Skip to main content

Codex Exceeded Retry Limit, Last Status: 429 Too Many Requests

“exceeded retry limit, last status: 429 Too Many Requests” is a 429 Codex couldn't classify, not your usage limit. First stop parallel chats and sub-agents.

LaoZhang AI TeamPublishedUpdated 18 min read
On this page
Codex request passing through authentication, limit, provider, and response-stream checkpoints

Codex ends a failed turn with this line when the server answers HTTP 429 and Codex cannot tell why:

■ exceeded retry limit, last status: 429 Too Many Requests, request id: ...

It means a request was refused with a 429 that carried no usage-limit or billing code Codex recognizes, so this is not the out-of-usage message, and the first thing to do is stop all parallel work and resend one short message.

What to do right now:

  1. Stop parallel work: leave one chat running, pause sub-agents, and resend one short message.
  2. Check the OpenAI status page for a Codex incident.
  3. If Codex talks to a relay or custom base URL, read that relay's log, not an OpenAI dashboard.
  4. Do not raise request_max_retries or stream_max_retries. Neither adds retries for a 429 that arrives without a Retry-After header.
  5. If that single short message still fails, switch model once with /model and resend it. If another model answers, the refusal is tied to the first model; mention that in /feedback.
  6. If one short message in a fresh session still fails with nothing else running and the status page is clear, stop retrying and send /feedback.

The line comes from Codex itself, whether you use the CLI, the IDE extension, the Codex desktop app, or Codex inside the ChatGPT desktop app with a ChatGPT sign-in, so seeing it “in ChatGPT” means the same Codex error. As of November 14, 2026, a read-only text search of the ChatGPT desktop app for macOS (version 26.1106.51428) found the line only inside the Codex CLI the app bundles (0.167.0-alpha.4.1), so in that app it appears only during a Codex task. The app's own interface code has no copy of it and phrases its 429s another way, such as “You’ve sent too many requests…” for verification codes and “Too many requests. Wait a moment, then try again.” for code review. In the current Codex client (rust-v0.169.0, released November 13, 2026), the real limit messages are worded differently. When your ChatGPT plan allowance runs out, Codex says “You’ve hit your usage limit…” and usually gives a reset time. When API credits or a spend cap run out, it says “Quota exceeded. Check your plan and billing details.” The bare retry-limit line means a 429 arrived without either label. Reports since late August 2026 cluster around three situations: too much parallel work on one account (sub-agents, several chats at once), a service-side problem, or a relay sitting between Codex and OpenAI.

401 Unauthorized and stream disconnected before completion fail for different reasons; each has its own section below.

What “exceeded retry limit, last status: 429” means in Codex

When Codex receives a 429, the current client reads the error body before choosing what to print. The mapping lives in the 429 branch of Codex's api_bridge.rs source file, and it produces these messages:

What Codex printedWhat the server's 429 saidFirst move
“You’ve hit your usage limit…”, often ending with “Try again at …”usage_limit_reached: a five-hour or weekly plan window is used upWait for the reset time shown, or use credits if your plan offers them
“Your workspace is out of credits…” or “You hit your spend cap…”The same usage-limit type, flagged as a workspace credit or spend capAsk the workspace owner, or raise the cap if you are the owner
“To use Codex with your ChatGPT plan, upgrade to Plus…”usage_not_included: the plan does not include CodexChange plans; retrying does nothing
“Quota exceeded. Check your plan and billing details.”insufficient_quota, credit_balance_exhausted, or an organization/project spend or usage limit codeFix billing for the account that made the request; retrying does nothing
exceeded retry limit, last status: 429 Too Many RequestsA 429 with none of the labels aboveWork through the checks in the next two sections

Two limits apply to that last row. First, it describes what Codex could parse, not the server's internal reason. A relay or proxy can wrap any upstream failure in its own 429, and Codex then shows the bare line. On October 26, 2026, a user running Codex through a third-party proxy captured the 429 body there in Codex GitHub issue #30471: code turn_rate_limit_exceeded, message “Codex is at capacity, please try again later.” Codex has no mapping for that code, so it showed only the bare line. Second, the bare line does not tell you how much allowance you have left. That user had used 2% of their weekly allowance, and the reporter of Codex GitHub issue #53978 on October 14, 2026 had about 70% left. Check /status to rule your plan limit in or out instead of guessing from this line.

