diff --git a/.env.example b/.env.example
index 356e109..f596830 100644
--- a/.env.example
+++ b/.env.example
@@ -1,8 +1,11 @@
 # HTTP server port
 PORT=80
 
-# Trigger background token refresh when usage reaches threshold percent
-USAGE_REFRESH_THRESHOLD=85
+# Prepare next ChatGPT account when active usage reaches threshold percent
+CHATGPT_PREPARE_THRESHOLD=85
+
+# Switch active ChatGPT account when usage reaches threshold percent
+CHATGPT_SWITCH_THRESHOLD=95
 
 # Persistent data directory (tokens, screenshots)
 DATA_DIR=/data
diff --git a/README.md b/README.md
index 7c38f5c..6f423f1 100644
--- a/README.md
+++ b/README.md
@@ -1,24 +1,17 @@
 # megapt
 
-HTTP service that returns an active ChatGPT access token.
-
-The service can:
-- restore/refresh a saved token from `/data`
-- auto-register a new ChatGPT account when needed
-- get verification email from a disposable mail provider (`temp-mail.org`)
-- expose token and usage info via HTTP endpoint
-
+Service for issuing ChatGPT OAuth tokens via browser automation with disposable email.
 
 ## Endpoints
 
-- `GET /token` - legacy route (defaults to `chatgpt` provider)
-- `GET /chatgpt/token` - explicit provider route
+- `GET /chatgpt/token`
+- `GET /token` (legacy alias, same as chatgpt)
 
-Example response:
+Response shape:
 
 ```json
 {
-  "token": "<access_token>",
+  "token": "...",
   "limit": {
     "used_percent": 0,
     "remaining_percent": 100,
@@ -37,67 +30,47 @@ Example response:
 }
 ```
 
+## Environment Variables
 
-## Environment variables
+- `PORT` - HTTP server port (default: `8080`)
+- `DATA_DIR` - persistent data directory for tokens/screenshots (default: `./data`)
+- `CHATGPT_PREPARE_THRESHOLD` - usage threshold to prepare `next_account` (default: `85`)
+- `CHATGPT_SWITCH_THRESHOLD` - usage threshold to switch active account to `next_account` (default: `95`)
 
-See `.env.example`.
+Example config is in `.env.example`.
 
-- `PORT` - HTTP port for the service
-- `USAGE_REFRESH_THRESHOLD` - percent threshold to trigger background token rotation
-- `DATA_DIR` - directory for persistent data (`chatgpt_tokens.json`, screenshots, etc.)
+## Token Lifecycle
 
+- **active account** - currently served token.
+- **next account** - pre-created account/token stored for fast switch.
 
-## Local run
+Behavior:
 
-Requirements:
-- Python 3.14+
-- Playwright Chromium dependencies
+1. If active token is valid, service returns it immediately.
+2. If active token is expired, service tries refresh under a single write lock.
+3. If refresh fails or token is missing, service registers a new account (up to 4 attempts).
+4. When usage reaches `CHATGPT_PREPARE_THRESHOLD`, service prepares `next_account`.
+5. When usage reaches `CHATGPT_SWITCH_THRESHOLD`, service switches active account to `next_account`.
 
-Install and run:
+## Startup Behavior
+
+On startup, service:
+
+1. Ensures active token exists and is usable.
+2. Ensures `next_account` is prepared for ChatGPT.
+
+## Data Files
+
+- `DATA_DIR/chatgpt_tokens.json` - token state with `active` and `next_account`.
+- `DATA_DIR/screenshots/` - automation failure screenshots.
+
+## Run Locally
 
 ```bash
-uv sync --frozen --no-dev
-./.venv/bin/python -m playwright install --with-deps chromium
-PYTHONPATH=./src ./.venv/bin/python src/server.py
+PYTHONPATH=./src python src/server.py
 ```
 
