+        );
+    }
+}
+
 #[test]
 fn billable_usage_should_require_valid_input_output_and_cache_facts() {
     for (raw, expected) in [
diff --git a/backend/crates/providers/openai/src/provider/observation.rs b/backend/crates/providers/openai/src/provider/observation.rs
index 9a1a8e428..1eff055f7 100644
--- a/backend/crates/providers/openai/src/provider/observation.rs
+++ b/backend/crates/providers/openai/src/provider/observation.rs
@@ -145,9 +145,9 @@ impl OpenAiResponseObservationState {
     }
 
     pub(super) fn effective_service_tier(&self) -> Option<&str> {
-        self.requested_service_tier
+        self.upstream_service_tier
             .as_deref()
-            .or(self.upstream_service_tier.as_deref())
+            .or(self.requested_service_tier.as_deref())
     }
 
     pub(super) fn merge_rate_limit_headers(&mut self, updates: &[(String, String)]) -> bool {
diff --git a/backend/crates/providers/openai/src/transport/canonical.rs b/backend/crates/providers/openai/src/transport/canonical.rs
index 6242b3f4c..7e542305f 100644
--- a/backend/crates/providers/openai/src/transport/canonical.rs
+++ b/backend/crates/providers/openai/src/transport/canonical.rs
@@ -158,7 +158,7 @@ impl CodexCanonicalDecoder {
         self
     }
 
-    /// 请求显式指定的档位优先用于计费；响应档位只在请求未指定时兜底。
+    /// 响应未报告实际档位时，才使用请求档位估算费用。
     #[must_use]
     pub fn with_requested_service_tier(mut self, service_tier: Option<&str>) -> Self {
         self.requested_service_tier = normalize_service_tier(service_tier);
@@ -862,18 +862,19 @@ impl CodexCanonicalDecoder {
             .unwrap_or(&self.fallback_model)
             .to_owned();
         let service_tier = self
-            .requested_service_tier
+            .response_service_tier
             .as_deref()
-            .or(self.response_service_tier.as_deref());
-        let web_search_calls = web_search_call_count(response);
+            .or(self.requested_service_tier.as_deref());
+        let tool_calls = billable_tool_calls(response);
         if let Some(breakdown) = usage
             .filter(|usage| billable_usage_is_complete(response, *usage))
             .and_then(|usage| {
-                let web_search_calls = web_search_calls?;
+                let (web_search_calls, file_search_calls) = tool_calls?;
                 openai_billing_breakdown(
                     &model,
                     OpenAiBillingUsage::from(usage)
-                        .with_web_search_calls(web_search_calls, self.web_search_pricing),
+                        .with_web_search_calls(web_search_calls, self.web_search_pricing)
+                        .with_file_search_calls(file_search_calls),
                     service_tier,
                 )
             })
@@ -1000,26 +1001,43 @@ fn core_usage(usage: TokenUsage) -> Usage {
     normalized
 }
 
-fn web_search_call_count(response: &Value) -> Option<u64> {
-    let mut calls = 0_u64;
-    for item in response
-        .get("output")
-        .and_then(Value::as_array)
-        .into_iter()
-        .flatten()
-        .filter(|item| item.get("type").and_then(Value::as_str) == Some("web_search_call"))
-    {
-        match item
-            .get("action")
-            .and_then(|action| action.get("type"))
-            .and_then(Value::as_str)
-        {
-            Some("search") => calls = calls.checked_add(1)?,
-            Some("open_page" | "find_in_page") => {}
-            Some(_) | None => return None,
+fn billable_tool_calls(response: &Value) -> Option<(u64, u64)> {
+    let mut web = 0_u64;
+    let mut file = 0_u64;
+    let items = match response.get("output") {
+        Some(value) => value.as_array()?.as_slice(),
+        None => &[],
+    };
+    for item in items {
+        match item.get("type").and_then(Value::as_str)? {
+            "web_search_call" => match item.pointer("/action/type").and_then(Value::as_str) {
+                Some("search") => web = web.checked_add(1)?,
+                Some("open_page" | "find_in_page") => {}
+                _ => return None,
+            },
+            "file_search_call"
+                if item.get("status").and_then(Value::as_str) == Some("completed") =>
+            {
+                file = file.checked_add(1)?;
+            }
+            // 这些输出没有独立的 OpenAI 工具调用费。
+            "message"
+            | "reasoning"
+            | "function_call"
+            | "custom_tool_call"
+            | "computer_call"
+            | "local_shell_call"
+            | "apply_patch_call"
+            | "mcp_call"
+            | "mcp_list_tools"
+            | "mcp_approval_request"
+            | "compaction" => {}
+            // 容器会话、生图及未知工具需要额外计费事实，不能把 token 小计
+            // 当作完整费用。
+            _ => return None,
         }
     }
-    Some(calls)
+    Some((web, file))
 }
 
 fn incomplete_finish_reason(response: &Value) -> FinishReason {
diff --git a/backend/crates/providers/openai/src/transport/usage.rs b/backend/crates/providers/openai/src/transport/usage.rs
index 00baab361..f66f90631 100644
--- a/backend/crates/providers/openai/src/transport/usage.rs
+++ b/backend/crates/providers/openai/src/transport/usage.rs
@@ -16,6 +16,7 @@ use super::{
 const LONG_CONTEXT_THRESHOLD: u64 = 272_000;
 const WEB_SEARCH_CALL_TICKS: u128 = 100_000_000;
 const WEB_SEARCH_PREVIEW_NON_REASONING_CALL_TICKS: u128 = 250_000_000;
+const FILE_SEARCH_CALL_TICKS: u128 = 25_000_000;
 
 /// OpenAI 公开 Token 价格计算所需的单次用量事实。
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
@@ -27,6 +28,7 @@ pub struct OpenAiBillingUsage {
     image_input_tokens: u64,
     image_output_tokens: u64,
     web_search_calls: u64,
+    file_search_calls: u64,
     web_search_pricing: Option<WebSearchPricing>,
 }
 
@@ -47,6 +49,7 @@ impl OpenAiBillingUsage {
             image_input_tokens: 0,
             image_output_tokens: 0,
             web_search_calls: 0,
+            file_search_calls: 0,
             web_search_pricing: None,
         }
     }
@@ -60,6 +63,11 @@ impl OpenAiBillingUsage {
         self.web_search_pricing = pricing;
         self
     }
+
+    pub(crate) const fn with_file_search_calls(mut self, calls: u64) -> Self {
+        self.file_search_calls = calls;
+        self
+    }
 }
 
 impl From<TokenUsage> for OpenAiBillingUsage {
@@ -72,6 +80,7 @@ impl From<TokenUsage> for OpenAiBillingUsage {
             image_input_tokens: usage.image_input_tokens,
             image_output_tokens: usage.image_output_tokens,
             web_search_calls: 0,
+            file_search_calls: 0,
             web_search_pricing: None,
         }
     }
@@ -125,6 +134,7 @@ struct ModelPricing {
     long_flex: TokenRates,
     long_fast: TokenRates,
     cache_write_percent: u32,
+    unpriced_long_context: bool,
 }
 
 impl ModelPricing {
@@ -137,6 +147,7 @@ impl ModelPricing {
             long_flex: TokenRates::ZERO,
             long_fast: TokenRates::ZERO,
             cache_write_percent: 0,
+            unpriced_long_context: false,
         }
     }
 
