diff --git a/desktop/electron/src/main.ts b/desktop/electron/src/main.ts
index b675a0ba62..ac92fec17b 100644
--- a/desktop/electron/src/main.ts
+++ b/desktop/electron/src/main.ts
@@ -1726,28 +1726,6 @@ function canonicalTierKey(name: string): string {
 }
 const ROUTER_PROFILE_IDS = new Set(['tokenrhythm', 'openrouter', 'dashscope', 'deepseek', 'gemini', 'volcengine', 'openai', 'zhipu', 'moonshot'])
 const TOKENRHYTHM_REGISTER_URL = 'https://tokenrhythm.studio/register'
-const DESKTOP_ENSEMBLE_PROFILES: Record<StaticEnsembleSelectionMode, {
-  provider: string
-  proposers: string[]
-  aggregator: string
-}> = {
-  static_tokenrhythm_b5: {
-    provider: 'tokenrhythm',
-    proposers: ['deepseek-v4-pro', 'glm-5.2', 'kimi-k2.7-code', 'qwen3.7-max'],
-    aggregator: 'glm-5.2',
-  },
-  static_openrouter_b5: {
-    provider: 'openrouter',
-    proposers: [
-      'deepseek/deepseek-v4-pro',
-      'z-ai/glm-5.2',
-      'moonshotai/kimi-k2.7-code',
-      'qwen/qwen3.7-max',
-    ],
-    aggregator: 'z-ai/glm-5.2',
-  },
-}
-
 const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
   {
     id: 'tokenrhythm',
@@ -2214,31 +2192,11 @@ function ensembleConfigTomlLines(credential: DesktopConnection): string[] {
   if (!selectionMode) {
     throw new Error(`LLM Ensemble is not supported for provider ${credential.provider}.`)
   }
-  const profile = DESKTOP_ENSEMBLE_PROFILES[selectionMode]
-  const candidates = profile.proposers.flatMap(model => [
-    '',
-    '[[llm_ensemble.candidates]]',
-    `provider = ${tomlString(profile.provider)}`,
-    `model = ${tomlString(model)}`,
-    'source = "custom"',
-    'enabled = true',
-    'role = "proposer"',
-  ])
-  candidates.push(
-    '',
-    '[[llm_ensemble.candidates]]',
-    `provider = ${tomlString(profile.provider)}`,
-    `model = ${tomlString(profile.aggregator)}`,
-    'source = "custom"',
-    'enabled = true',
-    'role = "aggregator"',
-  )
   return [
     '',
     '[llm_ensemble]',
     'enabled = true',
-    'selection_mode = "custom_b5"',
-    ...candidates,
+    `selection_mode = ${tomlString(selectionMode)}`,
   ]
 }
 
diff --git a/docs/features/LLM-ensemble-design.md b/docs/features/LLM-ensemble-design.md
index c0d5c1a2df..3efa53e808 100644
--- a/docs/features/LLM-ensemble-design.md
+++ b/docs/features/LLM-ensemble-design.md
@@ -95,15 +95,16 @@ Source: `_build_static_b5_members`, `STATIC_B5_PROFILES`
 Each preset is a `StaticB5Profile` — four fixed proposers plus one aggregator,
 all bound to a single provider:
 
-| Profile | Provider | Proposers | Aggregator |
-|---------|----------|-----------|------------|
-| `static_openrouter_b5` | `openrouter` | `deepseek/deepseek-v4-pro`, `z-ai/glm-5.2`, `moonshotai/kimi-k2.7-code`, `qwen/qwen3.7-max` | `z-ai/glm-5.2` |
-| `static_tokenrhythm_b5` | `tokenrhythm` | `deepseek-v4-pro`, `glm-5.2`, `kimi-k2.7-code`, `qwen3.7-max` | `glm-5.2` |
-
-The TokenRhythm profile is a mirror of the OpenRouter one: same aggregation
-shape and defaults, the same four models, only the provider and the model-id
-naming differ (OpenRouter-style `vendor/model` slugs vs. TokenRhythm's bare
-names).
+| Profile | Provider | Proposers | Aggregator | Thinking |
+|---------|----------|-----------|------------|----------|
+| `static_openrouter_b5` | `openrouter` | `deepseek/deepseek-v4.1-flash`, `z-ai/glm-5.3-flash`, `qwen/qwen3.8-flash`, `qwen/qwen3.8-max-0902` | `deepseek/deepseek-v4.1-flash` | `high` for every member |
+| `static_tokenrhythm_b5` | `tokenrhythm` | `deepseek-flash`, `glm-5.3-flash`, `qwen3.8-flash`, `qwen3.8-max` | `deepseek-flash` | `high` for every member |
+
+The OpenRouter profile uses the C5 lineup selected by the full DRACO evaluation.
+The TokenRhythm profile maps the same C5 model families to model IDs published
+by TokenRhythm. Both profiles explicitly set `high` thinking on all four
+proposers and the aggregator and use the same aggregation runtime defaults,
+while their model-ID conventions remain provider-specific.
 