-Then request token:
+## Docker Notes
 
-```bash
-curl http://127.0.0.1:8080/chatgpt/token
-```
-
-
-## Docker deployment
-
-Build image:
-
-```bash
-docker build -t megapt:latest .
-```
-
-Run container:
-
-```bash
-docker run -d \
-  --name megapt \
-  --restart unless-stopped \
-  --env-file .env \
-  -v ./data:/data \
-  -p 80:80 \
-  megapt:latest
-```
-
-Check logs:
-
-```bash
-docker logs -f megapt
-```
-
-
-## Notes
-
-- Service performs a startup token check and tries to recover token automatically.
-- Token write path is synchronized (single-writer lock) to avoid parallel re-registration.
-- Browser runs in virtual display (`Xvfb`) inside container.
-- Keep `/data` persistent between restarts.
+- Dockerfile sets `DATA_DIR=/data`.
+- `entrypoint.sh` starts Xvfb and runs `server.py`.
diff --git a/src/providers/chatgpt/provider.py b/src/providers/chatgpt/provider.py
index 6b70b9c..2971c27 100644
--- a/src/providers/chatgpt/provider.py
+++ b/src/providers/chatgpt/provider.py
@@ -1,19 +1,30 @@
 import asyncio
 import logging
-from typing import Callable
+import os
 from typing import Any
+from typing import Callable
 
 from playwright.async_api import BrowserContext
 
-from providers.base import Provider, ProviderTokens
 from email_providers import BaseProvider
 from email_providers import TempMailOrgProvider
-from .tokens import load_tokens, save_tokens, refresh_tokens
+from providers.base import Provider, ProviderTokens
+from .tokens import (
+    clear_next_tokens,
+    load_next_tokens,
+    load_state,
+    load_tokens,
+    promote_next_tokens,
+    refresh_tokens,
+    save_next_tokens,
+    save_tokens,
+)
 from .usage import get_usage_data
 from .registration import register_chatgpt_account
 
 logger = logging.getLogger(__name__)
-MAX_REGISTRATION_ATTEMPTS = 4
+CHATGPT_REGISTRATION_MAX_ATTEMPTS = 4
+CHATGPT_SWITCH_THRESHOLD = int(os.environ.get("CHATGPT_SWITCH_THRESHOLD", "95"))
 
 
 class ChatGPTProvider(Provider):
@@ -27,11 +38,11 @@ class ChatGPTProvider(Provider):
         self._token_write_lock = asyncio.Lock()
 
     async def _register_with_retries(self) -> bool:
-        for attempt in range(1, MAX_REGISTRATION_ATTEMPTS + 1):
+        for attempt in range(1, CHATGPT_REGISTRATION_MAX_ATTEMPTS + 1):
             logger.info(
                 "Registration attempt %s/%s",
                 attempt,
-                MAX_REGISTRATION_ATTEMPTS,
+                CHATGPT_REGISTRATION_MAX_ATTEMPTS,
             )
             success = await self.register_new_account()
             if success:
@@ -39,16 +50,75 @@ class ChatGPTProvider(Provider):
             logger.warning("Registration attempt %s failed", attempt)
         return False
 
+    async def _create_next_account_under_lock(self) -> bool:
+        active_before, next_before = load_state()
+        if next_before:
+            return True
+
+        logger.info("Creating next account")
+        success = await self._register_with_retries()
+        if not success:
+            return False
+
+        generated_active = load_tokens()
+        if not generated_active:
+            return False
+
+        # Registration writes new tokens as active; restore old active and keep
+        # generated account as next.
+        if active_before:
+            save_tokens(active_before)
+        else:
+            clear_next_tokens()
+        save_next_tokens(generated_active)
+        logger.info("Next account is ready")
+        return True
+
     async def force_recreate_token(self) -> str | None:
         async with self._token_write_lock:
             success = await self._register_with_retries()
             if not success:
                 return None
+            clear_next_tokens()
             tokens = load_tokens()
             if not tokens:
                 return None
             return tokens.access_token
 
+    async def ensure_next_account(self) -> bool:
+        next_tokens = load_next_tokens()
+        if next_tokens and not next_tokens.is_expired:
+            return True
+
+        async with self._token_write_lock:
+            next_tokens = load_next_tokens()
+            if next_tokens and not next_tokens.is_expired:
+                return True
+            return await self._create_next_account_under_lock()
+
+    async def maybe_switch_active_account(self, usage_percent: int) -> bool:
+        if usage_percent < CHATGPT_SWITCH_THRESHOLD:
+            return False
+
+        async with self._token_write_lock:
+            next_tokens = load_next_tokens()
+            if not next_tokens or next_tokens.is_expired:
+                logger.info(
+                    "Active usage >= %s%% and next account missing",
+                    CHATGPT_SWITCH_THRESHOLD,
+                )
+                created = await self._create_next_account_under_lock()
+                if not created:
+                    return False
+
+            switched = promote_next_tokens()
+            if switched:
+                logger.info(
+                    "Switched active account (usage >= %s%%)",
+                    CHATGPT_SWITCH_THRESHOLD,
+                )
+            return switched
+
     @property
     def name(self) -> str:
         return "chatgpt"
diff --git a/src/providers/chatgpt/tokens.py b/src/providers/chatgpt/tokens.py
index 4be1661..cc9a4f6 100644
--- a/src/providers/chatgpt/tokens.py
+++ b/src/providers/chatgpt/tokens.py
@@ -1,9 +1,11 @@
 import json
-import time
-import os
-import aiohttp
-from pathlib import Path
 import logging
+import os
+import time
+from pathlib import Path
+from typing import Any
+
+import aiohttp
 