The , request id: … suffix appears when the response carried a tracking ID. Keep it: it is the one value support can look up. The leading ■ is the CLI's error marker, not part of the server's message.

If the line reads exceeded retry limit, last status: 500 Internal Server Error instead, the situation is different. Codex does retry server errors and dropped connections, but in the current source this exact line comes from the retry limit inside Codex's network layer, which is reported as a 500 with no request ID. In a test with Codex CLI 0.162.0-alpha.17.2, real 500 and 503 responses were each retried 30 times and ended with different messages (see the local test below). The 500 is a stand-in status rather than proof that a server sent one, so check the status page and your connection rather than your plan.

Capture the route before changing it

Spend a minute recording what happened before you log out, edit config, or rerun a large task. In the CLI, start with two read-only commands:

bash
codex --version
codex login status

Then note:

  • the exact error line, including any prefix and the request ID;
  • the time and time zone;
  • the surface (CLI, desktop app, or IDE extension) and the model;
  • what else was running on the same account: other chats, sub-agents, other machines, automatic approval review;
  • your remaining limits. In an active CLI session, /status shows the chat ID, context usage, and rate limits; with ChatGPT sign-in, /usage opens usage and billing settings.

The place to look depends on how Codex is signed in:

How Codex is signed inWhere to lookWhat cannot answer the question
ChatGPT account/status or the usage dashboard, the active workspace, behavior in a fresh session, the status pageOpenAI Platform API balance or RPM/TPM limits
OpenAI API keyThe structured API error, the organization and project, billing and limits, response headersYour ChatGPT Plus or Pro meter
Custom provider, relay, or gatewayThe effective base URL, the provider account, the relay's own logs and trace IDAny OpenAI dashboard the relay doesn't use

Five-step Codex recovery flow that captures the active route, maps 401, 429, stream disconnect, and retry-limit evidence to a bounded action, and verifies the original path

This chart predates the October 2026 update: in the current client, credit and spend-limit 429s print “Quota exceeded” rather than the retry-limit line, and Codex does not retry a 429 on its own.

Looking at the wrong console can produce a perfectly accurate “usage remaining” figure for an account that never received the failed request.

If you build on the Codex App Server, the official Codex App Server error reference separates Unauthorized, upstream HTTP failures, ResponseStreamConnectionFailed, ResponseStreamDisconnected, and ResponseTooManyFailedAttempts. The bare retry-limit line arrives as ResponseTooManyFailedAttempts, with the 429 reported in httpStatusCode.

A Codex 429 is actionable only after you know who sent it

With the bare line in hand, four situations cover nearly every report. The quickest test separates the first one from the rest.

Several agents or chats running at once

Most reports since late August 2026 share one shape: a single stream works, several at once fail across every session on the account, and the usage meter still shows plenty left.

  • In Codex GitHub issue #41702 (ChatGPT Pro, open since August 30), one session with three child agents worked, but opening a second session with two more children brought 429s.
  • A related report, from a Pro account routed through a local OAuth gateway, said more than two or three simultaneous streams triggered failures and that sub-agents count as streams.
  • On October 14, a third concurrent desktop chat failed with the weekly meter at 57% (issue #54017), and a CLI user on WSL hit it with three sessions and 69% remaining.

OpenAI has not confirmed a concurrency limit or published a number for one. One reporter says OpenAI support told them sub-agent requests count toward rate limits and that no per-plan concurrency limit is documented.