 `_build_static_b5_members` simply materializes the profile: each proposer model
 becomes an `EnsembleMemberConfig` labeled `proposer_1..N`, the aggregator model
diff --git a/docs/features/MULTI-MODEL-FUSION-C5-RELEASE-NOTE.md b/docs/features/MULTI-MODEL-FUSION-C5-RELEASE-NOTE.md
new file mode 100644
index 0000000000..23b6796f17
--- /dev/null
+++ b/docs/features/MULTI-MODEL-FUSION-C5-RELEASE-NOTE.md
@@ -0,0 +1,74 @@
+# 多模型融合 C5 配置更新
+
+本次版本将默认多模型融合配置升级为 C5。更新目标是：在提高复杂任务质量的同时，显著降低生成成本和响应时间。
+
+## 1. 本次更新内容
+
+旧配置使用 DeepSeek V4 Pro、GLM-5.2、Kimi K2.7 Code、Qwen3.7 Max 起草，由 GLM-5.2 汇总。本次升级后的配置为：
+
+| 供应商 | 4 个起草模型 | 汇总模型 |
+| --- | --- | --- |
+| OpenRouter | DeepSeek V4.1 Flash、GLM-5.3 Flash、Qwen3.8 Flash、Qwen3.8 Max 0902 | DeepSeek V4.1 Flash |
+| TokenRhythm | `deepseek-flash`、`glm-5.3-flash`、`qwen3.8-flash`、`qwen3.8-max` | `deepseek-flash` |
+
+同时完成以下调整：
+
+- 所有起草模型和汇总模型均显式使用 `high` 思考档位。
+- 起草模型关闭工具调用，只负责生成独立候选答案。
+- 汇总模型结合原始问题和候选答案生成最终结果，并保留工具调用能力。
+- 起草和汇总超时分别为 120 秒和 180 秒；起草失败不额外重试，至少一个起草成功即可进入汇总。
+- OpenRouter、TokenRhythm、WebUI 和桌面端统一使用同一份静态配置，减少配置漂移。
+
+## 2. DRACO 100 题实验结果
+
+所有主指标均按同一套 DRACO 100 道任务计算。下表以旧配置 `old` 为基线；费用为平均每题生成费用，耗时不含判分时间。
+
+| 组别 | AvgQ（较 old） | AvgPass | Avg Gen$/题（较 old） | p50（较 old） | p95（较 old） |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| old | 59.188（基线） | 63.503 | $0.423032（基线） | 423.2s（基线） | 3241.7s（基线） |
+| A | 60.787（+2.70%） | 64.312 | $0.328690（-22.30%） | 372.0s（-12.11%） | 3037.7s（-6.30%） |
+| B | 62.575（+5.72%） | 65.936 | $0.132132（-68.77%） | 266.0s（-37.14%） | 2720.8s（-16.07%） |
+| C1 | 63.523（+7.32%） | 67.201 | $0.119361（-71.78%） | 185.4s（-56.20%） | 1910.5s（-41.06%） |
+| C2 | 63.807（+7.80%） | 67.624 | $0.073857（-82.54%） | 154.0s（-63.61%） | 934.6s（-71.17%） |
+| C3 | 60.395（+2.04%） | 64.186 | $0.069649（-83.54%） | 259.2s（-38.76%） | 2251.6s（-30.54%） |
+| C4 | 62.650（+5.85%） | 66.494 | $0.162033（-61.70%） | 193.6s（-54.26%） | 2329.4s（-28.14%） |
+| **C5** | **64.190（+8.45%）** | **67.845** | **$0.137097（-67.59%）** | **195.7s（-53.75%）** | **1579.3s（-51.28%）** |
+| C6 | 62.692（+5.92%） | 66.038 | $0.144662（-65.80%） | 215.8s（-49.00%） | 2256.7s（-30.39%） |
+| C7 | 54.189（-8.45%） | 59.312 | $0.163359（-61.38%） | 157.7s（-62.75%） | 1847.5s（-43.01%） |
+
+选择 C5 的原因：
+
+- **质量最高：** C5 的 AvgQ 为 64.190，是全部实验组最高值；相对 old 提高 5.002 个百分点，增幅 8.45%。
+- **成本明显下降：** 平均每题生成费用下降 67.59%，100 题总生成费用从 $42.30 降至 $13.71。
+- **速度明显提升：** p50 缩短 53.75%，p95 缩短 51.28%。
+- **结果更完整：** C5 的 100 道任务全部完成原生评分；old 为 98/100，历史参照 B 为 97/100。
+- **符合质量优先的产品取向：** C2、C3 更便宜，但质量低于 C5。C5 在保留大幅降本的同时取得最高质量，因此作为本次默认配置。
+
+相对第二轮历史参照 B，C5 的 AvgQ 再提高 1.615 个百分点，生成费用增加 3.76%，但 p50 和 p95 分别缩短 26.43% 和 41.96%。
+
+> 上述结果是本轮 100 题的观测值。不同运行时段、搜索结果、服务负载和缓存状态可能影响结果；C5 相对 B 的质量区间仍跨过零点。
+
+## 3. 三个代表性案例
+
+### 案例一：印度 NCD 投资分析
+
+任务要求识别 2025 年 12 月 8 日仍在开放的 NCD，比较评级、税后收益和发行人杠杆。
+
+- **old：48.506 分**。漏掉 Muthoot Mercantile 和 KLM Axiva，并使用了不符合日期条件的发行人；对 EFSL 杠杆的判断也不准确。
+- **C5：76.782 分，提升 28.276 个百分点**。识别出有效发行人，给出评级、11.73% 收益率、发行前后债务权益比，并明确提示高杠杆风险。
+
+### 案例二：巴黎萨克雷经济学硕士与 Charpak 奖学金
+
+任务要求整理项目录取条件、奖学金资格、申请时间和材料。
+
+- **old：39.894 分**。遗漏全日制限制、往届获奖者和博士生不可申请等关键条件，也没有覆盖 M2 申请及推荐信要求。
+- **C5：66.489 分，提升 26.596 个百分点**。补齐交换、实习、研究项目、博士生和往届获奖者等限制，并说明 M2、推荐信及英语证明要求。
+
+### 案例三：澳大利亚半退休城市选择
+
+任务要求比较 Toowoomba、Bundaberg 和 Cairns 的生活成本、会计岗位、房价及房颤专科医疗条件。
+
+- **old：49.863 分**。错误判断 Cairns 缺少本地电生理和消融能力，对 Toowoomba 的心脏服务及城市间价格口径说明也不完整。
+- **C5：68.493 分，提升 18.630 个百分点**。确认 Cairns 本地电生理和消融服务，补齐 Toowoomba 的介入、起搏器、ICD 和消融能力，并解释不同房价数据的时间与统计口径差异。
+
+本次配置更新及实验摘要见 [PR #1709](https://github.com/TokenRhythm/opensquilla/pull/1709)。完整实验报告及原始答案未包含在本仓库中。
diff --git a/opensquilla-webui/src/components/setup/SetupModelStrategyPanel.test.ts b/opensquilla-webui/src/components/setup/SetupModelStrategyPanel.test.ts
index c430e7c148..acf9384d01 100644
--- a/opensquilla-webui/src/components/setup/SetupModelStrategyPanel.test.ts
+++ b/opensquilla-webui/src/components/setup/SetupModelStrategyPanel.test.ts
@@ -1587,19 +1587,19 @@ describe('SetupModelStrategyPanel', () => {
         fixedProfile: {
           providerLabel: 'OpenRouter',
           proposers: [
-            { key: 'openrouter-fixed:proposer:openrouter:deepseek/deepseek-v4-pro', provider: 'openrouter', model: 'deepseek/deepseek-v4-pro', source: 'openrouter_fixed', enabled: true, role: '' },
-            { key: 'openrouter-fixed:proposer:openrouter:z-ai/glm-5.2', provider: 'openrouter', model: 'z-ai/glm-5.2', source: 'openrouter_fixed', enabled: true, role: '' },
-            { key: 'openrouter-fixed:proposer:openrouter:moonshotai/kimi-k2.7-code', provider: 'openrouter', model: 'moonshotai/kimi-k2.7-code', source: 'openrouter_fixed', enabled: true, role: '' },
-            { key: 'openrouter-fixed:proposer:openrouter:qwen/qwen3.7-max', provider: 'openrouter', model: 'qwen/qwen3.7-max', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:deepseek/deepseek-v4.1-flash', provider: 'openrouter', model: 'deepseek/deepseek-v4.1-flash', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:z-ai/glm-5.3-flash', provider: 'openrouter', model: 'z-ai/glm-5.3-flash', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:qwen/qwen3.8-flash', provider: 'openrouter', model: 'qwen/qwen3.8-flash', source: 'openrouter_fixed', enabled: true, role: '' },
+            { key: 'openrouter-fixed:proposer:openrouter:qwen/qwen3.8-max-0902', provider: 'openrouter', model: 'qwen/qwen3.8-max-0902', source: 'openrouter_fixed', enabled: true, role: '' },
           ],
-          aggregator: { key: 'openrouter-fixed:aggregator:openrouter:z-ai/glm-5.2', provider: 'openrouter', model: 'z-ai/glm-5.2', source: 'openrouter_fixed', enabled: true, role: 'aggregator' },
+          aggregator: { key: 'openrouter-fixed:aggregator:openrouter:deepseek/deepseek-v4.1-flash', provider: 'openrouter', model: 'deepseek/deepseek-v4.1-flash', source: 'openrouter_fixed', enabled: true, role: 'aggregator' },
         },
         showCandidateEditor: false,
       },
     }, { onUpdateEnsembleScheme })
 
