What to do when Codex is down: 10 ways to keep coding in 2026

Disclosure: Enjoy is our product. We checked every way below against its maker’s own pages on November 11, 2026, and we say where another way is the better pick, including the free ones that don’t need Enjoy.

The ways to keep coding, compared

When Codex stops answering, start with the message. In Codex’s own error list, Selected model is at capacity. Please try a different model. is its message for an overloaded model, and We’re currently experiencing high demand, which may cause temporary errors. its message for a server error at OpenAI. Both are about OpenAI’s servers, not your plan. A usage limit says “You’ve hit your usage limit” instead.

Most Codex incidents are short, but the long ones are long. In the 30 days to November 10, 2026, OpenAI’s status page marked one or more of Codex’s parts affected in five incidents. Three lasted an hour or less, by the times the page gives for the Codex parts. The other two took hours, and each was narrow: on October 13, “Degraded performance in GPT-6.1 Sol Codex and Work mode” lasted 8 hours and 18 minutes and named one model, and on October 14, 2 hours and 47 minutes of “Degraded performance in Codex and ChatGPT Work” marked one Codex part, Codex in the ChatGPT desktop app. So the first three ways below cost nothing and fit most incidents. The rest are for the wide ones, and for work that can’t wait.

Way	Cost, checked November 11, 2026	How soon you’re back	Best for
Check the status page, and wait	Free	When OpenAI recovers, most often within the hour	A short incident
Switch models with /model	Included in your ChatGPT plan	Right away, if the other model is healthy	An incident on one model
Another Codex app, such as the CLI	Included in your ChatGPT plan, from Plus	Right away, if that app isn’t affected	An incident on one app
An API key	API prices: GPT-6.1 Sol is $2 in and $10 out per million tokens	Right away, once you have a key	An incident that leaves OpenAI’s API up
OpenAI’s models on Amazon Bedrock, Google Cloud or Azure	Billed by your cloud provider	Right away, once it’s set up	Companies that already have a cloud account
OpenRouter	Each provider’s list price per token, plus 5.5% when you buy credits	Right away, once it’s set up	Keeping OpenAI’s models, through more than one provider
Claude Code, on a Claude plan	Claude Pro $20 a month, or $17 billed yearly; Max $100; Ultra $200	Right away, in a new conversation	A second lab’s agent
A local model, inside Codex	Free, on your own computer	Once it’s downloaded	Work that can’t depend on anyone’s servers
T3 Code	Free	Right away, in the same thread	Switching agents in one thread, free and open source
Enjoy	A monthly subscription, free for 7 days, with your phone and one computer (pricing)	Right away, in the same conversation	Handing a conversation to another lab’s agent

Prices are each company’s US list prices, before tax. Codex’s pricing puts Codex in every ChatGPT plan: Free is $0, Go $8 a month, Plus $20, and Pro “Plans at $100, $200, or $500 USD per month.” Waiting, switching models or apps, a local model and T3 Code cost nothing extra; the others add to what you pay for ChatGPT.

First, read the message

Not every stop is an outage. Codex’s messages sort them, and the fix differs for each:

What Codex shows	What it means	What to do
Reconnecting... 2/5	Codex is retrying a dropped response by itself	Wait; it tries five times by default
stream disconnected before completion	The response stopped partway, which Codex’s source treats as “a transient error” and retries	Let it retry, then check the status page
Selected model is at capacity. Please try a different model.	That model is busy on OpenAI’s side	Run /model
We’re currently experiencing high demand, which may cause temporary errors.	Codex’s message for a server error at OpenAI	Check the status page, wait a few minutes, send again
exceeded retry limit, last status: and a status code	Codex ran out of retries	Check the status page before you send again
You’ve hit your usage limit	Your plan’s usage ran out, which isn’t an outage	See our usage-limit guide
Quota exceeded. Check your plan and billing details.	Your plan or billing, not OpenAI’s service	Check your plan and billing

OpenAI’s status page lists Codex’s parts one by one: Codex Web, the Codex API, the CLI and the VS Code extension under Codex, and Codex in ChatGPT Desktop under ChatGPT. Each incident page lists its Affected components and when they were affected, so it tells you whether the trouble is one model, one app, or everything. The page also warns: “Availability metrics are reported at an aggregate level across all tiers, models and error types. Individual customer availability may vary depending on their subscription tier as well as the specific model and API features in use.”

Check the status page, and wait

Best for: a short incident, which three of the last five were.