The test: close other chats, pause sub-agents, and resend one short message. If that succeeds, bring parallel work back one stream at a time and note where it breaks. To cap how many sub-agents one session opens, set this in config.toml:

toml
[agents]
max_concurrent_threads_per_session = 1

The setting counts spawned agents only, excluding the main thread, and only within one session. It does not cover other sessions, machines, or apps on the same account, and it does not guarantee the server will accept the load.

A problem on OpenAI's side

The OpenAI status history lists “Degraded performance in GPT-6.1 Sol Codex and Work mode” on October 13, 2026, and “Degraded performance in Codex and ChatGPT Work” on October 14, both since resolved. Several “Elevated errors” incidents involving Codex appeared on October 6. None of those incident notes mention 429s, but the October 14 wave of 429 reports arrived the same day, so a link is possible and unconfirmed.

The same string has had service-side causes before. In January 2026, an OpenAI maintainer closed Codex GitHub issue #9135 saying the underlying cause was fixed and that later similar behavior would be a different problem. On June 3, 2026, a burst of identical reports appeared and closed within two days.

A newer wave has no explanation from OpenAI yet. Since October 20, 2026, open Codex GitHub issues #55507 and #57084 have collected the same line from Pro and Business accounts, some with 84–99% of their weekly allowance left. Some logs show the server telling Codex to wait minutes before retrying, as in retrying sampling request (1/5 in 385.99s), with several threads scheduled to retry at the same moment. As of November 14, 2026, neither issue has a reply from OpenAI, and #57084 has 88 comments.

Reporters read those minutes-long waits as delays the server asked for, but the #57084 reports say raw Retry-After headers were not captured, so that is not proven. Either way, if a Reconnecting... 1/5 sits for minutes, leave one thread waiting instead of restarting it or opening more. On November 13, OpenAI Support asked a reporter in the OpenAI community thread on recurring Codex 429s whether the error persisted after switching models, which is why one try with /model is worth making.

If the status page shows an open Codex incident, wait for it to clear, then resend one message. Nobody publishes a recovery time for these, so treat any fixed number of minutes as a guess.

Plan usage, credits, or API billing

If what you saw was “You’ve hit your usage limit…”, it is your plan window. As of October 15, 2026, the Codex pricing and usage limits page says Pro plans currently have no five-hour limit but weekly limits may apply, and gives Plus five-hour estimates that vary by model. For reset times and options, see ChatGPT Pro Codex Limit: Check Usage and Reset, ChatGPT Plus Codex Usage Limits and Credits, and Codex Weekly Limit Reset: Find Your Time or Reset Now.

With an API key, Codex is billed at API pricing and is not subject to ChatGPT plan limits. A Platform 429 can mean a temporary rate limit, an exhausted prepaid balance, or a spending or usage limit. For the types and fixes, see OpenAI API Error 429: Diagnose Rate Limits, Quota, and Spend Caps.

A relay, proxy, or account pool in between

If your model_providers entry points base_url at a relay, gateway, or account-pool proxy, the 429 may be that tool's own response. Two examples from 2026: CLIProxyAPI recorded an upstream OpenAI/Cloudflare 520 as its own 429, and a CC Switch user saw repeated 429s while the deeper error was a 400 from a tool_call_id conversion. Read the relay's log for the upstream status before changing anything in Codex. To check what Codex is actually pointed at, see Codex Custom Provider Setup: API Key, Base URL, and Auth.

“Error running remote compact task” and other 429 variants

Some lines wrap the same 429 in a prefix that tells you which request failed:

Line you sawWhat failedWhat to do
Error running remote compact task: exceeded retry limit, last status: 429 Too Many RequestsThe old server-side context compaction. On June 29, 2026, an OpenAI maintainer wrote in Codex GitHub issue #15468 that compaction now runs locally through the same endpoint as normal turns, which “eliminates ‘remote compaction’ errors.”Update Codex. If compaction still fails on a current version, send /feedback and open a new report.
Automatic approval review failed: exceeded retry limit, last status: 429 Too Many RequestsThe automatic approval review request (codex-auto-review), not your main chat. A Windows user reported it on October 12, 2026, with regular usage still available (issue #53324).The action was not executed, so there is nothing to undo. Turning automatic review off isolates the cause, but you then approve actions yourself; that is a safety trade-off, not a fix.
Codex account warmup failed. Reauthenticate the account and try again. (http_status:429)A third-party account-pool proxy (the opencodex family uses this exact text). openai/codex does not print it.Reauthenticate or replace the account inside that proxy, and check its pool status.
upstream request failedThe component that printed it could not reach the next service. Codex's own network proxy logs this text, and many gateways use similar wording.Read the log of whatever printed it; the cause sits one hop further along.