-    expect(el.textContent).toContain('deepseek/deepseek-v4-pro')
-    expect(el.textContent).toContain('moonshotai/kimi-k2.7-code')
+    expect(el.textContent).toContain('deepseek/deepseek-v4.1-flash')
+    expect(el.textContent).toContain('qwen/qwen3.8-max-0902')
     expect(el.textContent).toContain('Aggregator')
     expect(el.querySelector('.setup-model-strategy__ensemble > .control-section__head')).toBeNull()
     expect(el.textContent).not.toContain('Models draft in parallel')
diff --git a/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.test.ts b/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.test.ts
index cc70e01beb..4d40d2e01a 100644
--- a/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.test.ts
+++ b/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.test.ts
@@ -259,7 +259,11 @@ describe('useSetupEnsembleForm — scheme switching', () => {
     expect(aggregators[0]!.model).toBe(OPENROUTER_FIXED_ENSEMBLE_AGGREGATOR)
     const proposers = f.candidates.value.filter(c => c.role !== 'aggregator')
     expect(proposers.map(c => c.model)).toEqual([...OPENROUTER_FIXED_ENSEMBLE_PROPOSERS])
+    expect(f.candidates.value.map(c => c.thinking_level)).toEqual(Array(5).fill('high'))
     expect(f.payload().selectionMode).toBe(CUSTOM_B5_SELECTION_MODE)
+    expect(f.payload().candidates).toEqual(expect.arrayContaining([
+      expect.objectContaining({ thinking_level: 'high' }),
+    ]))
   })
 
   it('switching to custom seeds from the TokenRhythm lineup for tokenrhythm', () => {
@@ -276,6 +280,7 @@ describe('useSetupEnsembleForm — scheme switching', () => {
     expect(proposers.map(c => c.model)).toEqual([...TOKENRHYTHM_FIXED_ENSEMBLE_PROPOSERS])
     expect(f.candidates.value.find(c => c.role === 'aggregator')!.model)
       .toBe(TOKENRHYTHM_FIXED_ENSEMBLE_AGGREGATOR)
+    expect(f.candidates.value.map(c => c.thinking_level)).toEqual(Array(5).fill('high'))
   })
 
   it('switching back to preset restores the baseline candidate inputs', () => {
@@ -294,15 +299,12 @@ describe('useSetupEnsembleForm — scheme switching', () => {
     expect(f.isDirty.value).toBe(false)
   })
 
-  it('activateForProvider materializes a legacy preset-provider plan as custom', () => {
+  it('activateForProvider selects the provider preset for a legacy plan', () => {
     const f = useSetupEnsembleForm()
     f.initFromConfig({ selection_mode: 'router_dynamic' })
     f.activateForProvider('tokenrhythm')
-    expect(f.selectionMode.value).toBe(CUSTOM_B5_SELECTION_MODE)
-    expect(f.candidates.value.filter(c => c.role !== 'aggregator').map(c => c.model))
-      .toEqual([...TOKENRHYTHM_FIXED_ENSEMBLE_PROPOSERS])
-    expect(f.candidates.value.find(c => c.role === 'aggregator')?.model)
-      .toBe(TOKENRHYTHM_FIXED_ENSEMBLE_AGGREGATOR)
+    expect(f.selectionMode.value).toBe('static_tokenrhythm_b5')
+    expect(f.candidates.value).toEqual([])
   })
 
   it('activateForProvider gives other providers an explicit custom lineup seeded from tiers', () => {
diff --git a/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.ts b/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.ts
index 41b924c7ec..dcade86c34 100644
--- a/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.ts
+++ b/opensquilla-webui/src/composables/setup/useSetupEnsembleForm.ts
@@ -83,6 +83,7 @@ export interface EnsembleCandidateConfig {
   source?: 'custom' | 'legacy_model_options'
   enabled?: boolean
   role?: string
+  thinking_level?: string
 }
 
 export interface EnsembleRoutingModeState {
@@ -267,6 +268,8 @@ function normalizeCandidates(value: unknown): EnsembleCandidateConfig[] {
       enabled: raw.enabled === false ? false : true,
       role,
     }
+    const thinkingLevel = String(raw.thinking_level ?? raw.thinkingLevel ?? '').trim()
+    if (thinkingLevel) normalized.thinking_level = thinkingLevel
     const existingIndex = seen.get(key)
     if (existingIndex === undefined) {
       seen.set(key, out.length)
@@ -306,6 +309,7 @@ function customSeedFromProfile(profile: StaticB5Profile): EnsembleCandidateConfi
     source: 'custom',
     enabled: true,
     role: 'proposer',
+    ...(profile.thinkingLevel ? { thinking_level: profile.thinkingLevel } : {}),
   }))
   rows.push({
     provider: profile.provider,
@@ -313,6 +317,7 @@ function customSeedFromProfile(profile: StaticB5Profile): EnsembleCandidateConfi
     source: 'custom',
     enabled: true,
     role: 'aggregator',
+    ...(profile.thinkingLevel ? { thinking_level: profile.thinkingLevel } : {}),
   })
   return normalizeCandidates(rows)
 }
@@ -818,13 +823,14 @@ export function useSetupEnsembleForm() {
       && candidates.value.some(candidate => candidate.enabled !== false)
     ) return
     const presetMode = staticB5ModeForProvider(provider)
-    selectionMode.value = CUSTOM_B5_SELECTION_MODE
-    if (candidates.value.some(candidate => candidate.enabled !== false)) return
-    const profile = presetMode ? STATIC_B5_PROFILES[presetMode] : null
-    if (profile) {
-      candidates.value = customSeedFromProfile(profile)
+    if (presetMode) {
+      selectionMode.value = presetMode
+      modelOptions.value = []
+      candidates.value = []
       return
     }
+    selectionMode.value = CUSTOM_B5_SELECTION_MODE
+    if (candidates.value.some(candidate => candidate.enabled !== false)) return
     importTierCandidates(tierCandidates)
   }
 
@@ -909,6 +915,7 @@ export function useSetupEnsembleForm() {
       source: candidate.source || 'custom',
       enabled: candidate.enabled !== false,
       role: normalizeCandidateRole(candidate.role),
+      ...(candidate.thinking_level ? { thinking_level: candidate.thinking_level } : {}),
     }))
     if (minSuccessfulDirty.value) params.minSuccessfulProposers = minSuccessfulProposers.value
     if (allFailedPolicyDirty.value) params.allFailedPolicy = allFailedPolicy.value
diff --git a/opensquilla-webui/src/types/generated/router_tier_contract.test.ts b/opensquilla-webui/src/types/generated/router_tier_contract.test.ts
index 2b05b7fa4a..29114da00d 100644
--- a/opensquilla-webui/src/types/generated/router_tier_contract.test.ts
+++ b/opensquilla-webui/src/types/generated/router_tier_contract.test.ts
@@ -15,12 +15,17 @@ describe('generated router-tier SelectionMode contract', () => {
     ])
     expect(staticB5ModeForProvider('OpenRouter')).toBe('static_openrouter_b5')
     expect(staticB5ModeForProvider('tokenrhythm')).toBe('static_tokenrhythm_b5')
+    expect(STATIC_B5_PROFILES.static_openrouter_b5?.thinkingLevel).toBe('high')
+    expect(STATIC_B5_PROFILES.static_tokenrhythm_b5?.thinkingLevel).toBe('high')
   })
 
   it('keeps provider recommendations and ownership sets in generated data', () => {
     expect(PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES.tokenrhythm).toBe(
       'static_tokenrhythm_b5',
     )
+    expect(PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES.openrouter).toBe(
+      'static_openrouter_b5',
+    )
     expect(DORMANT_SHARED_SELECTION_MODES).toEqual([
       'static_openrouter_b5',
       'static_tokenrhythm_b5',
diff --git a/opensquilla-webui/src/types/generated/router_tier_contract.ts b/opensquilla-webui/src/types/generated/router_tier_contract.ts
index bfd49cc30a..0ea414fdb6 100644
--- a/opensquilla-webui/src/types/generated/router_tier_contract.ts
+++ b/opensquilla-webui/src/types/generated/router_tier_contract.ts
@@ -26,6 +26,7 @@ export interface StaticB5Profile {
   proposers: readonly string[]
   aggregator: string
   apiKeyEnv: string
+  thinkingLevel: string | null
   ownershipRole: string
 }
 