 from providers.base import ProviderTokens
 
@@ -16,33 +18,108 @@ CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
 TOKEN_URL = "https://auth.openai.com/oauth/token"
 
 
-def load_tokens() -> ProviderTokens | None:
-    if not TOKENS_FILE.exists():
+def _tokens_to_dict(tokens: ProviderTokens) -> dict[str, Any]:
+    return {
+        "access_token": tokens.access_token,
+        "refresh_token": tokens.refresh_token,
+        "expires_at": tokens.expires_at,
+    }
+
+
+def _dict_to_tokens(data: dict[str, Any] | None) -> ProviderTokens | None:
+    if not isinstance(data, dict):
         return None
     try:
-        with open(TOKENS_FILE) as f:
-            data = json.load(f)
         return ProviderTokens(
             access_token=data["access_token"],
             refresh_token=data["refresh_token"],
             expires_at=data["expires_at"],
         )
-    except json.JSONDecodeError, KeyError:
+    except KeyError, TypeError:
         return None
 
 
-def save_tokens(tokens: ProviderTokens):
+def _load_raw() -> dict[str, Any] | None:
+    if not TOKENS_FILE.exists():
+        return None
+    try:
+        with open(TOKENS_FILE) as f:
+            data = json.load(f)
+        if isinstance(data, dict):
+            return data
+        return None
+    except json.JSONDecodeError:
+        return None
+
+
+def _save_raw(data: dict[str, Any]) -> None:
     TOKENS_FILE.parent.mkdir(parents=True, exist_ok=True)
     with open(TOKENS_FILE, "w") as f:
-        json.dump(
-            {
-                "access_token": tokens.access_token,
-                "refresh_token": tokens.refresh_token,
-                "expires_at": tokens.expires_at,
-            },
-            f,
-            indent=2,
-        )
+        json.dump(data, f, indent=2)
+
+
+def _normalize_state(data: dict[str, Any] | None) -> dict[str, Any]:
+    if not data:
+        return {"active": None, "next_account": None}
+
+    if "active" in data or "next_account" in data:
+        return {
+            "active": data.get("active"),
+            "next_account": data.get("next_account"),
+        }
+
+    # Backward compatibility with old flat schema
+    return {"active": data, "next_account": None}
+
+
+def load_state() -> tuple[ProviderTokens | None, ProviderTokens | None]:
+    normalized = _normalize_state(_load_raw())
+    active = _dict_to_tokens(normalized.get("active"))
+    next_account = _dict_to_tokens(normalized.get("next_account"))
+    return active, next_account
+
+
+def save_state(
+    active: ProviderTokens | None, next_account: ProviderTokens | None
+) -> None:
+    payload = {
+        "active": _tokens_to_dict(active) if active else None,
+        "next_account": _tokens_to_dict(next_account) if next_account else None,
+    }
+    _save_raw(payload)
+
+
+def load_tokens() -> ProviderTokens | None:
+    active, _ = load_state()
+    return active
+
+
+def load_next_tokens() -> ProviderTokens | None:
+    _, next_account = load_state()
+    return next_account
+
+
+def save_tokens(tokens: ProviderTokens):
+    _, next_account = load_state()
+    save_state(tokens, next_account)
+
+
+def save_next_tokens(tokens: ProviderTokens):
+    active, _ = load_state()
+    save_state(active, tokens)
+
+
+def promote_next_tokens() -> bool:
+    active, next_account = load_state()
+    if not next_account:
+        return False
+    save_state(next_account, None)
+    return True
+
+
+def clear_next_tokens():
+    active, _ = load_state()
+    save_state(active, None)
 
 
 async def refresh_tokens(refresh_token: str) -> ProviderTokens | None:
@@ -55,7 +132,7 @@ async def refresh_tokens(refresh_token: str) -> ProviderTokens | None:
         async with session.post(TOKEN_URL, data=data) as resp:
             if not resp.ok:
                 text = await resp.text()
-                print(f"Token refresh failed: {resp.status} {text}")
+                logger.warning("Token refresh failed: %s %s", resp.status, text)
                 return None
             json_resp = await resp.json()
             expires_in = json_resp["expires_in"]
diff --git a/src/server.py b/src/server.py
index bb89e51..64ac91d 100644
--- a/src/server.py
+++ b/src/server.py
@@ -7,7 +7,7 @@ from aiohttp import web
 from providers.chatgpt import ChatGPTProvider
 