If the line instead reads Selected model is at capacity. Please try a different model., it is not a 429: Codex shows it when the server reports the model as overloaded (server_overloaded in the session log). Codex “Selected Model Is at Capacity”: Not Your Limit, What to Do covers that error.

Why more retries can make the evidence worse

The word “retry” in the message suggests Codex tried several times. For a 429, the current client may not have retried at all.

In the current client, the HTTP retry policy for model providers sets retry_429: false, while server errors and transport failures are retried. The Codex configuration reference lists request_max_retries (default 4) and stream_max_retries (default 5); both are capped at 100 and govern those other failures, plus one 429 case. At the turn level, Codex retries a 429 only when the server sends a Retry-After delay; it then waits that long and retries up to stream_max_retries times. Without one, the turn ends. One reporter found no Retry-After or rate-limit headers anywhere in their CLI logs. Codex GitHub issue #30471, open since June 28, 2026, asks OpenAI to retry 429s or reword the message. As of November 14, 2026, no maintainer has replied.

So any config change that raises retry counts cannot change how a 429 without Retry-After is handled. Retry loops you build around Codex have their own costs: they multiply traffic through gateways that also retry, lengthen a failure that will not change, and bury the first useful error under repeated attempts.

A stop rule that fits the evidence:

  1. Resend once after cutting parallel work to a single stream.
  2. If the status page shows a Codex incident, resend once more after it clears.
  3. Stop when one short message in a fresh session still returns the same bare 429 with nothing else running and the status page clear. At that point another attempt adds nothing; send /feedback with the request ID.
  4. If the message changes to “You’ve hit your usage limit…” or “Quota exceeded…”, the next step is waiting or billing, not retrying.

If you call models from your own code and need retry-versus-switch rules, see LLM API Retry vs Fallback: Five Gates Before Another Model Call.

Matrix mapping final Codex errors to authentication, limits, upstream provider, and response-stream layers with a sanitized support checklist

This chart predates the October 2026 update: in the current client, the credit and spend-limit codes in the 429 row print “Quota exceeded. Check your plan and billing details.”, not the retry-limit line.

Local test: Codex prints “exceeded retry limit” after a single 429

On October 15, 2026, Codex CLI 0.162.0-alpha.17.2 (the build bundled with the ChatGPT desktop app) was pointed at a small test server through a custom model_providers entry, with a throwaway CODEX_HOME. The server answered every POST /v1/responses with one fixed status, a JSON error body, and an x-request-id header, and it logged each request. Each scenario was a single run of codex exec --skip-git-repo-check "say hi" < /dev/null. A source comparison on November 14, 2026 found the 429 code path unchanged in rust-v0.169.0, so these results still describe the current client.

What was not run: the ChatGPT sign-in path, so no real OpenAI response is involved. The runs also left out API-key sign-in against the real API, relays, the desktop app UI, the IDE extension, 429 bodies that Codex recognizes (usage limit or quota), and dropped connections. The test shows how one client build reacts to a given response. It does not show what OpenAI's servers send to a ChatGPT-signed-in Codex.