@@ -33,17 +34,19 @@ export const STATIC_B5_PROFILES: Record<string, StaticB5Profile> = {
   "static_openrouter_b5": {
     provider: "openrouter",
     label: "OpenRouter",
-    proposers: ["deepseek/deepseek-v4-pro", "z-ai/glm-5.2", "moonshotai/kimi-k2.7-code", "qwen/qwen3.7-max"] as const,
-    aggregator: "z-ai/glm-5.2",
+    proposers: ["deepseek/deepseek-v4.1-flash", "z-ai/glm-5.3-flash", "qwen/qwen3.8-flash", "qwen/qwen3.8-max-0902"] as const,
+    aggregator: "deepseek/deepseek-v4.1-flash",
     apiKeyEnv: "OPENROUTER_API_KEY",
+    thinkingLevel: "high",
     ownershipRole: "static_profile",
   },
   "static_tokenrhythm_b5": {
     provider: "tokenrhythm",
     label: "TokenRhythm",
-    proposers: ["deepseek-v4-pro", "glm-5.2", "kimi-k2.7-code", "qwen3.7-max"] as const,
-    aggregator: "glm-5.2",
+    proposers: ["deepseek-flash", "glm-5.3-flash", "qwen3.8-flash", "qwen3.8-max"] as const,
+    aggregator: "deepseek-flash",
     apiKeyEnv: "TOKENRHYTHM_API_KEY",
+    thinkingLevel: "high",
     ownershipRole: "static_profile",
   },
 }
@@ -51,7 +54,7 @@ export const STATIC_B5_PROFILES: Record<string, StaticB5Profile> = {
 export const STATIC_B5_SELECTION_MODE_PROVIDERS:
   Readonly<Record<string, string>> = {"static_openrouter_b5": "openrouter", "static_tokenrhythm_b5": "tokenrhythm"}
 export const PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES:
-  Readonly<Record<string, string>> = {"tokenrhythm": "static_tokenrhythm_b5"}
+  Readonly<Record<string, string>> = {"openrouter": "static_openrouter_b5", "tokenrhythm": "static_tokenrhythm_b5"}
 export const SELECTION_MODE_OWNERSHIP_ROLES:
   Readonly<Record<string, string>> = {"static_openrouter_b5": "static_profile", "static_tokenrhythm_b5": "static_profile", "custom_b5": "custom_profile", "router_dynamic": "router_dynamic"}
 
diff --git a/scripts/generate_router_tier_contract.py b/scripts/generate_router_tier_contract.py
index b9c79bd1e6..c378fd3296 100644
--- a/scripts/generate_router_tier_contract.py
+++ b/scripts/generate_router_tier_contract.py
@@ -63,6 +63,7 @@ def _profile_block() -> str:
             + f"    proposers: {_json(list(profile.proposer_models))} as const,\n"
             + f"    aggregator: {_json(profile.aggregator_model)},\n"
             + f"    apiKeyEnv: {_json(profile.api_key_env)},\n"
+            + f"    thinkingLevel: {_json(profile.thinking_level)},\n"
             + f"    ownershipRole: {_json(profile.ownership_role)},\n"
             + "  },"
         )
@@ -111,6 +112,7 @@ def render() -> str:
   proposers: readonly string[]
   aggregator: string
   apiKeyEnv: string
+  thinkingLevel: string | null
   ownershipRole: string
 }}
 
diff --git a/src/opensquilla/gateway/model_routing.py b/src/opensquilla/gateway/model_routing.py
index fe21332138..7de5dcc5f6 100644
--- a/src/opensquilla/gateway/model_routing.py
+++ b/src/opensquilla/gateway/model_routing.py
@@ -235,6 +235,7 @@ def _custom_candidate(
     model: str,
     *,
     role: str = "proposer",
+    thinking_level: str | None = None,
 ) -> dict[str, Any]:
     return {
         "provider": provider,
@@ -242,7 +243,7 @@ def _custom_candidate(
         "source": "custom",
         "enabled": True,
         "role": role,
-        "thinking_level": "",
+        "thinking_level": str(thinking_level or "").strip(),
     }
 
 
@@ -271,7 +272,11 @@ def _provider_ensemble_candidates(config: Any) -> list[dict[str, Any]]:
     )
     if static_profile is not None:
         static_candidates = [
-            _custom_candidate(static_profile.provider_id, model)
+            _custom_candidate(
+                static_profile.provider_id,
+                model,
+                thinking_level=static_profile.thinking_level,
+            )
             for model in static_profile.proposer_models
         ]
         static_candidates.append(
@@ -279,6 +284,7 @@ def _provider_ensemble_candidates(config: Any) -> list[dict[str, Any]]:
                 static_profile.provider_id,
                 static_profile.aggregator_model,
                 role="aggregator",
+                thinking_level=static_profile.thinking_level,
             )
         )
         return static_candidates
diff --git a/src/opensquilla/provider/catalog_overrides.toml b/src/opensquilla/provider/catalog_overrides.toml
index 4e81df08fa..f110dfc517 100644
--- a/src/opensquilla/provider/catalog_overrides.toml
+++ b/src/opensquilla/provider/catalog_overrides.toml
@@ -115,6 +115,55 @@ max_output_tokens = 16384
 context_window = 256000
 max_output_tokens = 16384
 
+# C5 ensemble models. Budgets, capabilities, and base prices mirror
+# OpenRouter's public catalog; live catalog data remains authoritative.
+
+[openrouter."deepseek/deepseek-v4.1-flash"]
+context_window = 1048576
+max_output_tokens = 384000
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "openrouter"
+input_cost_per_mtok = 0.15
+output_cost_per_mtok = 0.6
+cache_read_cost_per_mtok = 0.003
+
+[openrouter."z-ai/glm-5.3-flash"]
+context_window = 1310720
+max_output_tokens = 131072
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "openrouter"
+input_cost_per_mtok = 0.09
+output_cost_per_mtok = 0.3
+cache_read_cost_per_mtok = 0.018
+
+[openrouter."qwen/qwen3.8-flash"]
+context_window = 1000000
+max_output_tokens = 131072
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "openrouter"
+input_cost_per_mtok = 0.15
+output_cost_per_mtok = 0.47
+cache_read_cost_per_mtok = 0.016
+cache_write_cost_per_mtok = 0.2
+
+[openrouter."qwen/qwen3.8-max-0902"]
+context_window = 1000000
+max_output_tokens = 131072
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "openrouter"
+input_cost_per_mtok = 2.0
+output_cost_per_mtok = 6.0
+cache_read_cost_per_mtok = 0.25
+cache_write_cost_per_mtok = 2.5
+
 # Official https://openrouter.ai/api/v1/models, verified 2026-09-17.
 # Use the smaller catalog/top-provider context when both are published.
 [openrouter."deepseek/deepseek-v4-flash"]
@@ -727,6 +776,42 @@ reasoning_format = "qwen_token_plan"
 # changing this neutral catalog shape. Vision is published for the kimi rows
 # and qwen3.8-max; explicit API declarations remain authoritative.
 
