diff --git a/server/claude-sdk.js b/server/claude-sdk.js
index e92e4f0..1d25b69 100644
--- a/server/claude-sdk.js
+++ b/server/claude-sdk.js
@@ -17,7 +17,7 @@ import crypto from 'crypto';
 import { promises as fs } from 'fs';
 import path from 'path';
 import os from 'os';
-import { CLAUDE_MODELS } from '../shared/modelConstants.js';
+import { CLAUDE_MODELS, getModelContextWindow } from '../shared/modelConstants.js';
 import { resolveClaudeCodeExecutablePath } from './shared/claude-cli-path.js';
 import {
   createNotificationEvent,
@@ -205,7 +205,14 @@ function mapCliOptionsToSDK(options = {}) {
   // Map model (default to sonnet)
   // Valid models: sonnet, opus, haiku, opusplan, sonnet[1m]
   sdkOptions.model = options.model || CLAUDE_MODELS.DEFAULT;
-  // Model logged at query start below
+
+  // Stamp per-model CONTEXT_WINDOW into the subprocess env so Claude Code
+  // self-limits its prompt budget to the model's actual capacity.
+  // Local/constrained models (e.g. claudecode, cc/qwen72b) declare a lower
+  // ceiling in MODEL_CONTEXT_OVERRIDES; cloud models use the global default.
+  const modelContextWindow = getModelContextWindow(sdkOptions.model);
+  sdkOptions.env.CONTEXT_WINDOW = String(modelContextWindow);
+  sdkOptions.env.VITE_CONTEXT_WINDOW = String(modelContextWindow);
 
   // Map system prompt configuration
   sdkOptions.systemPrompt = {
@@ -287,9 +294,10 @@ function transformMessage(sdkMessage) {
 /**
  * Extracts token usage from SDK result messages
  * @param {Object} resultMessage - SDK result message
+ * @param {string} modelValue - The model value used for this session
  * @returns {Object|null} Token budget object or null
  */
-function extractTokenBudget(resultMessage) {
+function extractTokenBudget(resultMessage, modelValue) {
   if (resultMessage.type !== 'result' || !resultMessage.modelUsage) {
     return null;
   }
@@ -312,11 +320,8 @@ function extractTokenBudget(resultMessage) {
   // Total used = input + output + cache tokens
   const totalUsed = inputTokens + outputTokens + cacheReadTokens + cacheCreationTokens;
 
-  // Use configured context window budget from environment (default 160000)
-  // This is the user's budget limit, not the model's context window
-  const contextWindow = parseInt(process.env.CONTEXT_WINDOW) || 160000;
-
-  // Token calc logged via token-budget WS event
+  // Use per-model context window so the budget meter reflects the model's actual ceiling.
+  const contextWindow = getModelContextWindow(modelValue);
 
   return {
     used: totalUsed,
@@ -682,7 +687,7 @@ async function queryClaudeSDK(command, options = {}, ws) {
         if (models.length > 0) {
           // Model info available in result message
         }
-        const tokenBudgetData = extractTokenBudget(message);
+        const tokenBudgetData = extractTokenBudget(message, sdkOptions.model);
         if (tokenBudgetData) {
           ws.send(createNormalizedMessage({ kind: 'status', text: 'token_budget', tokenBudget: tokenBudgetData, sessionId: capturedSessionId || sessionId || null, provider: 'claude' }));
         }