@@ -170,7 +181,15 @@ impl ModelPricing {
         self
     }
 
+    const fn with_unpriced_long_context(mut self) -> Self {
+        self.unpriced_long_context = true;
+        self
+    }
+
     fn rates(self, tier: PricingTier, long_context: bool) -> Option<TokenRates> {
+        if long_context && self.unpriced_long_context {
+            return None;
+        }
         // Only models with a published long-context column switch at the threshold.
         // A dash in that column is not a second, unavailable price tier for models
         // whose entire supported context is covered by the short-context price.
@@ -200,10 +219,12 @@ struct PricingRule {
     pricing: ModelPricing,
 }
 
-// Source: https://developers.openai.com/api/docs/pricing (verified 2026-08-14).
+// 价格来源：https://developers.openai.com/api/docs/pricing，核验日期 2026-09-09。
+// 使用常规价，不采用 Sol 的临时 $4/$20 优惠；缓存、Flex、Fast 和长上下文
+// 档位也统一按常规价计算。
 const PRICING_RULES: &[PricingRule] = &[
-    // Astra: https://developers.openai.com/api/docs/models/gpt-6-astra
-    // Rates verified against the pricing table on 2026-09-05.
+    // Astra：https://developers.openai.com/api/docs/models/gpt-6-astra
+    // 已于 2026-09-09 对照官方价目表核验。
     PricingRule {
         model: "gpt-6-astra",
         pricing: ModelPricing::new(100_000, 500_000, 10_000)
@@ -405,10 +426,60 @@ const PRICING_RULES: &[PricingRule] = &[
         model: "gpt-3.5-turbo",
         pricing: ModelPricing::new(5_000, 15_000, 0),
     },
+    // 各变体独立采用已公开的价格与档位，不能继承父型号的所有档位。
+    // Cyber 长上下文的官方来源存在差异，该区间暂不估价。
+    PricingRule {
+        model: "gpt-5.1-codex-mini",
+        pricing: ModelPricing::new(2_500, 20_000, 250),
+    },
+    PricingRule {
+        model: "gpt-5.3-chat-latest",
+        pricing: ModelPricing::new(17_500, 140_000, 1_750),
+    },
+    PricingRule {
+        model: "gpt-5.6-cyber",
+        pricing: ModelPricing::new(125_000, 750_000, 12_500)
+            .with_cache_write(125)
+            .with_unpriced_long_context(),
+    },
+    PricingRule {
+        model: "gpt-5.5-cyber",
+        pricing: ModelPricing::new(125_000, 750_000, 12_500).with_unpriced_long_context(),
+    },
+    PricingRule {
+        model: "chat-latest",
+        pricing: ModelPricing::new(50_000, 300_000, 5_000),
+    },
+    PricingRule {
+        model: "gpt-5-codex",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250),
+    },
+    PricingRule {
+        model: "gpt-5.1-codex",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250),
+    },
+    PricingRule {
+        model: "gpt-5.1-codex-max",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250),
+    },
+    PricingRule {
+        model: "gpt-5-chat-latest",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250),
+    },
+    PricingRule {
+        model: "gpt-5.1-chat-latest",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250),
+    },
+    PricingRule {
+        model: "gpt-5.2-codex",
+        pricing: ModelPricing::new(17_500, 140_000, 1_750),
+    },
+    PricingRule {
+        model: "gpt-5.2-chat-latest",
+        pricing: ModelPricing::new(17_500, 140_000, 1_750),
+    },
 ];
 
-const UNPRICED_MODELS: &[&str] = &["gpt-5.3-codex-spark"];
-
 #[derive(Clone, Copy)]
 struct TokenAmounts {
     input_ticks: u128,
@@ -439,13 +510,14 @@ fn openai_billing_breakdown_with_context(
     service_tier: Option<&str>,
     long_context: bool,
 ) -> Option<CalculatedCostBreakdown> {
-    if usage.cached_tokens > usage.input_tokens
+    if usage.cached_tokens.checked_add(usage.cache_write_tokens)? > usage.input_tokens
         || usage.image_input_tokens > 0
         || usage.image_output_tokens > 0
     {
         return None;
     }
-    let web_search_ticks = web_search_amount_ticks(usage)?;
+    let tool_ticks = web_search_amount_ticks(usage)?
+        .checked_add(u128::from(usage.file_search_calls).checked_mul(FILE_SEARCH_CALL_TICKS)?)?;
     let pricing = model_pricing(model)?;
     let normalized_tier = normalize_service_tier(service_tier);
     let tier = pricing_tier(normalized_tier.as_deref())?;
@@ -467,8 +539,8 @@ fn openai_billing_breakdown_with_context(
         usage.cached_tokens,
         usage.cache_write_tokens,
     )?;
-    standard.total_ticks = standard.total_ticks.checked_add(web_search_ticks)?;
-    selected.total_ticks = selected.total_ticks.checked_add(web_search_ticks)?;
+    standard.total_ticks = standard.total_ticks.checked_add(tool_ticks)?;
+    selected.total_ticks = selected.total_ticks.checked_add(tool_ticks)?;
     let multiplier_percent =
         effective_multiplier_percent(selected.total_ticks, standard.total_ticks)?;
     let cache_write_rate = cache_write_rate(selected_rates, pricing.cache_write_percent)?;
@@ -503,7 +575,12 @@ pub(crate) fn image_calculated_cost(request_body: &[u8], usage: &Value) -> Optio
     let request = serde_json::from_slice::<ImageModel>(request_body).ok()?;
     if !matches!(
         request.model.as_str(),
-        "gpt-image-2" | "gpt-image-2-2026-04-21"
+        "gpt-image-2"
+            | "gpt-image-2-2026-04-21"
+            | "gpt-image-2.5-sunburst"
+            | "gpt-image-2.5-sunburst-2026-09-08"
+            | "gpt-image-2.5-flare"
+            | "gpt-image-2.5-flare-2026-09-08"
     ) {
         return None;
     }
@@ -542,8 +619,9 @@ pub(crate) fn image_calculated_cost(request_body: &[u8], usage: &Value) -> Optio
         (_, _, cached) if cached == input => (text_input, image_input),
         _ => return None,
     };
-    // https://developers.openai.com/api/docs/pricing (2026-09-08)
-    // USD / 1M tokens: text 5 / cached 1.25; image 8 / cached 2 / output 30.
+    // 价格来源：https://developers.openai.com/api/docs/pricing，核验日期 2026-09-09。
+    // GPT Image 2 和两个 2.5 型号按相同的 token 单价计费，不使用按张估价。
+    // 每百万 token 的美元单价：文本 5 / 缓存 1.25；图片 8 / 缓存 2 / 输出 30。
     let text = token_amounts(
         TokenRates::new(50_000, 0, 12_500),
         0,
@@ -565,16 +643,10 @@ pub(crate) fn image_calculated_cost(request_body: &[u8], usage: &Value) -> Optio
 
 fn model_pricing(model: &str) -> Option<ModelPricing> {
     let normalized = normalize_model_name(model);
-    if UNPRICED_MODELS
-        .iter()
-        .any(|rule| model_matches_rule(&normalized, rule))
-    {
-        return None;
-    }
+    let model = pricing_model_name(&normalized);
     PRICING_RULES
         .iter()
-        .filter(|rule| model_matches_rule(&normalized, rule.model))
-        .max_by_key(|rule| rule.model.len())
+        .find(|rule| model == rule.model)
         .map(|rule| rule.pricing)
 }
 