+[tokenrhythm."deepseek-flash"]
+context_window = 1000000
+max_output_tokens = 384000
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "none"
+status = "testing"
+input_cost_per_mtok = 0.2867383512544803
+output_cost_per_mtok = 1.1469534050179212
+cache_read_cost_per_mtok = 0.005734767025089606
+
+[tokenrhythm."glm-5.3-flash"]
+context_window = 1048576
+max_output_tokens = 131072
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "none"
+status = "testing"
+input_cost_per_mtok = 0.11469534050179211
+output_cost_per_mtok = 0.4014336917562724
+cache_read_cost_per_mtok = 0.03297491039426523
+
+[tokenrhythm."qwen3.8-flash"]
+context_window = 1000000
+max_output_tokens = 131072
+supports_reasoning = true
+supports_tools = true
+supports_vision = true
+reasoning_format = "none"
+status = "testing"
+input_cost_per_mtok = 0.11469534050179211
+output_cost_per_mtok = 0.3870967741935484
+cache_read_cost_per_mtok = 0.014336917562724014
+
 [tokenrhythm."deepseek-v4-flash"]
 context_window = 1000000
 max_output_tokens = 384000
diff --git a/src/opensquilla/provider/ensemble.py b/src/opensquilla/provider/ensemble.py
index a298250080..a7a89f1f46 100644
--- a/src/opensquilla/provider/ensemble.py
+++ b/src/opensquilla/provider/ensemble.py
@@ -6356,13 +6356,14 @@ def _build_router_dynamic_members(
     return f"{ROUTER_DYNAMIC_SELECTION_MODE}/{routed_tier}", proposers, aggregator, plan
 
 
-def _static_b5_ref(provider_id: str, model: str) -> _DynamicModelRef:
+def _static_b5_ref(profile: StaticB5Profile, model: str) -> _DynamicModelRef:
     # Thinking belongs to the shared plan, not to the C3 tier that happened to
-    # activate it.  OpenRouter's shipped B5 plan preserves the historical C3
-    # high-reasoning default. TokenRhythm rejects thinking-toggle request
-    # fields, so leaving this unset preserves the provider's default payload.
-    thinking = "high" if provider_id == "openrouter" else None
-    return _DynamicModelRef(provider=provider_id, model=model, thinking=thinking)
+    # activate it. Each profile declares its own provider-compatible override.
+    return _DynamicModelRef(
+        provider=profile.provider_id,
+        model=model,
+        thinking=profile.thinking_level,
+    )
 
 
 def _static_default_if_legacy(
@@ -6387,7 +6388,7 @@ def _build_static_b5_members(
 ) -> tuple[str, list[EnsembleMemberConfig], EnsembleMemberConfig, dict[str, Any]]:
     proposers = [
         _member_from_ref(
-            _static_b5_ref(profile.provider_id, model),
+            _static_b5_ref(profile, model),
             config=config,
             inherited=inherited_provider_config,
             label=f"proposer_{index + 1}",
@@ -6397,7 +6398,7 @@ def _build_static_b5_members(
         for index, model in enumerate(profile.proposer_models)
     ]
     aggregator = _member_from_ref(
-        _static_b5_ref(profile.provider_id, profile.aggregator_model),
+        _static_b5_ref(profile, profile.aggregator_model),
         config=config,
         inherited=inherited_provider_config,
         label="aggregator",
@@ -6671,7 +6672,7 @@ def static_b5_credential_available(
             profile.provider_id,
             model,
             inherited_provider_config=inherited,
-            overrides=_static_b5_ref(profile.provider_id, model),
+            overrides=_static_b5_ref(profile, model),
             credential_pool_acquirer=credential_pool_acquirer,
             session_key=session_key,
         ).ready
diff --git a/src/opensquilla/router_tiers.py b/src/opensquilla/router_tiers.py
index 8d78aa52f8..d6d8b0ae38 100644
--- a/src/opensquilla/router_tiers.py
+++ b/src/opensquilla/router_tiers.py
@@ -37,6 +37,7 @@ class StaticB5Profile:
     aggregator_model: str
     label: str
     api_key_env: str
+    thinking_level: str | None = None
     ownership_role: str = "static_profile"
 
 
@@ -45,27 +46,29 @@ class StaticB5Profile:
         profile_name=STATIC_OPENROUTER_B5_SELECTION_MODE,
         provider_id="openrouter",
         proposer_models=(
-            "deepseek/deepseek-v4-pro",
-            "z-ai/glm-5.2",
-            "moonshotai/kimi-k2.7-code",
-            "qwen/qwen3.7-max",
+            "deepseek/deepseek-v4.1-flash",
+            "z-ai/glm-5.3-flash",
+            "qwen/qwen3.8-flash",
+            "qwen/qwen3.8-max-0902",
         ),
-        aggregator_model="z-ai/glm-5.2",
+        aggregator_model="deepseek/deepseek-v4.1-flash",
         label="OpenRouter",
         api_key_env="OPENROUTER_API_KEY",
+        thinking_level="high",
     ),
     STATIC_TOKENRHYTHM_B5_SELECTION_MODE: StaticB5Profile(
         profile_name=STATIC_TOKENRHYTHM_B5_SELECTION_MODE,
         provider_id="tokenrhythm",
         proposer_models=(
-            "deepseek-v4-pro",
-            "glm-5.2",
-            "kimi-k2.7-code",
-            "qwen3.7-max",
+            "deepseek-flash",
+            "glm-5.3-flash",
+            "qwen3.8-flash",
+            "qwen3.8-max",
         ),
-        aggregator_model="glm-5.2",
+        aggregator_model="deepseek-flash",
         label="TokenRhythm",
         api_key_env="TOKENRHYTHM_API_KEY",
+        thinking_level="high",
     ),
 }
 
@@ -80,6 +83,7 @@ class StaticB5Profile:
     ROUTER_DYNAMIC_SELECTION_MODE: "router_dynamic",
 }
 PROVIDER_RECOMMENDED_ENSEMBLE_SELECTION_MODES: dict[str, str] = {
+    "openrouter": STATIC_OPENROUTER_B5_SELECTION_MODE,
     "tokenrhythm": STATIC_TOKENRHYTHM_B5_SELECTION_MODE,
 }
 ENSEMBLE_SELECTION_MODE_ORDER = (
diff --git a/tests/test_desktop/test_electron_startup_contract.py b/tests/test_desktop/test_electron_startup_contract.py
index 19b245ce74..84fc60aee5 100644
--- a/tests/test_desktop/test_electron_startup_contract.py
+++ b/tests/test_desktop/test_electron_startup_contract.py
@@ -1561,9 +1561,9 @@ def test_desktop_tokenrhythm_single_page_onboarding_defaults_to_router() -> None
     assert "routerDefaultTier: 'c1'," in router_config
     assert "routerPresetBinding: 'follow_primary'," in router_config
     assert "[data-model-routing-mode]" not in onboarding_html
-    assert "'selection_mode = \"custom_b5\"'" in main_ts
-    assert "'[[llm_ensemble.candidates]]'" in main_ts
-    assert "DESKTOP_ENSEMBLE_PROFILES[selectionMode]" in main_ts
+    assert "`selection_mode = ${tomlString(selectionMode)}`" in main_ts
+    assert "'[[llm_ensemble.candidates]]'" not in main_ts
+    assert "DESKTOP_ENSEMBLE_PROFILES" not in main_ts
 
     expected_models = (
         "deepseek-v4-flash-0731",
diff --git a/tests/test_engine/test_agent_message_count_recovery.py b/tests/test_engine/test_agent_message_count_recovery.py
index 90cd847b3b..24e51b33ec 100644
--- a/tests/test_engine/test_agent_message_count_recovery.py
+++ b/tests/test_engine/test_agent_message_count_recovery.py
@@ -464,10 +464,10 @@ async def list_models(self) -> list[Any]:
 
 class _EnsembleBudgetCatalog:
     _WINDOWS = {
-        "deepseek-v4-pro": 1_000_000,
-        "glm-5.2": 1_000_000,
-        "kimi-k2.7-code": 256_000,
-        "qwen3.7-max": 1_000_000,
+        "deepseek-flash": 1_000_000,
+        "glm-5.3-flash": 1_048_576,
+        "qwen3.8-flash": 1_000_000,
+        "qwen3.8-max": 1_000_000,
     }
 