Test server's answerCodex configRequests receivedTime to the final lineFinal line Codex printed
429, body code rate_limit_exceeded, no Retry-Afterdefaults1about 0.3 sexceeded retry limit, last status: 429 Too Many Requests, request id: req_mock_429_test
Same 429request_max_retries = 10, stream_max_retries = 101about 0.15 sSame line
429 with Retry-After: 3defaults6, exactly 3 s apart; screen showed Reconnecting... 1/5 to 5/5about 15 sSame 429 line
429 with Retry-After: 3stream_max_retries = 23, 3 s apart; Reconnecting... 1/2, 2/2about 6 sSame 429 line
503defaults30about 24 sunexpected status 503 Service Unavailable: server overloaded, url: …/v1/responses, request id: …
500defaults30about 25 sWe’re currently experiencing high demand, which may cause temporary errors.

What the results show:

  • No retry happened before the bare 429 line. Codex sent one request, got a 429 it could not classify, and printed “exceeded retry limit” within a second. Raising both retry settings to 10 still produced one request. The request ID in the line is the server's x-request-id value, which is why it is worth keeping for support.
  • Retry-After is what turns 429 retries on. With the header, Codex waited exactly the number of seconds the server asked for and retried the turn stream_max_retries times (default 5) before printing the same line. In the CLI, a 429 line that appears at once, with no Reconnecting... n/N lines before it, means no retry happened. Those lines alone do not prove a Retry-After, because 5xx and dropped-stream retries print them too.
  • A real 5xx leaves a different trail. Each 503 or 500 run sent 30 requests (5 quickly backed-off attempts in each of 6 rounds) and ended with a message naming the server problem, not exceeded retry limit, last status: 500.

The test cannot tell you whether OpenAI's 429s to a ChatGPT-signed-in Codex carry Retry-After. The reporter in Codex GitHub issue #42685 found no such header in their CLI logs. If that holds, the first row is what those users get, the Codex retry limit settings cannot help, and the stop rule above is the practical path.

To repeat the test in a script, pass the prompt as an argument and close standard input with < /dev/null. Otherwise codex exec prints Reading additional input from stdin... and waits without sending any request.

Treat 401 as an authentication decision, not a retry prompt

A Codex 401 means the service that handled the request rejected its credentials. Waiting does not change that, so fix the credential, then make one short request.

On the OpenAI Platform API, a 401 can mean invalid authentication, an incorrect API key, missing organization membership, or an IP allowlist mismatch; the OpenAI API error codes guide lists these separately. Read the structured error and confirm the project before choosing a fix. The specific “Incorrect API key provided” case has its own walkthrough: Codex 401 Incorrect API Key: Outage, Stale Token, or Your Key?

For ChatGPT sign-in, check that codex login status shows the account you expect. Reauthenticating makes sense when the stored session is invalid or the wrong account is active, but codex logout clears stored credentials, so it is not a harmless first step. The Codex authentication guide also notes that workload identity follows the process environment, not login and logout.

For an API key, confirm the running process uses a key for the same organization, project, endpoint, and network policy you are inspecting. Do not print the key, dump the environment, or paste auth.json into an issue.

A custom gateway can reject its own credential or pass along an upstream 401. Match the gateway's request ID against its logs. If the gateway reached OpenAI and failed there, rotating the key Codex uses to reach the gateway fixes nothing.

“Stream disconnected” needs a controlled comparison

stream disconnected before completion means the response stopped before it finished. If text follows a colon, read it first. Reasons such as “you have no credits remaining” or a concurrency limit come from the server or relay and point to billing or parallel load, not to your network.

Without a reason, the stream could have been cut by the client, a proxy or TLS inspection, a managed network, a gateway, an upstream service, or a network change on your machine. Change one variable at a time:

  1. Start a fresh session with the same account, provider, and model, and send a short non-sensitive request.
  2. Note whether the failure comes before any output, after partial output, or with an explicit HTTP status.
  3. If policy permits, repeat that request once on another trusted network.
  4. Match the failure time and request ID against gateway or provider logs.
  5. If only one client or version fails, record the difference and stop broad retesting.

Success on another network narrows the problem to the path. It does not justify disabling your organization's security controls. Failure on both networks at once points back to the account, provider, gateway, or service.