@@ -613,16 +685,17 @@ fn fixed_block_web_search_model(model: &str) -> bool {
 
 fn reasoning_model(model: &str) -> bool {
     let normalized = normalize_model_name(model);
-    normalized.starts_with("gpt-5")
-        || model_matches_rule(&normalized, "gpt-6-astra")
-        || normalized.starts_with("o1")
-        || normalized.starts_with("o3")
-        || normalized.starts_with("o4")
+    let model = pricing_model_name(&normalized);
+    model.starts_with("gpt-5")
+        || model == "gpt-6-astra"
+        || model.starts_with("o1")
+        || model.starts_with("o3")
+        || model.starts_with("o4")
 }
 
 fn pricing_tier(service_tier: Option<&str>) -> Option<PricingTier> {
     match service_tier {
-        None | Some("auto" | "default" | "standard") => Some(PricingTier::Standard),
+        None | Some("default" | "standard") => Some(PricingTier::Standard),
         Some("flex") => Some(PricingTier::Flex),
         Some("fast" | "priority") => Some(PricingTier::Fast),
         Some(_) => None,
@@ -644,20 +717,23 @@ fn token_amounts(
     cached_tokens: u64,
     cache_write_tokens: u64,
 ) -> Option<TokenAmounts> {
+    if cached_tokens.checked_add(cache_write_tokens)? > input_tokens {
+        return None;
+    }
     let billed_cache_read = if rates.cache_read_ticks > 0 {
-        cached_tokens.min(input_tokens)
+        cached_tokens
     } else {
         0
     };
     let cache_write_rate = cache_write_rate(rates, cache_write_percent)?;
     let billed_cache_write = if cache_write_rate > 0 {
-        cache_write_tokens.min(input_tokens.saturating_sub(billed_cache_read))
+        cache_write_tokens
     } else {
         0
     };
     let uncached_input = input_tokens
-        .saturating_sub(billed_cache_read)
-        .saturating_sub(billed_cache_write);
+        .checked_sub(billed_cache_read)?
+        .checked_sub(billed_cache_write)?;
     let input_ticks = u128::from(uncached_input).checked_mul(rates.input_ticks)?;
     let output_ticks = u128::from(output_tokens).checked_mul(rates.output_ticks)?;
     let cache_read_ticks = u128::from(billed_cache_read).checked_mul(rates.cache_read_ticks)?;
@@ -703,13 +779,47 @@ fn normalize_model_name(model: &str) -> String {
         .to_ascii_lowercase()
 }
 
-fn model_matches_rule(model: &str, rule: &str) -> bool {
-    if model == rule {
-        return true;
+// 仅已核验的别名和快照可以共用价格；未知后缀、未来日期和微调模型
+// 不得直接继承父型号的价格。
+fn pricing_model_name(model: &str) -> &str {
+    match model {
+        "gpt-3.5-turbo-0125" => "gpt-3.5-turbo",
+        "gpt-4-0314" => "gpt-4",
+        "gpt-4-0613" => "gpt-4",
+        "gpt-4-turbo-2024-04-09" => "gpt-4-turbo",
+        "gpt-4.1-2025-04-14" => "gpt-4.1",
+        "gpt-4.1-mini-2025-04-14" => "gpt-4.1-mini",
+        "gpt-4.1-nano-2025-04-14" => "gpt-4.1-nano",
+        "gpt-4o-2024-08-06" => "gpt-4o",
+        "gpt-4o-2024-11-20" => "gpt-4o",
+        "gpt-4o-mini-2024-07-18" => "gpt-4o-mini",
+        "gpt-5-2025-08-07" => "gpt-5",
+        "gpt-5-mini-2025-08-07" => "gpt-5-mini",
+        "gpt-5-nano-2025-08-07" => "gpt-5-nano",
+        "gpt-5-pro-2025-10-06" => "gpt-5-pro",
+        "gpt-5.1-2025-11-13" => "gpt-5.1",
+        "gpt-5.2-2025-12-11" => "gpt-5.2",
+        "gpt-5.2-pro-2025-12-11" => "gpt-5.2-pro",
+        "gpt-5.4-2026-03-05" => "gpt-5.4",
+        "gpt-5.4-mini-2026-03-17" => "gpt-5.4-mini",
+        "gpt-5.4-nano-2026-03-17" => "gpt-5.4-nano",
+        "gpt-5.4-pro-2026-03-05" => "gpt-5.4-pro",
+        "gpt-5.5-2026-04-23" => "gpt-5.5",
+        "gpt-5.5-pro-2026-04-23" => "gpt-5.5-pro",
+        "gpt-daybreak-blue-latest" => "gpt-5.6-sol",
+        "gpt-daybreak-red-latest" => "gpt-5.6-cyber",
+        "o1-2024-12-17" => "o1",
+        "o1-pro-2025-03-19" => "o1-pro",
+        "o3-2025-04-16" => "o3",
+        "o3-mini-2025-01-31" => "o3-mini",
+        "o3-pro-2025-06-10" => "o3-pro",
+        "o4-mini-2025-04-16" => "o4-mini",
+        _ => model,
     }
-    model
-        .strip_prefix(rule)
-        .is_some_and(|suffix| matches!(suffix.as_bytes().first(), Some(b'-' | b'.' | b':')))
+}
+
+fn model_matches_rule(model: &str, rule: &str) -> bool {
+    pricing_model_name(model) == rule
 }
 
 /// 规范化请求或响应携带的服务档位，供观测与计费共用。
diff --git a/backend/crates/providers/openai/tests/provider/contract.rs b/backend/crates/providers/openai/tests/provider/contract.rs
index 92b3a48f9..783c72d8c 100644
--- a/backend/crates/providers/openai/tests/provider/contract.rs
+++ b/backend/crates/providers/openai/tests/provider/contract.rs
@@ -1001,6 +1001,18 @@ async fn image_endpoints_bypass_only_the_text_catalog_and_preserve_the_current_c
             br#"{"model":"gpt-image-2","images":[{"image_url":"data:image/png;base64,AAEC"}],"prompt":"add fog","prompt":"duplicate remains opaque"}"#.as_slice(),
             br#"{"created":1787212801,"data":[{"b64_json":"AwQF"}],"quality":"high"}"#.as_slice(),
         ),
+        (
+            ImageRequestKind::Generation,
+            "/codex/images/generations",
+            br#"{"model":"gpt-image-2.5-flare","prompt":"a lighthouse","quality":"xhigh"}"#.as_slice(),
+            br#"{"created":1788900000,"data":[{"b64_json":"AAEC"}],"quality":"xhigh"}"#.as_slice(),
+        ),
+        (
+            ImageRequestKind::Edit,
+            "/codex/images/edits",
+            br#"{"model":"gpt-image-2.5-sunburst","images":[{"image_url":"data:image/png;base64,AAEC"}],"prompt":"add fog","quality":"max"}"#.as_slice(),
+            br#"{"created":1788900001,"data":[{"b64_json":"AwQF"}],"quality":"max"}"#.as_slice(),
+        ),
     ];
     for (_, endpoint, body, response_body) in &cases {
         Mock::given(method("POST"))
@@ -1225,6 +1237,10 @@ async fn image_prices_should_use_modality_rates_and_precede_delivery() {
         for (kind, model) in [
             (ImageRequestKind::Generation, "gpt-image-2"),
             (ImageRequestKind::Edit, "gpt-image-2-2026-04-21"),
+            (ImageRequestKind::Generation, "gpt-image-2.5-sunburst"),
+            (ImageRequestKind::Edit, "gpt-image-2.5-sunburst-2026-09-08"),
+            (ImageRequestKind::Generation, "gpt-image-2.5-flare"),
+            (ImageRequestKind::Edit, "gpt-image-2.5-flare-2026-09-08"),
         ] {
             let events = image_metering_events(kind, &response, model).await;
             let costs = events
@@ -1240,9 +1256,13 @@ async fn image_prices_should_use_modality_rates_and_precede_delivery() {
                         })
                 })
                 .collect::<Vec<_>>();
-            assert_eq!(costs.len(), 1);
+            assert_eq!(costs.len(), 1, "model={model}");
             let (cost_index, cost) = costs[0];
-            assert_eq!(cost.total().amount().scaled(), expected_ticks);
+            assert_eq!(
+                cost.total().amount().scaled(),
+                expected_ticks,
+                "model={model}"
+            );
             assert_eq!(cost.into_estimate().source().as_str(), "calculated");
             let wire_index = events
                 .iter()
@@ -1272,6 +1292,9 @@ async fn image_prices_should_remain_unknown_when_modality_or_model_is_uncertain(
     let mut cases = vec![
         ("gpt-image-future", valid.clone()),
         ("gpt-image-2-future", valid.clone()),
+        ("gpt-image-2.5", valid.clone()),
+        ("gpt-image-2.5-sunburst-future", valid.clone()),
+        ("gpt-image-2.5-flare-future", valid.clone()),
         ("gpt-image-2", json!(null)),
         (
             "gpt-image-2",
@@ -4820,7 +4843,7 @@ async fn completed_response_persists_session_affinity_before_stream_consumer_sto
     drop(stream);
 