     def resolve_context_window_with_source(
@@ -1829,18 +1829,17 @@ def _build_member(config: ProviderConfig) -> _CountAwareEnsembleMemberProvider:
     assert sum(call["wire_messages"] > 100 for call in member_calls) == 4
     assert sum(call["wire_messages"] <= 90 for call in member_calls) == 5
     assert sum(call["aggregator"] for call in member_calls) == 1
-    assert any(
-        call["model"] == "kimi-k2.7-code"
-        and call["max_tokens"] == 16_000
-        and call["request_cap"] == 880_000
-        for call in member_calls
-    )
-    assert any(
-        call["model"] == "glm-5.2"
-        and call["max_tokens"] == 128_000
-        and call["request_cap"] == 3_408_000
+    # Failed drafts, compacted retries, and aggregation must retain C5 member
+    # budgets instead of inheriting the outer Kimi route's 256k context limit.
+    assert {
+        (call["model"], call["max_tokens"], call["request_cap"])
         for call in member_calls
-    )
+    } == {
+        ("deepseek-flash", 384_000, 2_384_000),
+        ("glm-5.3-flash", 131_072, 3_590_016),
+        ("qwen3.8-flash", 131_072, 3_395_712),
+        ("qwen3.8-max", 131_072, 3_395_712),
+    }
     assert any(getattr(event, "kind", None) == "done" for event in events)
     assert not any(
         isinstance(event, ErrorEvent)
diff --git a/tests/test_gateway/test_rpc_model_routing.py b/tests/test_gateway/test_rpc_model_routing.py
index 89b8eb6016..3f9d670cc5 100644
--- a/tests/test_gateway/test_rpc_model_routing.py
+++ b/tests/test_gateway/test_rpc_model_routing.py
@@ -63,7 +63,7 @@ def test_fresh_tokenrhythm_ensemble_activation_materializes_recommended_plan() -
     assert "llm_ensemble.candidates" not in changed.force_persist_paths()
 
 
-def test_fresh_openrouter_ensemble_activation_materializes_custom_lineup() -> None:
+def test_fresh_openrouter_ensemble_activation_uses_recommended_static_plan() -> None:
     cfg = GatewayConfig(
         llm={
             "provider": "openrouter",
@@ -73,13 +73,10 @@ def test_fresh_openrouter_ensemble_activation_materializes_custom_lineup() -> No
 
     changed = upsert_llm_ensemble(cfg, enabled=True).config
 
-    assert changed.llm_ensemble.selection_mode == "custom_b5"
-    assert len(changed.llm_ensemble.candidates) == 5
-    assert {
-        candidate.provider for candidate in changed.llm_ensemble.candidates
-    } == {"openrouter"}
-    assert changed.llm_ensemble.candidates[-1].model == "z-ai/glm-5.2"
-    assert changed.llm_ensemble.candidates[-1].role == "aggregator"
+    assert changed.llm_ensemble.selection_mode == "static_openrouter_b5"
+    assert changed.llm_ensemble.candidates == []
+    assert "llm_ensemble.selection_mode" in changed.force_persist_paths()
+    assert "llm_ensemble.candidates" not in changed.force_persist_paths()
 
 
 def test_explicit_cross_provider_ensemble_selection_is_preserved() -> None:
diff --git a/tests/test_llm_ensemble_config.py b/tests/test_llm_ensemble_config.py
index 5a4f6e5863..198f66e4a8 100644
--- a/tests/test_llm_ensemble_config.py
+++ b/tests/test_llm_ensemble_config.py
@@ -47,12 +47,14 @@ def test_llm_ensemble_defaults_to_disabled_for_model_router_first_install() -> N
     )
     assert provider.profile_name == "static_openrouter_b5"
     assert [member.provider_config.model for member in provider.proposers] == [
-        "deepseek/deepseek-v4-pro",
-        "z-ai/glm-5.2",
-        "moonshotai/kimi-k2.7-code",
-        "qwen/qwen3.7-max",
+        "deepseek/deepseek-v4.1-flash",
+        "z-ai/glm-5.3-flash",
+        "qwen/qwen3.8-flash",
+        "qwen/qwen3.8-max-0902",
     ]
-    assert provider.aggregator.provider_config.model == "z-ai/glm-5.2"
+    assert provider.aggregator.provider_config.model == "deepseek/deepseek-v4.1-flash"
+    assert {member.thinking for member in provider.proposers} == {"high"}
+    assert provider.aggregator.thinking == "high"
     # The fresh/default shared policy is a one-draft admission floor.
     assert provider.min_successful_proposers == 1
     assert provider.target_successful_proposers == 1
@@ -85,15 +87,15 @@ def test_static_openrouter_b5_does_not_need_model_options() -> None:
 
     assert provider.profile_name == "static_openrouter_b5"
     assert [member.provider_config.model for member in provider.proposers] == [
-        "deepseek/deepseek-v4-pro",
-        "z-ai/glm-5.2",
-        "moonshotai/kimi-k2.7-code",
-        "qwen/qwen3.7-max",
+        "deepseek/deepseek-v4.1-flash",
+        "z-ai/glm-5.3-flash",
+        "qwen/qwen3.8-flash",
+        "qwen/qwen3.8-max-0902",
     ]
-    assert provider.aggregator.provider_config.model == "z-ai/glm-5.2"
+    assert provider.aggregator.provider_config.model == "deepseek/deepseek-v4.1-flash"
 
 
-def test_static_tokenrhythm_b5_mirrors_the_openrouter_lineup() -> None:
+def test_static_tokenrhythm_b5_keeps_its_provider_specific_lineup() -> None:
     cfg = GatewayConfig(
         llm_ensemble={
             "enabled": True,
@@ -114,16 +116,18 @@ def test_static_tokenrhythm_b5_mirrors_the_openrouter_lineup() -> None:
 
     assert provider.profile_name == "static_tokenrhythm_b5"
     assert [member.provider_config.model for member in provider.proposers] == [
-        "deepseek-v4-pro",
-        "glm-5.2",
-        "kimi-k2.7-code",
-        "qwen3.7-max",
+        "deepseek-flash",
+        "glm-5.3-flash",
+        "qwen3.8-flash",
+        "qwen3.8-max",
     ]
     assert all(
         member.provider_config.provider == "tokenrhythm" for member in provider.proposers
     )
+    assert all(member.thinking == "high" for member in provider.proposers)
     assert provider.aggregator.provider_config.provider == "tokenrhythm"
-    assert provider.aggregator.provider_config.model == "glm-5.2"
+    assert provider.aggregator.provider_config.model == "deepseek-flash"
+    assert provider.aggregator.thinking == "high"
     # Same aggregation defaults as the static OpenRouter profile.
     assert provider.min_successful_proposers == 1
     assert provider.proposer_timeout_seconds == 120.0
@@ -443,10 +447,10 @@ def test_static_openrouter_b5_ensemble_locks_members_across_routed_tiers() -> No
         provider_routing={"z-ai/glm-5.2": "z-ai"},
     )
     expected_proposers = [
-        "deepseek/deepseek-v4-pro",
-        "z-ai/glm-5.2",
-        "moonshotai/kimi-k2.7-code",
-        "qwen/qwen3.7-max",
+        "deepseek/deepseek-v4.1-flash",
+        "z-ai/glm-5.3-flash",
+        "qwen/qwen3.8-flash",
+        "qwen/qwen3.8-max-0902",
     ]
 
     for tier in ("c0", "c1", "c2", "c3"):
@@ -459,12 +463,12 @@ def test_static_openrouter_b5_ensemble_locks_members_across_routed_tiers() -> No
 