Codex is already trying for you. Its configuration reference gives the defaults: “Retry count for HTTP requests to the provider (default: 4)” and “Retry count for SSE streaming interruptions (default: 5).” While it retries a dropped response, Codex shows a count such as Reconnecting... 2/5, as its retry loop puts it, “so the user understands what is happening instead of staring at a seemingly frozen screen.” status.openai.com posts each incident with updates until it’s resolved, and has a Subscribe to updates button.

If you close Codex while you wait, nothing is lost. In the CLI, /resume lets you “Resume a saved chat from your session list,” and “Codex reloads the selected chat’s transcript so you can pick up where you left off,” say its slash command docs.

Price: free.

Watch for: the status page can trail the problem. On October 20, an incident marked the CLI, the Codex API, Codex Web and the VS Code extension, and its only update was written after the fact: “On October 20, from 9:35 to 10:35 a.m. PDT, some users near Illinois experienced authentication errors with the API, ChatGPT, and Codex. We rolled back a configuration change, and service was restored.” And the retry counts are settings for a custom model provider, under model_providers.<id>; for OpenAI’s own, the advanced configuration says “you can’t override built-in provider IDs.”

Switch models

Best for: an incident on one model, like October 13’s.

When a model is the problem, Codex says so, and says what to do: “Selected model is at capacity. Please try a different model.” The slash command docs give the steps: “Type /model and press Enter,” then “Choose a model such as gpt-6-luna or gpt-6.1-sol from the popup,” and “Run /status to verify the change.” The status page’s titles tell you which model to leave. October 13’s named GPT-6.1 Sol, and OpenAI closed it with “We have resolved the issue causing degraded performance with GPT-6.1 Sol in Codex and ChatGPT Work.”

Price: included in your ChatGPT plan.

Watch for: “The picker only shows models available to your account and workspace,” and on Free and Go, Codex’s pricing lists only “GPT-6 Luna at Standard speed in the desktop app, subject to rollout,” so there may be nothing to switch to. And some incidents are wider than a model: on October 21, “Degraded performance in ChatGPT, Codex, and APIs” marked the CLI, Codex Web, the Codex API and the desktop app, along with ChatGPT and the API, though only for 22 minutes.

Try another Codex app

Best for: an incident on one app, like October 14’s.

Codex is several apps, and the status page tracks each one. On October 14, the incident marked Codex in ChatGPT Desktop and ChatGPT Work, and not the CLI, the VS Code extension or Codex Web. On November 5, “ChatGPT Work and Dots Errors” marked Codex Web and no other Codex part, for 31 minutes. When your app is the one listed, another may still work. The local ones share their settings: OpenAI’s Bedrock page notes that “The ChatGPT desktop app, Codex CLI, IDE extension, and SDK read the same local configuration layers.”

Price: included in your ChatGPT plan. Plus includes “Codex on the web, in the CLI, in the IDE extension, and on iOS.”

Watch for: on Free and Go, Codex is in the desktop app only, so there’s no second app to move to. Install the CLI and sign in before you need it. And a sign-in incident can reach several apps at once: October 20’s marked all four parts under the status page’s Codex heading.

Sign in with an API key

Best for: an incident in Codex that leaves OpenAI’s API running.

Codex doesn’t have to run on your ChatGPT plan. Its authentication docs say “You can also sign in to the ChatGPT desktop app, Codex CLI, or IDE extension with an API key,” from the OpenAI dashboard. In the CLI:

printenv OPENAI_API_KEY | codex login --with-api-key

In the desktop app, “On the signed-out screen, select Use API Key.” The status page lists the API’s own parts, such as Responses and Chat Completions, under a separate APIs heading, and the October 13, October 14 and November 5 incidents didn’t mark them.

Price: “OpenAI bills API key usage through your OpenAI Platform account at standard API rates.” OpenAI’s API pricing lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output, GPT-6.1 Terra at $1 and $5, and GPT-6 Luna at $0.10 and $0.50.

Watch for: “When you sign in with an API key, Codex uses standard API pricing instead of included ChatGPT plan credits,” so sign back in with ChatGPT when the incident is over. Codex’s pricing lists “No cloud-based features (GitHub code review, Slack, etc.)” on an API key, and “Model availability follows the API models available to your key.” The API has bad hours of its own: on November 9, “Elevated Errors On API” lasted 42 minutes.

OpenAI’s models on Amazon Bedrock, Google Cloud or Azure

Best for: companies that already have an AWS, Google Cloud or Azure account, and need OpenAI’s models in particular.