     assert_eq!(affinity.binding_count(), 1);
-    assert_eq!(observed_service_tier.as_deref(), Some("priority"));
+    assert_eq!(observed_service_tier.as_deref(), Some("default"));
     assert_eq!(upstream_service_tier.as_deref(), Some("default"));
 }
 
diff --git a/backend/crates/providers/openai/tests/transport/canonical.rs b/backend/crates/providers/openai/tests/transport/canonical.rs
index 6dc5dd82d..5e6775293 100644
--- a/backend/crates/providers/openai/tests/transport/canonical.rs
+++ b/backend/crates/providers/openai/tests/transport/canonical.rs
@@ -219,7 +219,7 @@ fn decoder_should_emit_calculated_cost_for_complete_known_model_usage() {
 }
 
 #[test]
-fn decoder_should_prefer_requested_service_tier_over_response_tier_for_billing() {
+fn decoder_should_prefer_response_service_tier_over_requested_tier_for_billing() {
     let body = concat!(
         "event: response.created\n",
         "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_fast_cost\",\"model\":\"gpt-5.4\",\"service_tier\":\"default\"}}\n\n",
@@ -236,7 +236,7 @@ fn decoder_should_prefer_requested_service_tier_over_response_tier_for_billing()
     assert!(canonical_facts(&events).into_iter().any(|event| matches!(
         event,
         GatewayEvent::CalculatedCost(cost)
-            if cost.total().amount().scaled() == 6_875_000
+            if cost.total().amount().scaled() == 3_437_500
     )));
 }
 
@@ -346,7 +346,7 @@ fn decoder_should_fail_closed_for_fixed_block_web_search_content() {
 }
 
 #[test]
-fn websocket_decoder_should_prefer_requested_service_tier_over_response_tier_for_billing() {
+fn websocket_decoder_should_prefer_response_service_tier_over_requested_tier_for_billing() {
     let created = websocket_event_to_sse_frame(
         r#"{"type":"response.created","response":{"id":"resp_ws_fast_cost","model":"gpt-5.4","service_tier":"default"}}"#,
     )
@@ -366,7 +366,7 @@ fn websocket_decoder_should_prefer_requested_service_tier_over_response_tier_for
     assert!(canonical_facts(&events).into_iter().any(|event| matches!(
         event,
         GatewayEvent::CalculatedCost(cost)
-            if cost.total().amount().scaled() == 6_875_000
+            if cost.total().amount().scaled() == 3_437_500
     )));
 }
 
@@ -936,3 +936,96 @@ fn canonical_facts(events: &[ProviderEvent]) -> Vec<&GatewayEvent> {
         .flat_map(ProviderEvent::canonical_facts)
         .collect()
 }
+
+fn pricing_response_cost(
+    requested: Option<&str>,
+    actual: Option<&str>,
+    output: serde_json::Value,
+    tools: &[serde_json::Value],
+) -> Option<u128> {
+    let mut response = json!({
+        "id": "resp_pricing", "model": "gpt-6-astra", "status": "completed",
+        "output": output, "usage": {"input_tokens": 10_000, "output_tokens": 0, "total_tokens": 10_000}
+    });
+    if let Some(tier) = actual {
+        response["service_tier"] = json!(tier);
+    }
+    let created = json!({"type": "response.created", "response": {"id": "resp_pricing", "model": "gpt-6-astra"}});
+    let completed = json!({"type": "response.completed", "response": response});
+    let body = format!(
+        "event: response.created\ndata: {created}\n\nevent: response.completed\ndata: {completed}\n\n"
+    );
+    let events = CodexCanonicalDecoder::new("gpt-6-astra")
+        .with_requested_service_tier(requested)
+        .with_request_tool_pricing("gpt-6-astra", Some(tools))
+        .with_raw_sse_passthrough()
+        .push(body.as_bytes())
+        .expect("计费不确定时仍须正常透传响应");
+    assert!(
+        canonical_facts(&events)
+            .iter()
+            .any(|event| matches!(event, GatewayEvent::Completed(_)))
+    );
+    canonical_facts(&events)
+        .into_iter()
+        .find_map(|event| match event {
+            GatewayEvent::CalculatedCost(cost) => Some(cost.total().amount().scaled()),
+            _ => None,
+        })
+}
+
+#[test]
+fn billing_should_follow_actual_tier_and_only_fall_back_when_it_is_absent() {
+    for (requested, actual, expected) in [
+        (Some("fast"), Some("default"), Some(1_000_000_000)),
+        (Some("auto"), Some("fast"), Some(2_000_000_000)),
+        (Some("default"), Some("flex"), Some(500_000_000)),
+        (Some("fast"), None, Some(2_000_000_000)),
+        (Some("auto"), None, None),
+        (Some("default"), Some("future"), None),
+    ] {
+        assert_eq!(
+            pricing_response_cost(requested, actual, json!([]), &[]),
+            expected,
+            "{requested:?} -> {actual:?}"
+        );
+    }
+}
+
+#[test]
+fn billing_should_add_file_and_web_search_fees_without_tier_markup() {
+    let output = json!([
+        {"type": "file_search_call", "id": "fs_1", "status": "completed"},
+        {"type": "file_search_call", "id": "fs_2", "status": "completed"},
+        {"type": "web_search_call", "id": "ws_1", "status": "completed", "action": {"type": "search"}}
+    ]);
+    for (tier, expected) in [("default", 1_150_000_000), ("fast", 2_150_000_000)] {
+        assert_eq!(
+            pricing_response_cost(
+                None,
+                Some(tier),
+                output.clone(),
+                &[json!({"type": "web_search"})]
+            ),
+            Some(expected)
+        );
+    }
+}
+
+#[test]
+fn billing_should_not_emit_partial_totals_for_unpriced_tool_outputs() {
+    for output in [
+        json!([{"type": "code_interpreter_call", "id": "ci_1", "status": "completed"}]),
+        json!([{"type": "shell_call", "id": "sh_1", "status": "completed"}]),
+        json!([{"type": "image_generation_call", "id": "ig_1", "status": "completed"}]),
+        json!([{"type": "future_paid_tool_call", "id": "tool_1"}]),
+        json!([{"type": "file_search_call", "id": "fs_1", "status": "failed"}]),
+        json!({"malformed": true}),
+    ] {
+        assert_eq!(
+            pricing_response_cost(None, None, output.clone(), &[]),
+            None,
+            "{output}"
+        );
+    }
+}
diff --git a/backend/crates/providers/openai/tests/transport/usage.rs b/backend/crates/providers/openai/tests/transport/usage.rs
index f304bf782..e5b6556fe 100644
--- a/backend/crates/providers/openai/tests/transport/usage.rs
+++ b/backend/crates/providers/openai/tests/transport/usage.rs
@@ -239,6 +239,90 @@ async fn exact_limit_success_body_should_be_accepted() {
     assert!(usage["rate_limit"].is_object());
 }
 