Do not jump to raising stream_idle_timeout_ms (default 300,000 ms, five minutes). A longer timeout only changes how long the client waits. It cannot fix a wrong base URL, depleted credits, or a connection a gateway closes on purpose.

Build a support packet that can be correlated

Escalate when a minimal request fails in fresh sessions, when your usage display contradicts the error, when the same failure persists after a stated reset, or when only one client and provider combination reproduces it. In the CLI, /feedback opens the feedback dialog and can attach logs. For a public report, open an issue on the openai/codex GitHub repository.

Include:

  • Codex version and surface: CLI, desktop app, or IDE extension;
  • operating system and any managed-network context;
  • sign-in method and provider name;
  • first and most recent failure times, with time zone;
  • the full error line, including any prefix and the request ID;
  • whether the failure came before first output or after partial output;
  • what ran in parallel: sessions, sub-agents, automatic approval review;
  • your remaining usage at the time, from /status or the dashboard;
  • one controlled comparison and its result, such as single stream versus several;
  • the smallest sanitized log excerpt that keeps the first failure.

Leave out API keys, tokens, authorization headers, full auth or config files, environment dumps, proprietary source, and prompt content. Support needs a request ID, a time, and a provider to match, not the state of your whole machine.

Test recovery on the original setup: same account, provider, and client, with one short request that completes. Switching accounts, models, or networks can unblock you, but it does not show that the original path recovered.

Codex 429 questions: how long it lasts and whose fault it is

How long does a Codex 429 error last?

It depends on which message you saw. “You’ve hit your usage limit…” usually states its reset time. The bare exceeded retry limit, last status: 429 Too Many Requests has no published duration: concurrency-shaped failures stop when parallel work drops, and incident-related ones clear when the incident does.

Is the Codex 429 error my fault?

Usually not in the sense of a mistake. The bare line means the server rejected a request without saying it was your plan limit or billing. Parallel load you start, such as several chats or sub-agents, can contribute, which is why cutting to one stream is the first test.

Will switching Codex to an API key stop the 429?

It moves the session to OpenAI Platform rate limits and API pricing, so it can work around a ChatGPT-account problem. Usage is billed at API prices, and the Platform API has its own 429s for rate limits, prepaid balance, and spend caps.

Sources18

External pages this guide links to, in the order they appear. Last updated Nov 14, 2026.

  1. 1.OpenAI status pagestatus.openai.com/history
  2. 2.429 branch of Codex's api_bridge.rs source filegithub.com/openai/codex/blob/rust-v0.169.0/codex-rs/codex-api/src/api_bridge.rs
  3. 3.Codex GitHub issue #30471github.com/openai/codex/issues/30471
  4. 4.Codex GitHub issue #53978github.com/openai/codex/issues/53978
  5. 5.Codex App Server error referencelearn.chatgpt.com/docs/app-server
  6. 6.Codex GitHub issue #41702github.com/openai/codex/issues/41702
  7. 7.issue #54017github.com/openai/codex/issues/54017
  8. 8.Codex GitHub issue #9135github.com/openai/codex/issues/9135
  9. 9.#55507github.com/openai/codex/issues/55507
  10. 10.#57084github.com/openai/codex/issues/57084
  11. 11.OpenAI community thread on recurring Codex 429scommunity.openai.com/t/recurring-codex-http-429-errors-are-disrupting-development/1415467
  12. 12.Codex pricing and usage limits pagelearn.chatgpt.com/docs/pricing
  13. 13.Codex GitHub issue #15468github.com/openai/codex/issues/15468
  14. 14.issue #53324github.com/openai/codex/issues/53324
  15. 15.Codex configuration referencelearn.chatgpt.com/docs/config-file/config-reference
  16. 16.Codex GitHub issue #42685github.com/openai/codex/issues/42685
  17. 17.OpenAI API error codes guidedevelopers.openai.com/api/docs/guides/error-codes
  18. 18.Codex authentication guidelearn.chatgpt.com/docs/auth
More in ChatGPT & OpenAI