limboinf · limboinf · Sep 16, 2026 · Sep 16, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -13,6 +13,13 @@ for reviewers.
 
 Add user-visible changes here before running `bun run release:prepare <version>`.
 
+### Removed
+
+- **The "Extended Context (1M)" toggle is gone.** It opted Anthropic requests
+  into the 1M-token context beta, which bills over-200K requests at higher
+  rates and fails outright on lower-tier API keys. Models now always run with
+  the standard 200K context window.
+
 ## [0.13.0] - 2026-09-13
 
 ### Added

diff --git a/apps/electron/src/renderer/pages/settings/AiSettingsPage.tsx b/apps/electron/src/renderer/pages/settings/AiSettingsPage.tsx
@@ -639,7 +639,6 @@ export default function AiSettingsPage() {
   // Default settings state (app-level)
   const [defaultThinking, setDefaultThinking] = useState<ThinkingLevel>(DEFAULT_THINKING_LEVEL)
   const [extendedPromptCache, setExtendedPromptCache] = useState(false)
-  const [enable1MContext, setEnable1MContext] = useState(false)
   const [rtkEnabled, setRtkEnabled] = useState(false)
   const [rtkStatus, setRtkStatus] = useState<{ installed: boolean; path: string | null; version: string | null } | null>(null)
   const [rtkRechecking, setRtkRechecking] = useState(false)
@@ -673,9 +672,6 @@ export default function AiSettingsPage() {
         const extendedCache = await window.electronAPI.getExtendedPromptCache()
         setExtendedPromptCache(extendedCache)
 
-        const enable1M = await window.electronAPI.getEnable1MContext()
-        setEnable1MContext(enable1M)
-
         const rtkOn = await window.electronAPI.getRtkEnabled()
         setRtkEnabled(rtkOn)
 
@@ -979,11 +975,6 @@ export default function AiSettingsPage() {
     await window.electronAPI?.setExtendedPromptCache(enabled)
   }, [])
 
-  const handleEnable1MContextChange = useCallback(async (enabled: boolean) => {
-    setEnable1MContext(enabled)
-    await window.electronAPI?.setEnable1MContext(enabled)
-  }, [])
-
   const handleRtkToggle = useCallback(async (enabled: boolean) => {
     setRtkEnabled(enabled)
     await window.electronAPI?.setRtkEnabled(enabled)
@@ -1125,12 +1116,6 @@ export default function AiSettingsPage() {
               {/* Performance */}
               <SettingsSection title={t("settings.ai.performance")} description={t("settings.ai.performanceDesc")}>
                 <SettingsCard>
-                  <SettingsToggle
-                    label={t("settings.ai.extendedContext")}
-                    description={t("settings.ai.extendedContextDesc")}
-                    checked={enable1MContext}
-                    onCheckedChange={handleEnable1MContextChange}
-                  />
                   <SettingsToggle
                     label={t("settings.ai.extendedPromptCache")}
                     description={t("settings.ai.extendedPromptCacheDesc")}

diff --git a/apps/electron/src/shared/types.ts b/apps/electron/src/shared/types.ts
@@ -544,11 +544,9 @@ export interface ElectronAPI {
   getRichToolDescriptions(): Promise<boolean>
   setRichToolDescriptions(enabled: boolean): Promise<void>
 
-  // Prompt caching & context
+  // Prompt caching
   getExtendedPromptCache(): Promise<boolean>
   setExtendedPromptCache(enabled: boolean): Promise<void>
-  getEnable1MContext(): Promise<boolean>
-  setEnable1MContext(enabled: boolean): Promise<void>
 
   // RTK token optimization
   getRtkEnabled(): Promise<boolean>

diff --git a/apps/electron/src/transport/channel-map.ts b/apps/electron/src/transport/channel-map.ts
@@ -203,8 +203,6 @@ export const CHANNEL_MAP = {
   setRichToolDescriptions: invoke(RPC_CHANNELS.appearance.SET_RICH_TOOL_DESCRIPTIONS),
   getExtendedPromptCache: invoke(RPC_CHANNELS.caching.GET_EXTENDED_PROMPT_CACHE),
   setExtendedPromptCache: invoke(RPC_CHANNELS.caching.SET_EXTENDED_PROMPT_CACHE),
-  getEnable1MContext: invoke(RPC_CHANNELS.caching.GET_ENABLE_1M_CONTEXT),
-  setEnable1MContext: invoke(RPC_CHANNELS.caching.SET_ENABLE_1M_CONTEXT),
   getRtkEnabled: invoke(RPC_CHANNELS.rtk.GET_ENABLED),
   setRtkEnabled: invoke(RPC_CHANNELS.rtk.SET_ENABLED),
   getRtkStatus: invoke(RPC_CHANNELS.rtk.GET_STATUS),