+#[test]
+fn billing_should_use_explicit_variant_prices_and_verified_snapshots() {
+    // 每个用例包含一个普通输入 token 和一个输出 token，金额单位为 USD ticks。
+    for (model, expected) in [
+        ("gpt-5.1-codex-mini", 22_500),
+        ("gpt-5.3-chat-latest", 157_500),
+        ("gpt-5.6-cyber", 875_000),
+        ("gpt-5.5-cyber", 875_000),
+        ("gpt-daybreak-red-latest", 875_000),
+        ("gpt-daybreak-blue-latest", 350_000),
+        ("chat-latest", 350_000),
+        ("gpt-5.1-codex-max", 112_500),
+        ("gpt-5.2-codex", 157_500),
+        ("gpt-5.4-2026-03-05", 175_000),
+        (" OpenAI/GPT-4o-2024-08-06 ", 125_000),
+        ("gpt-4o-2024-05-13", 200_000),
+        ("gpt-3.5-turbo-0125", 20_000),
+        ("gpt-3.5-turbo-1106", 30_000),
+    ] {
+        let billing =
+            openai_billing_breakdown(model, billing_usage(1, 1, 0, 0), None).expect(model);
+        assert_eq!(
+            billing.total_amount().amount().scaled(),
+            expected,
+            "{model}"
+        );
+    }
+}
+
+#[test]
+fn billing_should_not_inherit_prices_for_unknown_models_or_tiers() {
+    for model in [
+        "gpt-6-astra-future",
+        "gpt-6-astra-2099-01-01",
+        "gpt-5.6-sol-wm",
+        "gpt-5.6-cyber-future",
+        "gpt-5.4-cyber",
+        "gpt-5.3-codex-spark",
+        "gpt-4.5-preview",
+        "gpt-4-32k",
+        "gpt-5.4:custom",
+        "gpt-5.4.1",
+    ] {
+        assert!(
+            openai_billing_breakdown(model, billing_usage(1, 1, 0, 0), None).is_none(),
+            "{model}"
+        );
+    }
+    for (model, input, tier) in [
+        ("gpt-5.6-cyber", 272_001, None),
+        ("gpt-5.5-cyber", 272_001, None),
+        ("gpt-5.6-cyber", 1, Some("fast")),
+        ("gpt-5.1-codex-mini", 1, Some("flex")),
+        ("gpt-6-astra", 1, Some("auto")),
+        ("gpt-6-astra", 1, Some("ultrafast")),
+    ] {
+        assert!(
+            openai_billing_breakdown(model, billing_usage(input, 1, 0, 0), tier).is_none(),
+            "{model} {tier:?}"
+        );
+    }
+}
+
+#[test]
+fn billing_should_reject_overlapping_or_overflowing_cache_counts() {
+    for (input, cached, written) in [(100, 20, 90), (100, 0, 101), (u64::MAX, u64::MAX, 1)] {
+        assert!(
+            openai_billing_breakdown(
+                "gpt-6-astra",
+                billing_usage(input, 0, cached, written),
+                None
+            )
+            .is_none()
+        );
+    }
+    let cyber = openai_billing_breakdown("gpt-5.6-cyber", billing_usage(100, 10, 20, 10), None)
+        .expect("已公开的 Cyber 缓存写价格");
+    assert_eq!(
+        cyber.cache_write_price_per_million().amount().to_string(),
+        "15.6250000000"
+    );
+    assert_eq!(cyber.total_amount().amount().scaled(), 18_062_500);
+}
+
 #[tokio::test]
 async fn fetch_should_use_wham_usage_headers_only() {
     let server = MockServer::start().await;
diff --git a/backend/crates/providers/xai/src/admin.rs b/backend/crates/providers/xai/src/admin.rs
index f1f9f9e2d..b92426e61 100644
--- a/backend/crates/providers/xai/src/admin.rs
+++ b/backend/crates/providers/xai/src/admin.rs
@@ -56,7 +56,7 @@ use crate::credential::{
     RotateManagedGrokCredential, SecretValue, VerifiedGrokAccount, VerifiedTokenSet,
 };
 use crate::transport::profile::{GrokCliReleaseSnapshot, GrokCliReleaseStatus};
-use crate::transport::{GROK_CLI_BASE_URL, XAI_PROVIDER_NAME, grok_billing_breakdown};
+use crate::transport::{GROK_CLI_BASE_URL, XAI_PROVIDER_NAME, grok_billing_breakdown_with_tier};
 
 const PENDING_SCHEMA_VERSION: u64 = 3;
 const PENDING_TTL: TimeDelta = TimeDelta::minutes(30);
@@ -390,11 +390,12 @@ impl ProviderAdmin for XaiAdminProvider {
         else {
             return Ok(None);
         };
-        let Some(breakdown) = grok_billing_breakdown(
+        let Some(breakdown) = grok_billing_breakdown_with_tier(
             &input.upstream_model_id,
             input_tokens,
             output_tokens,
             input.cached_tokens.unwrap_or_default(),
+            input.service_tier.as_deref(),
         ) else {
             return Ok(None);
         };
diff --git a/backend/crates/providers/xai/src/lib.rs b/backend/crates/providers/xai/src/lib.rs
index c48114e8c..539ddc114 100644
--- a/backend/crates/providers/xai/src/lib.rs
+++ b/backend/crates/providers/xai/src/lib.rs
@@ -79,7 +79,7 @@ pub use transport::{
     MAX_GROK_BILLING_BYTES, MAX_GROK_MODEL_CATALOG_BYTES, OfficialGrokEndpointPolicy,
     ReqwestGrokInferenceTransport, ReqwestGrokModelCatalogTransport, ReqwestOAuthTransport,
     SelectedGrokSession, XAI_PROVIDER_NAME, build_grok_headers, grok_billing_breakdown,
-    parse_grok_billing, parse_grok_model_catalog,
+    grok_billing_breakdown_with_tier, parse_grok_billing, parse_grok_model_catalog,
 };
 
 /// xAI 初始化后交给组装根的最小能力集。
diff --git a/backend/crates/providers/xai/src/provider/stream.rs b/backend/crates/providers/xai/src/provider/stream.rs
index f359ddd62..ba8d81b5a 100644
--- a/backend/crates/providers/xai/src/provider/stream.rs
+++ b/backend/crates/providers/xai/src/provider/stream.rs
@@ -185,7 +185,8 @@ pub(super) fn cold_compaction_http_sse_stream(
                 }
             }
         };
-        yield ProviderEvent::observation(accepted.observation);
+        let mut observation = accepted.observation;
+        yield ProviderEvent::observation(observation.clone());
 
         let mut body = accepted.response.into_body();
         let mut canonical = GrokCanonicalDecoder::new(upstream_model.as_str());
@@ -251,6 +252,10 @@ pub(super) fn cold_compaction_http_sse_stream(
             }
         }
 
+        if let Some(tier) = canonical.response_service_tier() {
+            observation = observation.with_service_tier_if_valid(tier.to_owned());
+            yield ProviderEvent::observation(observation);
+        }
         let started = facts
             .started
             .ok_or_else(|| mark_transient_compaction_failure(protocol_sent()))?;
@@ -460,7 +465,7 @@ pub(super) fn cold_http_sse_stream(
             }
         };
 