         assert provider.profile_name == "static_openrouter_b5"
         assert [member.provider_config.model for member in provider.proposers] == expected_proposers
-        assert provider.aggregator.provider_config.model == "z-ai/glm-5.2"
+        assert provider.aggregator.provider_config.model == "deepseek/deepseek-v4.1-flash"
         assert provider.selection_plan == {
             "strategy": "static_openrouter_b5",
             "profile": "static_openrouter_b5",
             "proposer_models": expected_proposers,
-            "aggregator_model": "z-ai/glm-5.2",
+            "aggregator_model": "deepseek/deepseek-v4.1-flash",
             "proposer_count": 4,
             "configured_min_successful_proposers": 9,
             "effective_min_successful_proposers": 4,
diff --git a/tests/test_provider/test_catalog_layers.py b/tests/test_provider/test_catalog_layers.py
index 5eaf965e13..5db3cbc570 100644
--- a/tests/test_provider/test_catalog_layers.py
+++ b/tests/test_provider/test_catalog_layers.py
@@ -321,6 +321,10 @@ def test_packaged_corrections_file_parses_with_expected_tables() -> None:
     assert set(payload["openrouter"]) == {
         "anthropic/claude-opus-4.8",
         "anthropic/claude-sonnet-4.6",
+        "deepseek/deepseek-v4.1-flash",
+        "qwen/qwen3.8-flash",
+        "qwen/qwen3.8-max-0902",
+        "z-ai/glm-5.3-flash",
         "x-ai/grok-4.3",
         "stepfun/step-3.5-flash",
         "deepseek/deepseek-v4-flash",
diff --git a/tests/test_provider_ensemble.py b/tests/test_provider_ensemble.py
index e7a00b3033..993968a462 100644
--- a/tests/test_provider_ensemble.py
+++ b/tests/test_provider_ensemble.py
@@ -401,10 +401,11 @@ def __init__(
         windows: dict[str, tuple[int, str] | Exception] | None = None,
     ) -> None:
         self.windows = windows or {
-            "deepseek-v4-pro": (1_000_000, "catalog"),
-            "glm-5.2": (1_000_000, "catalog"),
+            "deepseek-flash": (1_000_000, "catalog"),
+            "glm-5.3-flash": (1_048_576, "catalog"),
+            "qwen3.8-flash": (1_000_000, "catalog"),
+            "qwen3.8-max": (1_000_000, "catalog"),
             "kimi-k2.7-code": (256_000, "catalog"),
-            "qwen3.7-max": (1_000_000, "catalog"),
         }
 
     def _resolve(self, model_id: str) -> tuple[int, str]:
@@ -475,7 +476,7 @@ def test_ensemble_budget_uses_each_tokenrhythm_authority_window(
 
 
 def _tokenrhythm_budget_registry() -> _FakeRegistry:
-    models = ("deepseek-v4-pro", "glm-5.2", "kimi-k2.7-code", "qwen3.7-max")
+    models = ("deepseek-flash", "glm-5.3-flash", "qwen3.8-flash", "qwen3.8-max")
     return _FakeRegistry(
         {
             model: _FakePlan(
@@ -3288,28 +3289,35 @@ async def test_tokenrhythm_ensemble_rebinds_request_cap_per_member_context(
     ]
 
     calls_by_model = {call["model"]: call["config"] for call in registry.calls}
-    # Kimi's 256k window and 16k output yield 880,000 chars; GLM's 1m window yields
-    # 3,408,000. Parameterizing the inherited cap pins both widening and
-    # tightening instead of relying on the outer route's model.
-    assert calls_by_model["kimi-k2.7-code"].provider_request_max_chars == 880_000
-    assert calls_by_model["glm-5.2"].provider_request_max_chars == 3_408_000
+    # The models share roughly 1m context windows but reserve different output
+    # budgets plus the static profile's explicit high-thinking allowance.
+    # Parameterizing the inherited cap pins both widening and tightening
+    # instead of relying on the outer route's model.
+    assert calls_by_model["qwen3.8-flash"].provider_request_max_chars == 3_375_712
+    assert calls_by_model["deepseek-flash"].provider_request_max_chars == 2_364_000
+    assert all(call["config"].thinking is True for call in registry.calls)
+    assert all(call["config"].thinking_level == "high" for call in registry.calls)
 
     done = next(event for event in events if isinstance(event, DoneEvent))
     assert done.ensemble_trace is not None
-    kimi_trace = next(
+    qwen_trace = next(
         candidate["execution"]
         for candidate in done.ensemble_trace["candidates"]
-        if candidate["model"] == "kimi-k2.7-code"
-    )
-    assert kimi_trace["effective_context_window_tokens"] == 256_000
-    assert kimi_trace["effective_context_window_source"] == "catalog"
-    assert kimi_trace["effective_provider_request_max_chars"] == 880_000
-    assert kimi_trace["provider_request_max_chars_source"] == "member_context"
+        if candidate["model"] == "qwen3.8-flash"
+    )
+    assert qwen_trace["effective_context_window_tokens"] == 1_000_000
+    assert qwen_trace["effective_context_window_source"] == "catalog"
+    assert qwen_trace["effective_provider_request_max_chars"] == 3_375_712
+    assert qwen_trace["provider_request_max_chars_source"] == "member_context"
+    assert qwen_trace["effective_thinking"] is True
+    assert qwen_trace["effective_thinking_level"] == "high"
     aggregator_trace = done.ensemble_trace["final_request"]["execution"]
     assert aggregator_trace["effective_context_window_tokens"] == 1_000_000
     assert aggregator_trace["effective_context_window_source"] == "catalog"
-    assert aggregator_trace["effective_provider_request_max_chars"] == 3_408_000
+    assert aggregator_trace["effective_provider_request_max_chars"] == 2_364_000
     assert aggregator_trace["provider_request_max_chars_source"] == "member_context"
+    assert aggregator_trace["effective_thinking"] is True
+    assert aggregator_trace["effective_thinking_level"] == "high"
 
 
 @pytest.mark.parametrize(
@@ -3409,10 +3417,10 @@ async def test_ensemble_member_context_precedence_is_override_then_global_then_c
     monkeypatch.setattr("opensquilla.provider.ensemble._build_provider", registry.provider_for)
     catalog = _BudgetCatalog(
         {
-            "deepseek-v4-pro": (1_000_000, "catalog"),
-            "glm-5.2": (1_000_000, "catalog"),
-            "kimi-k2.7-code": (300_000, "override"),
-            "qwen3.7-max": (1_000_000, "catalog"),
+            "deepseek-flash": (1_000_000, "catalog"),
+            "glm-5.3-flash": (1_048_576, "catalog"),
+            "qwen3.8-flash": (300_000, "override"),
+            "qwen3.8-max": (1_000_000, "catalog"),
         }
     )
     provider = _build_tokenrhythm_budget_provider(
@@ -3433,17 +3441,17 @@ async def test_ensemble_member_context_precedence_is_override_then_global_then_c
     ]
 
