diff --git a/src/llm.ts b/src/llm.ts
index 84408f0..6290dc2 100644
--- a/src/llm.ts
+++ b/src/llm.ts
@@ -6,13 +6,15 @@ import { v4 as uuidv4 } from 'uuid';
 // Unified LLM client — OpenAI-compatible message format throughout
 //
 // Two backends:
-//   VertexOpenAIClient  — for GLM-5, Claude Sonnet, etc. via Vertex global endpoint
-//   GeminiFlashClient   — for Gemini Flash/Pro via @google/genai SDK
+//   VertexOpenAIClient  — for GLM-5 and other Vertex MaaS models (openai-compat endpoint)
+//   GeminiClient        — for Gemini Flash/Pro via @google/genai SDK (API key)
 //
-// Model tier defaults (overridable via env):
-//   Tier A: gemini-2.5-flash  ($0.15/$0.60 per 1M)  — routing, summaries, log parsing
-//   Tier B: zai-org/glm-5-maas ($1.00/$3.20 per 1M) — coding, feature work
-//   Tier C: zai-org/glm-5-maas ($1.00/$3.20 per 1M) — complex decisions, escalation
+// Model tier defaults (overridable via TIER_A/B/C_MODEL env vars):
+//   Tier A: gemini-2.5-flash    — routing, summaries, log parsing (API key, high quota)
+//   Tier B: zai-org/glm-5-maas  — coding, feature work (Vertex MaaS, retry on 429)
+//   Tier C: zai-org/glm-5-maas  — complex decisions, escalation
+//
+// Claude models are NOT available in this GCP project — do not use anthropic/* IDs.
 // =============================================================================
 
 // ---------------------------------------------------------------------------
@@ -294,8 +296,8 @@ export type ModelTier = 'A' | 'B' | 'C';
 
 const TIER_MODELS: Record<ModelTier, string> = {
     A: process.env.TIER_A_MODEL ?? 'gemini-2.5-flash',
-    B: process.env.TIER_B_MODEL ?? 'claude-sonnet-4-6',
-    C: process.env.TIER_C_MODEL ?? 'claude-sonnet-4-6'
+    B: process.env.TIER_B_MODEL ?? 'zai-org/glm-5-maas',
+    C: process.env.TIER_C_MODEL ?? 'zai-org/glm-5-maas'
 };
 
 export function createLLM(modelOrTier: string | ModelTier, opts?: { temperature?: number }): LLMClient {