-        let observation = xai_response_observation(&response)?;
+        let mut observation = xai_response_observation(&response)?;
         let base_timings = observation.timings();
         let mut first_token_ms: Option<u64> = None;
         yield ProviderEvent::observation(observation.clone());
@@ -529,6 +534,14 @@ pub(super) fn cold_http_sse_stream(
                         }),
                 );
             }
+            if let Some(tier) = decoder.response_service_tier()
+                && observation.service_tier() != Some(tier)
+            {
+                observation = observation.with_service_tier_if_valid(tier.to_owned());
+                yield ProviderEvent::observation(observation.clone().with_timings(
+                    ProviderResponseTimings { first_token_ms, ..base_timings },
+                ));
+            }
             let completed = events
                 .iter()
                 .flat_map(ProviderEvent::canonical_facts)
@@ -581,6 +594,14 @@ pub(super) fn cold_http_sse_stream(
                     }),
             );
         }
+        if let Some(tier) = decoder.response_service_tier()
+            && observation.service_tier() != Some(tier)
+        {
+            observation = observation.with_service_tier_if_valid(tier.to_owned());
+            yield ProviderEvent::observation(observation.clone().with_timings(
+                ProviderResponseTimings { first_token_ms, ..base_timings },
+            ));
+        }
         attach_xai_session_update(&mut final_events, &mut session_capture)?;
         if let Some(capture) = reasoning_replay_capture.as_mut() {
             capture.observe(&final_events);
diff --git a/backend/crates/providers/xai/src/transport/canonical.rs b/backend/crates/providers/xai/src/transport/canonical.rs
index 18c3149ff..a0a77c868 100644
--- a/backend/crates/providers/xai/src/transport/canonical.rs
+++ b/backend/crates/providers/xai/src/transport/canonical.rs
@@ -38,6 +38,20 @@ struct ModelPricing {
     long: TokenRates,
 }
 
+// 价格来源：https://docs.x.ai/developers/pricing，核验日期 2026-09-09。
+const GROK_46_PRICING: ModelPricing = ModelPricing {
+    short: TokenRates {
+        input_ticks: 20_000,
+        cached_input_ticks: 5_000,
+        output_ticks: 60_000,
+    },
+    long: TokenRates {
+        input_ticks: 40_000,
+        cached_input_ticks: 10_000,
+        output_ticks: 120_000,
+    },
+};
+
 const GROK_45_PRICING: ModelPricing = ModelPricing {
     short: TokenRates {
         input_ticks: 20_000,
@@ -85,6 +99,23 @@ pub fn grok_billing_breakdown(
     output_tokens: u64,
     cached_tokens: u64,
 ) -> Option<CalculatedCostBreakdown> {
+    grok_billing_breakdown_with_tier(model, input_tokens, output_tokens, cached_tokens, None)
+}
+
+/// 按响应确认的实际档位计算 xAI Token 费用；未知档位不估算。
+#[must_use]
+pub fn grok_billing_breakdown_with_tier(
+    model: &str,
+    input_tokens: u64,
+    output_tokens: u64,
+    cached_tokens: u64,
+    service_tier: Option<&str>,
+) -> Option<CalculatedCostBreakdown> {
+    let (tier, multiplier) = match service_tier.map(str::trim) {
+        None | Some("default" | "standard") => ("default", 1_u128),
+        Some("priority") => ("priority", 2),
+        Some(_) => return None,
+    };
     let pricing = model_pricing(model)?;
     let rates = if input_tokens >= LONG_CONTEXT_THRESHOLD {
         pricing.long
@@ -101,21 +132,21 @@ pub fn grok_billing_breakdown(
         .checked_add(output_amount_ticks)?;
     Some(CalculatedCostBreakdown::new(
         CalculatedCostAmounts::new(
-            usd_money(input_amount_ticks)?,
-            usd_money(output_amount_ticks)?,
-            usd_money(cache_read_amount_ticks)?,
+            usd_money(input_amount_ticks.checked_mul(multiplier)?)?,
+            usd_money(output_amount_ticks.checked_mul(multiplier)?)?,
+            usd_money(cache_read_amount_ticks.checked_mul(multiplier)?)?,
             usd_money(0)?,
             usd_money(standard_amount_ticks)?,
-            usd_money(standard_amount_ticks)?,
+            usd_money(standard_amount_ticks.checked_mul(multiplier)?)?,
         ),
         CalculatedCostRates::new(
-            usd_price_per_million(rates.input_ticks)?,
-            usd_price_per_million(rates.output_ticks)?,
-            usd_price_per_million(rates.cached_input_ticks)?,
+            usd_price_per_million(rates.input_ticks.checked_mul(multiplier)?)?,
+            usd_price_per_million(rates.output_ticks.checked_mul(multiplier)?)?,
+            usd_price_per_million(rates.cached_input_ticks.checked_mul(multiplier)?)?,
             usd_money(0)?,
         ),
-        Some("default".to_owned()),
-        100,
+        Some(tier.to_owned()),
+        if multiplier == 2 { 200 } else { 100 },
     ))
 }
 
@@ -144,6 +175,8 @@ pub struct GrokCanonicalDecoder {
     tool_arguments_seen: BTreeSet<u32>,
     usage_emitted: bool,
     output_start_seen: bool,
+    response_service_tier: Option<String>,
+    requires_provider_cost: bool,
 }
 
 impl GrokCanonicalDecoder {
@@ -159,6 +192,8 @@ impl GrokCanonicalDecoder {
             tool_arguments_seen: BTreeSet::new(),
             usage_emitted: false,
             output_start_seen: false,
+            response_service_tier: None,
+            requires_provider_cost: false,
         }
     }
 
@@ -168,10 +203,30 @@ impl GrokCanonicalDecoder {
     pub fn for_request(fallback_model: impl Into<String>, request: &GrokResponsesRequest) -> Self {
         Self {
             response_transform: request.response_transform(),
+            requires_provider_cost: request.body().get("tool_choice").and_then(Value::as_str)
+                != Some("none")
+                && request
+                    .body()
+                    .get("tools")
+                    .and_then(Value::as_array)
+                    .is_some_and(|tools| {
+                        tools.iter().any(|tool| {
+                            !matches!(
+                                tool.get("type").and_then(Value::as_str),
+                                Some("function" | "custom" | "mcp")
+                            )
+                        })
+                    }),
             ..Self::new(fallback_model)
         }
     }
 
+    /// 上游报告的实际服务档位，供费用与持久观测共用。
+    #[must_use]
+    pub fn response_service_tier(&self) -> Option<&str> {
+        self.response_service_tier.as_deref()
+    }
+
     pub fn push(&mut self, chunk: &[u8]) -> Result<Vec<ProviderEvent>, ProviderError> {
         let events = self.decoder.push(chunk).map_err(protocol_error)?;
         self.decode(events)
@@ -259,6 +314,16 @@ impl GrokCanonicalDecoder {
                 continue;
             };
             let event_type = event_type.to_owned();
+            // 工具转换可能隐藏注入的调用，计费事实必须从转换前的上游事件读取。
+            if let Some(response) = value.get("response") {
+                if let Some(tier) = response.get("service_tier").and_then(Value::as_str) {
+                    self.response_service_tier = Some(tier.trim().to_owned());
+                }
+                self.requires_provider_cost |= !token_only_response(response);
+            }
+            if let Some(item) = value.get("item") {
+                self.requires_provider_cost |= !token_only_output(item);
+            }
             // 无法转换的上游帧（多为畸形帧）直接丢弃，保持已开始的客户端流不中断。
             let Ok(transformed) = self
                 .response_transform
@@ -590,7 +655,10 @@ impl GrokCanonicalDecoder {
             .to_owned();
         if let Some(cost) = provider_reported_cost(response)? {
             output.push(GatewayEvent::ProviderCost(cost));
-        } else if let Some(cost) = usage.and_then(|usage| calculated_cost(response, &model, usage))
+        } else if !self.requires_provider_cost
+            && let Some(cost) = usage.and_then(|usage| {
+                calculated_cost(response, &model, usage, self.response_service_tier())
+            })
         {
             output.push(GatewayEvent::CalculatedCost(cost));
         }
@@ -729,29 +797,71 @@ fn provider_reported_cost(response: &Value) -> Result<Option<ProviderReportedCos
         return Ok(None);
     };
     let ticks = value.as_u64().ok_or_else(protocol_error_marker)?;
-    if ticks == 0 {
-        return Ok(None);
-    }
+    // 明确返回的零费用也是上游账单事实，不能当作缺失而改用本地估算。
     ProviderReportedCost::from_usd_ticks(u128::from(ticks))
         .map(Some)
         .map_err(protocol_error)
 }
 