diff --git a/packages/server-core/src/handlers/rpc/settings.ts b/packages/server-core/src/handlers/rpc/settings.ts
@@ -31,8 +31,6 @@ export const HANDLED_CHANNELS = [
   RPC_CHANNELS.appearance.SET_RICH_TOOL_DESCRIPTIONS,
   RPC_CHANNELS.caching.GET_EXTENDED_PROMPT_CACHE,
   RPC_CHANNELS.caching.SET_EXTENDED_PROMPT_CACHE,
-  RPC_CHANNELS.caching.GET_ENABLE_1M_CONTEXT,
-  RPC_CHANNELS.caching.SET_ENABLE_1M_CONTEXT,
   RPC_CHANNELS.sessions.GET_MODEL,
   RPC_CHANNELS.sessions.SET_MODEL,
   RPC_CHANNELS.settings.GET_DEFAULT_THINKING_LEVEL,
@@ -296,18 +294,6 @@ export function registerSettingsHandlers(server: RpcServer, deps: HandlerDeps):
     setExtendedPromptCache(enabled)
   })
 
-  // Get 1M context window setting
-  server.handle(RPC_CHANNELS.caching.GET_ENABLE_1M_CONTEXT, async () => {
-    const { getEnable1MContext } = await import('@bitlab/shared/config/storage')
-    return getEnable1MContext()
-  })
-
-  // Set 1M context window setting
-  server.handle(RPC_CHANNELS.caching.SET_ENABLE_1M_CONTEXT, async (_ctx, enabled: boolean) => {
-    const { setEnable1MContext } = await import('@bitlab/shared/config/storage')
-    setEnable1MContext(enabled)
-  })
-
   // ============================================================
   // RTK Token-Optimization Settings
   // ============================================================

diff --git a/packages/shared/src/agent/backend/types.ts b/packages/shared/src/agent/backend/types.ts
@@ -203,9 +203,6 @@ export interface CoreBackendConfig {
    */
   onImageResize?: (filePath: string, maxSizeBytes: number) => Promise<string | null>;
 
-  /** Enable 1M context window for current Opus models. Default: true. Set false to use 200K and conserve usage limits. */
-  enable1MContext?: boolean;
-
 }
 
 // ============================================================

diff --git a/packages/shared/src/config/storage.ts b/packages/shared/src/config/storage.ts
@@ -93,9 +93,8 @@ export interface StoredConfig {
   /** Whether skills may pre-approve tools for their turn. Absent means enabled. */
   skillToolApproval?: boolean;
   mcpSettings?: BitlabMcpSettings;
-  // Prompt caching & context
+  // Prompt caching
   extendedPromptCache?: boolean;  // Use 1h prompt cache TTL instead of 5m (default: false)
-  enable1MContext?: boolean;  // Enable 1M context window for supported models (default: false — opt-in; requires Anthropic Tier 4+)
   // Token optimization
   rtkEnabled?: boolean;  // Route Bash commands through rtk to compress tool output (default: false). https://github.com/rtk-ai/rtk
   // Network proxy
@@ -588,28 +587,6 @@ export function setMcpSettings(settings: BitlabMcpSettings): void {
   saveConfig(config);
 }
 
-/**
- * Get whether 1M context window is enabled.
- * When disabled, models use 200K context and the interceptor strips the context-1m beta header.
- * Defaults to false — the 1M beta requires Anthropic Tier 4+, and enabling it by default
- * causes 400 "Invalid Request" for lower-tier API keys on large contexts (issue #567).
- * Users opt in via AI Settings → Performance → Extended Context (1M).
- */
-export function getEnable1MContext(): boolean {
-  const config = loadStoredConfig();
-  return config?.enable1MContext === true;
-}
-
-/**
- * Set whether 1M context window is enabled.
- */
-export function setEnable1MContext(enabled: boolean): void {
-  const config = loadStoredConfig();
-  if (!config) return;
-  config.enable1MContext = enabled;
-  saveConfig(config);
-}
-
 /**
  * Get whether rtk Bash-output compression is enabled.
  * When enabled, the PreToolUse pipeline rewrites Bash commands to their `rtk` equivalents