     calls_by_model = {call["model"]: call["config"] for call in registry.calls}
-    assert calls_by_model["kimi-k2.7-code"].provider_request_max_chars == 1_056_000
-    assert calls_by_model["glm-5.2"].provider_request_max_chars == 1_408_000
+    assert calls_by_model["qwen3.8-flash"].provider_request_max_chars == 575_712
+    assert calls_by_model["deepseek-flash"].provider_request_max_chars == 364_000
     done = next(event for event in events if isinstance(event, DoneEvent))
     assert done.ensemble_trace is not None
-    kimi_trace = next(
+    qwen_trace = next(
         candidate["execution"]
         for candidate in done.ensemble_trace["candidates"]
-        if candidate["model"] == "kimi-k2.7-code"
+        if candidate["model"] == "qwen3.8-flash"
     )
-    assert kimi_trace["effective_context_window_source"] == "override"
-    assert kimi_trace["effective_context_window_tokens"] == 300_000
+    assert qwen_trace["effective_context_window_source"] == "override"
+    assert qwen_trace["effective_context_window_tokens"] == 300_000
     aggregator_trace = done.ensemble_trace["final_request"]["execution"]
     assert aggregator_trace["effective_context_window_source"] == "config"
     assert aggregator_trace["effective_context_window_tokens"] == 500_000
@@ -4044,10 +4052,10 @@ async def test_ensemble_default_context_rebinds_but_catalog_failure_retains_oute
     monkeypatch.setattr("opensquilla.provider.ensemble._build_provider", registry.provider_for)
     catalog = _BudgetCatalog(
         {
-            "deepseek-v4-pro": (1_000_000, "catalog"),
-            "glm-5.2": RuntimeError("catalog unavailable"),
-            "kimi-k2.7-code": (256_000, "default"),
-            "qwen3.7-max": (1_000_000, "catalog"),
+            "deepseek-flash": RuntimeError("catalog unavailable"),
+            "glm-5.3-flash": (1_048_576, "catalog"),
+            "qwen3.8-flash": (1_000_000, "default"),
+            "qwen3.8-max": (1_000_000, "catalog"),
         }
     )
     provider = _build_tokenrhythm_budget_provider(catalog=catalog)
@@ -4065,20 +4073,20 @@ async def test_ensemble_default_context_rebinds_but_catalog_failure_retains_oute
     ]
 
     calls_by_model = {call["model"]: call["config"] for call in registry.calls}
-    # Automatic output follows this deployment's fallback, without borrowing
-    # a different authority's catalog. The member's context is still rebound.
-    assert calls_by_model["kimi-k2.7-code"].max_tokens == 16_000
-    assert calls_by_model["kimi-k2.7-code"].provider_request_max_chars == 880_000
-    assert calls_by_model["glm-5.2"].provider_request_max_chars == 555_555
+    # The C5 member keeps its model-specific output ceiling while rebinding
+    # the default context; a catalog error still preserves the outer cap.
+    assert calls_by_model["qwen3.8-flash"].max_tokens == 131_072
+    assert calls_by_model["qwen3.8-flash"].provider_request_max_chars == 3_375_712
+    assert calls_by_model["deepseek-flash"].provider_request_max_chars == 555_555
     done = next(event for event in events if isinstance(event, DoneEvent))
     assert done.ensemble_trace is not None
-    kimi_trace = next(
+    qwen_trace = next(
         candidate["execution"]
         for candidate in done.ensemble_trace["candidates"]
-        if candidate["model"] == "kimi-k2.7-code"
+        if candidate["model"] == "qwen3.8-flash"
     )
-    assert kimi_trace["effective_context_window_source"] == "default"
-    assert kimi_trace["provider_request_max_chars_source"] == "member_context"
+    assert qwen_trace["effective_context_window_source"] == "default"
+    assert qwen_trace["provider_request_max_chars_source"] == "member_context"
     aggregator_trace = done.ensemble_trace["final_request"]["execution"]
     assert aggregator_trace["effective_context_window_source"] == "error"
     assert aggregator_trace["provider_request_max_chars_source"] == "inherited"
@@ -4088,7 +4096,7 @@ async def test_ensemble_default_context_rebinds_but_catalog_failure_retains_oute
 async def test_rebinding_rebinds_fallback_chat_config(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    models = ("deepseek-v4-pro", "glm-5.2", "kimi-k2.7-code", "qwen3.7-max")
+    models = ("deepseek-flash", "glm-5.3-flash", "qwen3.8-flash", "qwen3.8-max")
     registry = _FakeRegistry(
         {
             model: _FakePlan([ErrorEvent(message="synthetic failure", code="500")])
diff --git a/tests/test_provider_model_catalog.py b/tests/test_provider_model_catalog.py
index a9facb49b0..ed0f9baf68 100644
--- a/tests/test_provider_model_catalog.py
+++ b/tests/test_provider_model_catalog.py
@@ -63,6 +63,27 @@ def test_deepseek_direct_current_flash_alias_and_pro_limits() -> None:
         assert caps.reasoning_format == "deepseek"
 
 
+def test_openrouter_c5_models_have_offline_budgets_and_capabilities() -> None:
+    catalog = ModelCatalog()
+    expected = {
+        "deepseek/deepseek-v4.1-flash": (1_048_576, 384_000),
+        "z-ai/glm-5.3-flash": (1_310_720, 131_072),
+        "qwen/qwen3.8-flash": (1_000_000, 131_072),
+        "qwen/qwen3.8-max-0902": (1_000_000, 131_072),
+    }
+
+    for model, (context_window, max_tokens) in expected.items():
+        assert catalog.resolve_context_window_with_source(
+            model, provider="openrouter"
+        ) == (context_window, "catalog")
+        assert catalog.resolve_max_tokens(model, provider="openrouter") == max_tokens
+        capabilities = catalog.get_capabilities(model, provider_name="openrouter")
+        assert capabilities.supports_reasoning is True
+        assert capabilities.supports_tools is True
+        assert capabilities.supports_vision is True
+        assert capabilities.reasoning_format == "openrouter"
+
+
 @pytest.mark.parametrize(
     ("model", "window", "output"),
     [
@@ -108,6 +129,21 @@ def test_provider_scoped_corrections_budget_outranks_snapshot_merge() -> None:
     them (see test_provider/test_live_catalog.py)."""
     catalog = ModelCatalog()
 
+    expected_c5_budgets = {
+        "deepseek-flash": (1_000_000, 384_000),
+        "glm-5.3-flash": (1_048_576, 131_072),
+        "qwen3.8-flash": (1_000_000, 131_072),
+        "qwen3.8-max": (1_000_000, 131_072),
+    }
+    for model, (context_window, max_tokens) in expected_c5_budgets.items():
+        assert catalog.resolve_context_window_with_source(
+            model, provider="tokenrhythm"
+        ) == (context_window, "catalog")
+        assert (
+            catalog.resolve_max_tokens(model, provider="tokenrhythm")
+            == max_tokens
+        )
+
     assert catalog.resolve_context_window_with_source(
         "deepseek-v4-flash", provider="tokenrhythm"
     ) == (1_000_000, "catalog")
diff --git a/tests/test_router_tier_contract.py b/tests/test_router_tier_contract.py
index 3ed331508b..0bbc7ece9f 100644
--- a/tests/test_router_tier_contract.py
+++ b/tests/test_router_tier_contract.py
@@ -58,6 +58,15 @@ def test_selection_mode_metadata_has_one_canonical_profile_owner() -> None:
     assert tokenrhythm is STATIC_B5_PROFILES["static_tokenrhythm_b5"]
     assert openrouter is not None and openrouter.provider_id == "openrouter"
     assert tokenrhythm is not None and tokenrhythm.provider_id == "tokenrhythm"
+    assert tokenrhythm.proposer_models == (
+        "deepseek-flash",
+        "glm-5.3-flash",
+        "qwen3.8-flash",
+        "qwen3.8-max",
+    )
+    assert tokenrhythm.aggregator_model == "deepseek-flash"
+    assert openrouter.thinking_level == "high"
+    assert tokenrhythm.thinking_level == "high"
     assert openrouter.ownership_role == "static_profile"
     assert tokenrhythm.api_key_env == "TOKENRHYTHM_API_KEY"