Codex can send its requests to another cloud’s copy of OpenAI’s models. For Amazon Bedrock, a built-in provider, OpenAI says “the OpenAI-hosted Responses API isn’t in the request path,” and you “authenticate with a Bedrock API key or AWS IAM credentials.” To start, “Start codex. On the sign-in screen, choose Use Amazon Bedrock,” or in the desktop app, “On Sign in another way, select Use Amazon Bedrock.” Its model list has GPT-6 Astra, GPT-6 Sol and GPT-6 Luna.

Codex also has a built-in Google Cloud provider, set with model_provider = "gcp" in its advanced configuration, and that page has a block for Azure. Microsoft’s Codex guide says “You can run this coding agent entirely on Azure infrastructure,” with models such as gpt-6-astra and gpt-6-sol.

Price: billed by the cloud, not your ChatGPT plan. For Bedrock, OpenAI sends questions about “quotas, billing, regional availability” to “the customer’s AWS administrator or AWS Support.” On Google Cloud, “all usage appears on a single Google Cloud invoice,” says Google’s announcement.

Watch for: this is set-up work for a calm day. On Bedrock, “Hosted ChatGPT Work on the web, Codex cloud, and features that depend on OpenAI-hosted cloud services or cloud-managed discovery aren’t currently available,” and “Fast Mode isn’t available with Amazon Bedrock.” On Google Cloud, Google says OpenAI’s models are “rolling out now in private preview to select enterprise customers,” as our memo on the deal explains. Microsoft’s guide says that on Windows you “install and configure WSL2,” and that “Entra ID support is currently not available for Codex.”

OpenRouter

Best for: keeping OpenAI’s models when OpenAI’s own service is the part that’s failing.

OpenRouter puts one account in front of models from many providers, billed per token. GPT-6.1 Sol’s page lists OpenAI, Amazon Bedrock and Azure as its providers, and its routing docs set allow_fallbacks to true by default: “Whether to allow backup providers when the primary is unavailable.”

Its Codex guide promises “automatic provider failover” and adds a [model_providers.openrouter] block to ~/.codex/config.toml, with base_url = "https://openrouter.ai/api/v1", then sets model_provider = "openrouter". The model “must be a complete OpenRouter slug including the provider prefix,” such as openai/gpt-6.1-sol.

Price: “No separate OpenAI plan is required.” OpenRouter’s pricing says “Inference is billed at the provider’s list price on every plan,” and its fee “is charged when you buy credits, 5.5% on Standard and 8% on Business, never on individual requests.” The guide adds that “Failed requests aren’t billed.”

Watch for: you pay per token here, whatever ChatGPT plan you have. The guide’s examples still name GPT-5.6 models. “model_provider and model_providers only take effect in your user-level ~/.codex/config.toml,” and with a plain env_key, “Codex won’t use the OpenRouter model catalog.” Set it up before you need it.

Claude Code, on a Claude plan

Best for: a second lab’s agent, on a service of its own.

Claude Code is Anthropic’s coding agent, and it runs on Anthropic’s service, not OpenAI’s. Check status.claude.com before you move. Your project comes with you: Anthropic’s docs say “Claude can also read a repository’s AGENTS.md files in place of CLAUDE.md,” so the instructions you wrote for Codex still apply.

Price: Claude’s pricing says “Claude Code is included in all paid plans”: Pro is $20 a month, or $17 billed yearly, Max $100 and Ultra $200. Its Free plan leaves Claude Code out.

Watch for: Anthropic’s docs don’t describe importing Codex’s chats, so start Claude Code with where things stand. Claude has limits too: “Most accounts have a limit that resets on a rolling five-hour session window, and paid plans add weekly limits on top.” It has outages of its own, which our Claude Code guide covers. Install it and sign in before the day you need it. Our Claude Code vs Codex guide compares the two.

A local model, inside Codex

Best for: work that can’t depend on anyone’s servers, on a computer with memory to spare.

You can stay in Codex and change only where the model runs. “Codex can run against a local ‘open source’ provider such as Ollama or LM Studio when you pass --oss,” says its advanced configuration. Ollama’s Codex guide starts it with one command:

ollama launch codex

“Type /model inside Codex to switch between your installed, tool-capable Ollama models without restarting.” For a particular model, run codex --oss -m and its name.

Price: free. Ollama’s pricing starts at $0, which includes “Run models locally.”

Watch for: Ollama says to “Use a context window of at least 64k tokens for Codex,” and to “Pull additional models with ollama pull <model> before launching.” Its cloud models, such as gpt-oss:120b-cloud, aren’t local: its pricing says “Pay per token for every cloud model.” Download a local one before an outage, not during one. Our guide to local models covers which to try and what they need.

