Possible Codex usage metering anomaly on Plus (very small tasks consuming large 5h + weekly quota)

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #13167
• #12759
• #12728
• #11508
• #11785

Powered by Codex Action

As I mentioned on other threads, we have not changed anything on the server side related to usage accounting or metering. If you're noticing a change in usage, it's likely in how you're using Codex.

If you suspect this might be a client-side regression, you could try downgrading to an earlier version of the CLI to see if you observe the same behavior.

As I mentioned on other threads, we have not changed anything on the server side related to usage accounting or metering. If you're noticing a change in usage, it's likely in how you're using Codex. If you suspect this might be a client-side regression, you could try downgrading to an earlier version of the CLI to see if you observe the same behavior.

Is there no chance a bug has developed that has impacted this in some way unintentionally, as opposed to a direct change?

You can surely agree that the above example (I can screenshot the entire session) should not use 2% on gpt-5.3-codex-medium, especially on 2x?

Usually I'm messaging on high/xhigh with tools, recursive tasks etc. and I'm surprised to see it go below 90% after extensive use. Today I've taken it to below 35% in a 5 hour window of pretty casual usage and with lower reasoning.

I'll see if I can downgrade, but I think there's been enough people suddenly comment on the issue in the last 24-48 hours that _something_ might be playing up somewhere - even if unintentionally. I just don't want to assume "it's me" and it lead to a bigger issue as right now I feel loathe to use it because it decimates my usage allowance. Prior to today, even on Plus it felt like I had near unlimited usage based on the stats.

I'm also noticing a strange synchronisation issue with my weekly % which I haven't previously noticed for some time. In the attached screenshot, for example Codex CLI initially said 56% which is aligned with the website, then less than a minute later while processing my prompt it suddenly changes to 53% weekly in Codex CLI (screenshot), while the website still says 56% weekly, and then Codex CLI moments later also goes back to 56% weekly. I'm not sure if this in some way related to some issues others are reporting about a sudden change in consumption of the quota, but thought I'd share this in case it's useful or important in some way.

I'm on Pro plan.

<img width="691" height="637" alt="Image" src="https://github.com/user-attachments/assets/4102baa7-6b39-4cb4-a373-41e51325c7fe" />

As I mentioned on other threads, we have not changed anything on the server side related to usage accounting or metering. If you're noticing a change in usage, it's likely in how you're using Codex. If you suspect this might be a client-side regression, you could try downgrading to an earlier version of the CLI to see if you observe the same behavior.

Im really tired of your useless replies in every similar issue when this is full of users reporting the same thing here and on Reddit... so STOP with your 0 help and say to your supreme leader Sam Altman to stop laugh at his customers because we are paying big amounts of money and restore back our normal usage limit and fix your apps... this is unaceptable and reading people like you is annoying.

正如我在其他帖子中提到的，我們沒有對伺服器端與使用量統計或計量相關的任何設定進行更改。如果您發現使用量發生變化，很可能是您使用 Codex 的方式改變了。 如果您懷疑這可能是客戶端回歸問題，您可以嘗試降級到早期版本的 CLI，看看是否會出現相同的行為。

You only need to fix this bug where there's remaining credit but it can't be used; you don't need to worry about whether our usage method has changed.

Going to add, Using Win 10/vscode extension. Usage of 5h/weekly was insanely high compared to previous week. There's clearly something wrong.

I also ran into this problem. On February 24, everything was fine. And the next time I used Codex on February 28, the limits started to be used up very quickly. Today I also tried to install an older version of the Codex CLI, but the problem still remains. Apparently, the problem is on the server side, the principle of counting tokens/limits has changed.

Same here.

Definitely something going on here - what used to last 2/3 days now lasts 2/3 hours.

same here

seems likely? https://www.reddit.com/r/codex/comments/1rhw43o/my_codex_tokens_suddenly_have_been_exhausted_over/
i consumed 30 credits on an rg and two reads totalling 800 loc and a git diff earlier.

### Token usage regression between codex-cli 0.102.0 and 0.106.0 (same model/settings)

I tested both versions with the same setup:

• Model: gpt-5.3-codex
• Reasoning effort: xhigh
• Same prompt/workflow style

### What I found

1. Per-request overhead increased in 0.106

• Request payload is about +6.6% vs 0.102 (same model + same reasoning setting).

1. Total weekly burn can increase much more than 6.6%

• In practice, 0.106 also increases request fan-out (more child-agent related calls in my workload).
• Effective total increase looked closer to ~+21% to +32% in real usage.

### Why this likely happens

• 0.106 exposes larger tool schemas in prompts.
• With multi_agent=true and sqlite enabled/default, extra agent-job tooling (e.g. spawn_agents_on_csv) can be

included.

• That increases prompt/tool overhead and can increase multi-agent call frequency.