 PORT = int(os.environ.get("PORT", "8080"))
-USAGE_REFRESH_THRESHOLD = int(os.environ.get("USAGE_REFRESH_THRESHOLD", "85"))
+CHATGPT_PREPARE_THRESHOLD = int(os.environ.get("CHATGPT_PREPARE_THRESHOLD", "85"))
 LIMIT_EXHAUSTED_PERCENT = 100
 
 logging.basicConfig(level=logging.INFO)
@@ -37,7 +37,7 @@ def build_limit(usage_percent: int) -> dict[str, int | bool]:
         "used_percent": usage_percent,
         "remaining_percent": remaining,
         "exhausted": usage_percent >= LIMIT_EXHAUSTED_PERCENT,
-        "needs_refresh": usage_percent >= USAGE_REFRESH_THRESHOLD,
+        "needs_refresh": usage_percent >= CHATGPT_PREPARE_THRESHOLD,
     }
 
 
@@ -59,6 +59,9 @@ async def ensure_provider_token_ready(provider_name: str):
         logger.error("[%s] Could not prepare token at startup", provider_name)
         return
 
+    if isinstance(provider, ChatGPTProvider):
+        await provider.ensure_next_account()
+
     usage_info = await provider.get_usage_info(token)
     if "error" not in usage_info:
         logger.info("[%s] Startup token is ready", provider_name)
@@ -157,6 +160,27 @@ async def token_handler(request: web.Request) -> web.Response:
     usage_percent = usage_info.get("used_percent", 0)
     remaining_percent = usage_info.get("remaining_percent", max(0, 100 - usage_percent))
 
+    if isinstance(provider, ChatGPTProvider):
+        switched = await provider.maybe_switch_active_account(usage_percent)
+        if switched:
+            token = await provider.get_token()
+            if not token:
+                return web.json_response(
+                    {"error": "Failed to get active token after account switch"},
+                    status=503,
+                )
+            usage_info = await provider.get_usage_info(token)
+            if "error" in usage_info:
+                return web.json_response(
+                    {"error": usage_info["error"]},
+                    status=503,
+                )
+            usage_percent = usage_info.get("used_percent", 0)
+            remaining_percent = usage_info.get(
+                "remaining_percent", max(0, 100 - usage_percent)
+            )
+            logger.info("[%s] Active account switched before response", provider_name)
+
     logger.info(
         "[%s] token issued, used=%s%% remaining=%s%%",
         provider_name,
@@ -184,11 +208,14 @@ async def token_handler(request: web.Request) -> web.Response:
         )
 
     # Trigger background refresh if needed
-    if usage_percent >= USAGE_REFRESH_THRESHOLD:
-        trigger_background_refresh(
-            provider_name,
-            f"usage {usage_percent}% >= threshold {USAGE_REFRESH_THRESHOLD}%",
-        )
+    if usage_percent >= CHATGPT_PREPARE_THRESHOLD:
+        if isinstance(provider, ChatGPTProvider):
+            await provider.ensure_next_account()
+        else:
+            trigger_background_refresh(
+                provider_name,
+                f"usage {usage_percent}% >= threshold {CHATGPT_PREPARE_THRESHOLD}%",
+            )
 
     return web.json_response(
         {
@@ -214,7 +241,10 @@ def create_app() -> web.Application:
 
 if __name__ == "__main__":
     logger.info("Starting token service on port %s", PORT)
-    logger.info("Usage refresh threshold: %s%%", USAGE_REFRESH_THRESHOLD)
+    logger.info(
+        "ChatGPT prepare-next threshold: %s%%",
+        CHATGPT_PREPARE_THRESHOLD,
+    )
     logger.info("Available providers: %s", ", ".join(PROVIDERS.keys()))
     app = create_app()
     web.run_app(app, host="0.0.0.0", port=PORT)