-fn calculated_cost(response: &Value, model: &str, usage: TokenUsage) -> Option<CalculatedCost> {
-    if !billable_usage_is_complete(response, usage) {
+// cost_in_usd_ticks 已包含服务端工具费；缺少该字段时，只有全部按 token
+// 计费的响应才能确定总价。
+fn token_only_output(item: &Value) -> bool {
+    matches!(
+        item.get("type").and_then(Value::as_str),
+        Some(
+            "message"
+                | "reasoning"
+                | "function_call"
+                | "custom_tool_call"
+                | "mcp_call"
+                | "mcp_list_tools"
+                | "mcp_approval_request"
+                | "compaction"
+        )
+    )
+}
+
+fn token_only_response(response: &Value) -> bool {
+    let empty = |value: &Value| value.as_object().is_some_and(serde_json::Map::is_empty);
+    response.get("tool_usage").is_none_or(empty)
+        && response
+            .pointer("/usage/server_side_tool_usage")
+            .is_none_or(empty)
+        && response
+            .pointer("/usage/num_server_side_tool_calls")
+            .is_none_or(|value| value.as_u64() == Some(0))
+        && response.get("output").is_none_or(|value| {
+            value
+                .as_array()
+                .is_some_and(|items| items.iter().all(token_only_output))
+        })
+}
+
+fn calculated_cost(
+    response: &Value,
+    model: &str,
+    usage: TokenUsage,
+    service_tier: Option<&str>,
+) -> Option<CalculatedCost> {
+    if !billable_usage_is_complete(response, usage)
+        || usage.image_input_tokens > 0
+        || usage.image_output_tokens > 0
+    {
         return None;
     }
-    let breakdown = grok_billing_breakdown(
+    let breakdown = grok_billing_breakdown_with_tier(
         model,
         usage.input_tokens,
         usage.output_tokens,
         usage.cached_tokens,
+        service_tier,
     )?;
     Some(breakdown.calculated_cost())
 }
 
 fn model_pricing(model: &str) -> Option<ModelPricing> {
     match model {
+        "grok-4.6" | "grok-4.6-latest" => Some(GROK_46_PRICING),
         "grok-4.5" | "grok-4.5-latest" | "grok-4.5-build-free" | "grok-build-latest" => {
             Some(GROK_45_PRICING)
         }
diff --git a/backend/crates/providers/xai/src/transport/mod.rs b/backend/crates/providers/xai/src/transport/mod.rs
index fcdc79307..abcd9365b 100644
--- a/backend/crates/providers/xai/src/transport/mod.rs
+++ b/backend/crates/providers/xai/src/transport/mod.rs
@@ -11,7 +11,9 @@ pub(crate) mod profile;
 mod request;
 mod session;
 
-pub use canonical::{GrokCanonicalDecoder, grok_billing_breakdown};
+pub use canonical::{
+    GrokCanonicalDecoder, grok_billing_breakdown, grok_billing_breakdown_with_tier,
+};
 pub use catalog::{
     GROK_BILLING_URL, GROK_MODEL_CATALOG_URL, GrokBillingClient, GrokBillingError,
     GrokBillingRequest, GrokBillingSnapshot, GrokBillingTransport, GrokBillingTransportError,
diff --git a/backend/crates/providers/xai/tests/admin.rs b/backend/crates/providers/xai/tests/admin.rs
index 66b176d8c..d8a743683 100644
--- a/backend/crates/providers/xai/tests/admin.rs
+++ b/backend/crates/providers/xai/tests/admin.rs
@@ -158,6 +158,26 @@ async fn xai_admin_provider_validates_known_billing_breakdown() {
     assert_eq!(billing.total_amount.amount.as_str(), "0.0002175");
     assert_eq!(billing.input_price_per_million.amount.as_str(), "2");
     assert_eq!(billing.output_price_per_million.amount.as_str(), "6");
+
+    let priority = admin
+        .calculated_billing(&ProviderBillingInput {
+            upstream_model_id: "grok-4.6".to_owned(),
+            service_tier: Some("priority".to_owned()),
+            input_tokens: Some(100),
+            output_tokens: Some(10),
+            cached_tokens: Some(20),
+            cache_write_tokens: Some(0),
+            total: CurrencyCost {
+                currency: "USD".to_owned(),
+                amount: "0.00046".parse().expect("amount"),
+            },
+        })
+        .expect("计费明细")
+        .expect("已知的 Priority 价格");
+    assert_eq!(priority.service_tier.as_deref(), Some("priority"));
+    assert_eq!(priority.standard_amount.amount.as_str(), "0.00023");
+    assert_eq!(priority.input_price_per_million.amount.as_str(), "4");
+    assert_eq!(priority.multiplier_percent, 200);
 }
 
 #[tokio::test]
diff --git a/backend/crates/providers/xai/tests/provider/contract.rs b/backend/crates/providers/xai/tests/provider/contract.rs
index 236a28ab7..7d72677b7 100644
--- a/backend/crates/providers/xai/tests/provider/contract.rs
+++ b/backend/crates/providers/xai/tests/provider/contract.rs
@@ -1239,6 +1239,56 @@ async fn transport_error_observation_should_retain_available_metrics() {
     assert_eq!(error.kind(), ProviderErrorKind::Unavailable);
 }
 
+#[tokio::test]
+async fn billing_observation_should_persist_the_final_response_tier_before_completion() {
+    for (actual, expected_cost) in [("priority", 400_000_000), ("default", 200_000_000)] {
+        let body = format!(
+            concat!(
+                "event: response.created\ndata: {{\"type\":\"response.created\",\"response\":{{\"id\":\"resp_tier\",\"model\":\"grok-4.6\",\"service_tier\":\"priority\"}}}}\n\n",
+                "event: response.completed\ndata: {{\"type\":\"response.completed\",\"response\":{{\"id\":\"resp_tier\",\"model\":\"grok-4.6\",\"service_tier\":\"{}\",\"status\":\"completed\",\"output\":[],\"usage\":{{\"input_tokens\":10000,\"output_tokens\":0}}}}}}\n\n"
+            ),
+            actual
+        );
+        let transport =
+            StubInferenceTransport::sequence([InferenceMode::SuccessBody(body.into_bytes())]);
+        let provider = provider(StubSelector::success(), transport).await;
+        let mut stream = provider
+            .execute(
+                provider_request("xai"),
+                context(CancellationToken::new(), None),
+            )
+            .await
+            .expect("provider stream");
+        let mut tier = None;
+        let mut cost = None;
+        let mut completed = false;
+        while let Some(event) = stream.next().await {
+            let event = event.expect("provider event");
+            if let Some(value) = event
+                .response_observation()
+                .and_then(|value| value.service_tier())
+            {
+                tier = Some(value.to_owned());
+            }
+            for fact in event.canonical_facts() {
+                match fact {
+                    GatewayEvent::CalculatedCost(value) => {
+                        cost = Some(value.total().amount().scaled())
+                    }
+                    GatewayEvent::Completed(_) => completed = true,
+                    _ => {}
+                }
+            }
+            if completed {
+                break;
+            }
+        }
+        assert!(completed);
+        assert_eq!(tier.as_deref(), Some(actual));
+        assert_eq!(cost, Some(expected_cost));
+    }
+}
+
 #[tokio::test]
 async fn compaction_stream_should_emit_openai_wire_and_keep_only_metering_canonical() {
     let summary = valid_compaction_summary("validated summary");
diff --git a/backend/crates/providers/xai/tests/transport/canonical.rs b/backend/crates/providers/xai/tests/transport/canonical.rs
index a464fbeb5..22b837416 100644
--- a/backend/crates/providers/xai/tests/transport/canonical.rs
+++ b/backend/crates/providers/xai/tests/transport/canonical.rs
@@ -5,7 +5,10 @@ use gateway_core::policy::ClientApiKeyId;
 use gateway_protocol::openai::sse::{encode_sse_event, encode_sse_event_with_metadata};
 use serde_json::Value;
 