diff --git a/packages/shared/src/i18n/locales/en.json b/packages/shared/src/i18n/locales/en.json
@@ -584,8 +584,6 @@
   "settings.ai.defaultSectionDesc": "Settings for new chats when no workspace override is set.",
   "settings.ai.description": "Model, thinking, connections",
   "settings.ai.enterConnectionName": "Enter connection name...",
-  "settings.ai.extendedContext": "Extended Context (1M)",
-  "settings.ai.extendedContextDesc": "Use 1M token context window for current Opus models. Disable to use 200K and conserve usage limits.",
   "settings.ai.extendedPromptCache": "Extended prompt cache (1 hour)",
   "settings.ai.extendedPromptCacheDesc": "Cache prompts for 1 hour instead of 5 minutes. Only applies to Claude models via Anthropic API. Reduces cost for long sessions but increases cache write cost.",
   "settings.ai.inheritFromApp": "Inherit from app settings",

diff --git a/packages/shared/src/i18n/locales/zh-Hans.json b/packages/shared/src/i18n/locales/zh-Hans.json
@@ -584,8 +584,6 @@
   "settings.ai.defaultSectionDesc": "未设置 Workspace 覆盖时新聊天的设置。",
   "settings.ai.description": "模型、思考、连接",
   "settings.ai.enterConnectionName": "输入连接名称...",
-  "settings.ai.extendedContext": "扩展上下文 (1M)",
-  "settings.ai.extendedContextDesc": "为 Opus 4.7 使用 1M token 上下文窗口。禁用则使用 200K 以节省用量配额。",
   "settings.ai.extendedPromptCache": "扩展提示缓存 (1 小时)",
   "settings.ai.extendedPromptCacheDesc": "将提示缓存从 5 分钟延长到 1 小时。仅适用于通过 Anthropic API 的 Claude 模型。可降低长会话成本，但会增加缓存写入成本。",
   "settings.ai.inheritFromApp": "继承应用设置",

diff --git a/packages/shared/src/interceptor-common.ts b/packages/shared/src/interceptor-common.ts
@@ -138,18 +138,6 @@ export function isExtendedPromptCacheEnabled(): boolean {
   return config?.extendedPromptCache === true;
 }
 
-/**
- * Check if 1M context window is enabled.
- * When disabled, the interceptor strips the context-1m beta header.
- * Defaults to false — the 1M beta requires Anthropic Tier 4+, so it's opt-in
- * to avoid 400 "Invalid Request" on lower-tier API keys (issue #567).
- * Must stay in sync with getEnable1MContext() in config/storage.ts.
- */
-export function is1MContextEnabled(): boolean {
-  const config = getInterceptorConfig();
-  return config?.enable1MContext === true;
-}
-
 // ============================================================================
 // LAST API ERROR
 // ============================================================================

diff --git a/packages/shared/src/protocol/channels.ts b/packages/shared/src/protocol/channels.ts
@@ -284,8 +284,6 @@ export const RPC_CHANNELS = {
   caching: {
     GET_EXTENDED_PROMPT_CACHE: 'caching:getExtendedPromptCache',
     SET_EXTENDED_PROMPT_CACHE: 'caching:setExtendedPromptCache',
-    GET_ENABLE_1M_CONTEXT: 'caching:getEnable1MContext',
-    SET_ENABLE_1M_CONTEXT: 'caching:setEnable1MContext',
   },
   rtk: {
     GET_ENABLED: 'rtk:getEnabled',

diff --git a/packages/shared/src/unified-network-interceptor.ts b/packages/shared/src/unified-network-interceptor.ts
@@ -24,7 +24,6 @@ import {
   debugLog,
   isRichToolDescriptionsEnabled,
   isExtendedPromptCacheEnabled,
-  is1MContextEnabled,
   setStoredError,
   toolMetadataStore,
   displayNameSchema,
@@ -790,16 +789,14 @@ const anthropicAdapter: ApiAdapter = {
     sanitizeEmptyTextCacheControl(body);
     upgradePromptCacheTtl(body);
 
-    // Strip SDK-injected 1M context beta when setting disables it.
-    // The SDK adds this header automatically for Opus/Sonnet 4.6 models,
-    // but the user may want 200K context to conserve usage limits.
-    if (!is1MContextEnabled()) {
-      debugLog('[Anthropic] Stripping context-1m beta header (enable1MContext=false)');
-      init = {
-        ...init,
-        headers: stripBetaHeader(init?.headers as HeadersInitType | undefined, 'context-1m-2025-08-07'),
-      };
-    }
+    // Always strip the SDK-injected 1M context beta. The SDK adds this header
+    // automatically for Opus/Sonnet 4.6 models, but it 400s on lower-tier
+    // Anthropic API keys (issue #567) and burns usage limits faster, so all
+    // requests stay on the 200K context window.
+    init = {
+      ...init,
+      headers: stripBetaHeader(init?.headers as HeadersInitType | undefined, 'context-1m-2025-08-07'),
+    };
 
     const fastMode = shouldEnableFastMode(body.model);
     if (fastMode) {