T3 Code

Best for: switching agents inside one thread, free and open source.

T3 Code calls itself “The open-source control plane for coding agents”: “Orchestrate Claude Code, Codex, Antigravity, OpenCode, Cursor, and Grok from one surface. Bring your own subscription.” When Codex is down, you switch the thread to another agent, and T3 Code “transfers conversation context when you switch providers.” When you switch back, “Codex receives historical messages directly when its installed version supports that operation.” Its FAQ says it works on Windows, Mac and Linux.

Price: free. “There’s no way to pay us to use T3 Code. You bring your own sub. It is free and open source (MIT).”

Watch for: “A handoff does not copy the outgoing provider’s reasoning, tool-call state, or attachments,” and for a long conversation it’s “a budgeted selection, not an agent-written summary.” Its README says “We are very very early in this project. Expect bugs.” You still need a second agent installed and signed in.

Enjoy

Best for: handing the conversation you’re in to another lab’s agent, and back to Codex when it returns.

Enjoy runs Codex, Claude Code, Grok Build, Antigravity, Muse Code, and local models through OpenCode, on Mac, Windows and Linux, on the plans you already have. When Codex can’t finish a request:

• You see why. The conversation’s status changes to Failed, and Enjoy shows “a one-line reason, the agent’s own error under Details, and Retry.”
• Another agent retries it. Change the agent in the message box, and “a second button appears, starting with Retry with and naming what changed.” The new agent “gets the conversation so far, including replies from other agents.” Changing only the model works the same way, for an incident on one model.
• Or a local model does. Switch to OpenCode with a model on your own computer.
• Codex picks it up later. “Each agent keeps its own place in a conversation. When you switch back to an agent, it continues where it left off.”

Price: Enjoy is one monthly subscription, and the first seven days are free, with a card to start. It includes your phone and any browser and one computer; each other computer is $9.99 a month, and each teammate you invite needs a subscription of their own. See pricing.

Watch for: Enjoy doesn’t bring OpenAI’s service back, and it doesn’t switch agents for you when one fails: the request waits until you choose. Enjoy runs the Codex CLI, which OpenAI includes from Plus. The second agent has to be installed and signed in, on a plan of its own or a free one, and that’s quicker done before an outage. Enjoy isn’t open source, and it’s a monthly subscription on top of your AI plans once the free week ends.

When to stay with Codex and wait

Most of the time. Three of the last five incidents that touched Codex were over within about an hour, and Codex retries in the meantime. The two that weren’t each named one model or one app, so /model and another Codex app are worth trying before anything that costs money. Setting up a cloud account or a gateway in the middle of an outage can take longer than the outage does.

Set up a second way when an hour matters: before a deadline, for a script that runs unattended, or when you already pay for Claude. The cheapest insurance is free: the Codex CLI signed in beside the desktop app, Claude Code signed in on a plan you have, or a local model already downloaded.

A bigger ChatGPT plan isn’t a documented fix. Codex’s pricing lists “Priority request processing” on Enterprise and Edu, but neither it nor the status page says how that fits an incident.

And don’t treat a limit as an outage. “You’ve hit your usage limit” is about your plan, and our usage-limit guide compares the ways through it.

How we compared

We read OpenAI’s status page and its incident history, Codex’s error messages and retry loop on GitHub, its docs on configuration, slash commands, authentication, Amazon Bedrock and pricing, OpenAI’s API pricing, Google Cloud’s announcement, Microsoft’s Codex guide, OpenRouter’s Codex guide, model page, routing docs and pricing, Claude’s pricing and Claude Code’s docs, Ollama’s Codex guide and pricing, T3 Code’s site and docs, and Enjoy’s own help on November 11, 2026, and every one is linked below.

The incident counts are ours. We counted the incidents on OpenAI’s status page posted from October 12 to November 10, 2026, by UTC time, that mark one or more of Codex Web, the Codex API, the CLI, the VS Code extension and Codex in ChatGPT Desktop as affected, and took each one’s length from the time the page gives for those parts. An October 27 incident, “Elevated error rates for most hosted Code Reviews and Security Reviews on private GitHub repositories,” lists no affected part, so it isn’t in the count.

We quote each maker’s own words for its limits and leave out what it hasn’t documented, and we didn’t run these setups ourselves. Status pages and plans change often, so we recheck this guide against its sources at least once a month, and the date at the foot of the page says when we last did.