-use provider_xai::{GrokCanonicalDecoder, GrokResponsesRequest, grok_billing_breakdown};
+use provider_xai::{
+    GrokCanonicalDecoder, GrokResponsesRequest, grok_billing_breakdown,
+    grok_billing_breakdown_with_tier,
+};
 
 fn terminal_cost_events(
     model: &str,
@@ -231,17 +234,139 @@ fn billing_breakdown_should_match_the_calculated_terminal_cost() {
 
 #[test]
 fn decoder_should_price_official_grok_45_build_free_variant() {
-    let events = terminal_cost_events("grok-4.5-build-free", 201, 0, 76, Some(0));
+    let events = terminal_cost_events("grok-4.5-build-free", 201, 0, 76, None);
 
     assert_eq!(calculated_cost_ticks(&events), Some(8_580_000));
 }
 
 #[test]
-fn zero_provider_cost_should_fall_back_to_calculated_cost() {
+fn zero_provider_cost_should_remain_an_authoritative_zero() {
     let events = terminal_cost_events("grok-4.5", 1, 0, 1, Some(0));
 
-    assert_eq!(calculated_cost_ticks(&events), Some(80_000));
-    assert_eq!(provider_cost_ticks(&events), None);
+    assert_eq!(calculated_cost_ticks(&events), None);
+    assert_eq!(provider_cost_ticks(&events), Some(0));
+}
+
+#[test]
+fn billing_should_price_grok_46_cache_and_priority_at_the_context_boundary() {
+    // 短、长上下文分别检查普通输入、缓存、输出的单价，Priority 只采用已确认档位。
+    for (input, tier, expected) in [
+        (199_999, None, ["2", "0.5", "6"]),
+        (200_000, None, ["4", "1", "12"]),
+        (199_999, Some("priority"), ["4", "1", "12"]),
+        (200_000, Some("priority"), ["8", "2", "24"]),
+    ] {
+        let billing = grok_billing_breakdown_with_tier("grok-4.6", input, 10, 20, tier)
+            .expect("Grok 4.6 价格");
+        let rates = [
+            billing.input_price_per_million(),
+            billing.cache_read_price_per_million(),
+            billing.output_price_per_million(),
+        ]
+        .map(|money| {
+            money
+                .amount()
+                .to_string()
+                .trim_end_matches('0')
+                .trim_end_matches('.')
+                .to_owned()
+        });
+        assert_eq!(rates, expected);
+    }
+    assert!(grok_billing_breakdown_with_tier("grok-4.6", 10, 1, 0, Some("future")).is_none());
+    let priority = grok_billing_breakdown_with_tier("grok-4.6", 100, 10, 20, Some("priority"))
+        .expect("Priority 价格");
+    assert_eq!(priority.standard_amount().amount().scaled(), 2_300_000);
+    assert_eq!(priority.total_amount().amount().scaled(), 4_600_000);
+    assert_eq!(priority.multiplier_percent(), 200);
+}
+
+fn pricing_events(
+    mut response: Value,
+    request: Option<&GrokResponsesRequest>,
+) -> Vec<GatewayEvent> {
+    response["id"] = serde_json::json!("resp_pricing");
+    response["model"] = serde_json::json!("grok-4.6");
+    response["status"] = serde_json::json!("completed");
+    let created = serde_json::json!({"type": "response.created", "response": {"id": "resp_pricing", "model": "grok-4.6", "service_tier": "priority"}});
+    let completed = serde_json::json!({"type": "response.completed", "response": response});
+    let body = format!(
+        "event: response.created\ndata: {created}\n\nevent: response.completed\ndata: {completed}\n\n"
+    );
+    let mut decoder = request.map_or_else(
+        || GrokCanonicalDecoder::new("grok-4.6"),
+        |request| GrokCanonicalDecoder::for_request("grok-4.6", request),
+    );
+    let events = decoder
+        .push(body.as_bytes())
+        .expect("计费未知时仍应交付响应");
+    let facts: Vec<_> = events
+        .into_iter()
+        .flat_map(|event| event.into_parts().0)
+        .collect();
+    assert!(
+        facts
+            .iter()
+            .any(|event| matches!(event, GatewayEvent::Completed(_)))
+    );
+    facts
+}
+
+#[test]
+fn billing_should_use_the_final_response_tier_and_preserve_provider_cost() {
+    for (tier, ticks, calculated, provider) in [
+        ("priority", None, Some(400_000_000), None),
+        ("default", None, Some(200_000_000), None),
+        ("future", None, None, None),
+        ("priority", Some(0), None, Some(0)),
+        ("priority", Some(123), None, Some(123)),
+    ] {
+        let mut response = serde_json::json!({"service_tier": tier, "output": [], "usage": {"input_tokens": 10_000, "output_tokens": 0}});
+        if let Some(ticks) = ticks {
+            response["usage"]["cost_in_usd_ticks"] = serde_json::json!(ticks);
+        }
+        let events = pricing_events(response, None);
+        assert_eq!(
+            (calculated_cost_ticks(&events), provider_cost_ticks(&events)),
+            (calculated, provider),
+            "{tier}"
+        );
+    }
+}
+
+#[test]
+fn billing_should_require_provider_totals_for_server_tools() {
+    for kind in [
+        "web_search_call",
+        "x_search_call",
+        "code_interpreter_call",
+        "file_search_call",
+        "image_generation_call",
+        "future_tool_call",
+    ] {
+        let response = serde_json::json!({"output": [{"type": kind, "id": "tool_1", "status": "completed"}], "usage": {"input_tokens": 100, "output_tokens": 1}});
+        assert_eq!(
+            calculated_cost_ticks(&pricing_events(response, None)),
+            None,
+            "{kind}"
+        );
+    }
+    let request =
+        tool_request(serde_json::json!({"input": "查资料", "tools": [{"type": "web_search"}]}));
+    // 工具结果可能被上游省略，允许调用收费工具时不能仅靠空 output 认定没有附加费。
+    let mut response =
+        serde_json::json!({"output": [], "usage": {"input_tokens": 100, "output_tokens": 1}});
+    assert_eq!(
+        calculated_cost_ticks(&pricing_events(response.clone(), Some(&request))),
+        None
+    );
+    response["usage"]["cost_in_usd_ticks"] = serde_json::json!(0);
+    assert_eq!(
+        provider_cost_ticks(&pricing_events(response, Some(&request))),
+        Some(0)
+    );
+    let response = serde_json::json!({"output": [], "usage": {"input_tokens": 100, "output_tokens": 1, "num_server_side_tool_calls": 1}});
+    assert_eq!(calculated_cost_ticks(&pricing_events(response, None)), None);
 }
 
 #[test]