### Easy workaround (recommended)

Run 0.106 but disable costly features by default:

```bash
codex --disable multi_agent --disable sqlite

  Or set permanently in ~/.codex/config.toml:

  [features]
  multi_agent = false
  sqlite = false

  Enable them only when needed:

  codex --enable multi_agent --enable sqlite

  ### Is this a bug?

  Maybe not a functional bug, but it does look like a cost/performance regression for users who care about weekly token
  limits.

  ### Optimization opportunities (for maintainers)

  - Avoid injecting heavy tool schemas unless needed.
  - Lazy-register multi-agent/batch-agent tools.
  - Decouple sqlite default behavior from tool exposure where possible.
  - Provide a “low-token mode” preset for budget-sensitive users.

@etraut-openai 

yeah the codex vscode extension on windows is giving very odd usage limits - i think this is a bug

Disable sqlite is a bad idea @yf-river , some new features require it

It is true that we have more tools that consume a bit more usage. Will reflect on this to try to find a solution

Disable is a bad idea @yf-river , some new features require itsqlite It is true that we have more tools that consume a bit more usage. Will reflect on this to try to find a solution

It's not that, even when using it in opencode, my usage is consumed faster; it's not about the harness

I'm also experiencing this. Usually I am able to go several days on my plus plan with less than 10% of my weekly limit. Today I have used >50% of my weekly limit in a few hours.

Needless to say there has been zero change in my actual usage.

I too experienced a very significant increase (several times higher than normal) in token consumption for the last 3-5 days, unsure exactly when it started.
Usually my Pro account has more than plenty of quota to use Codex all week long without ever getting even close to the limit. But in just the last 2-3 days it burned straight through all my credits in no time with exactly the same kind of prompts I usually do in the same solution I always work with.

I tried purchasing additional credits for 40 USD today.
It burned straight through all of that purchased credit in just a couple of hours - something is definitely not right.

(pro user and xhigh)

Getting really odd usage limit counts here (ChatGPT Plus), too. Tested after fresh reset with a quick hello world that reads readme.md (1k words, 8KB). 5h burn is at 1% per 3k tokens. Not sure what was the baseline usage burn before this but many times lower.

Two tests - first:

› hello world / ping test - just read a root readme.md and report back if file access works


• I’ll quickly read the repository root readme.md to verify file access, then report exactly what I can see.

• Ran Get-Content -Path readme.md -TotalCount 40
  └ [REDACTED]

  So your ping test passed.

/status

╭─────────────────────────────────────────────────────────────────────────────────╮
│  >_ OpenAI Codex (v0.106.0)                                                     │
│                                                                                 │
│ Visit https://chatgpt.com/codex/settings/usage for up-to-date                   │
│ information on rate limits and credits                                          │
│                                                                                 │
│  Model:                gpt-5.3-codex (reasoning medium, summaries auto)         │
│  Directory:            [REDACTED]                                               │
│  Permissions:          Custom (workspace-write with network access, on-request) │
│  Agents.md:            AGENTS.md                                                │
│  Account:              [REDACTED]                                               │
│  Collaboration mode:   Default                                                  │
│  Session:              019cb1b7-850a-73b2-b19a-5278ab7c3ec2                     │
│                                                                                 │
│  Context window:       99% left (13.5K used / 258K)                             │
│  5h limit:             [███████████████████░] 97% left (resets 09:15)           │
│  Weekly limit:         [████████████████████] 99% left (resets 04:15 on 10 Mar) │
╰─────────────────────────────────────────────────────────────────────────────────╯
Token usage: total=9,150 input=9,131 (+ 4,864 cached) output=19
To continue this session, run codex resume 019cb1b1-862c-7d30-a3ba-6f1ee8174b0d

Second run right after the one above:

› hello world / ping test - just read a root readme.md and report back if file access works


• I’ll quickly check the repo root readme.md and confirm file access by reading a few lines from it.

• Ran Get-Content -Path .\readme.md -TotalCount 20
  └ [REDACTED]
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

• File access works. I successfully read root readme.md, and it starts with:

[REDACTED]

/status

╭─────────────────────────────────────────────────────────────────────────────────╮
│  >_ OpenAI Codex (v0.106.0)                                                     │
│                                                                                 │
│ Visit https://chatgpt.com/codex/settings/usage for up-to-date                   │
│ information on rate limits and credits                                          │
│                                                                                 │
│  Model:                gpt-5.3-codex (reasoning medium, summaries auto)         │
│  Directory:            [REDACTED]                                               │
│  Permissions:          Custom (workspace-write with network access, on-request) │
│  Agents.md:            AGENTS.md                                                │
│  Account:              [REDACTED]                                               │
│  Collaboration mode:   Default                                                  │
│  Session:              019cb1b9-ab44-78c0-a572-6ef36ff7a3a7                     │
│                                                                                 │
│  Context window:       99% left (13.3K used / 258K)                             │
│  5h limit:             [███████████████████░] 94% left (resets 09:15)           │
│  Weekly limit:         [████████████████████] 98% left (resets 04:15 on 10 Mar) │
╰─────────────────────────────────────────────────────────────────────────────────╯
Token usage: total=9,926 input=9,677 (+ 16,128 cached) output=249 (reasoning 88)
To continue this session, run codex resume 019cb1b9-ab44-78c0-a572-6ef36ff7a3a7

No config changes on my side between now and before the increased burn rate. Windows OS. Upgrated to 106.0 but no idea if it started happening during update or just a coincidence.

In my main chatgpt plus subscription I have the same problem, I tried a second account with plus and there the use is normal, it does not have that problem

And some other people are reporting usage limits not being used at all:

https://www.reddit.com/r/codex/comments/1rjeav6/lmao_now_the_usage_hasnt_been_decreasing_for/

Now this may be caused by quota counter not updating for them yet (possibly will get refreshed and show correct state later) but I find it kind of weird that some accounts get normal burn reflected right away, some zero, some x3-5 the normal (rough estimate?).

If this is account specific, I can run more debug tests or provide extra details privately.

Poster above me is probably a better repro case if they get different burn rates in the same environment, just with two different accounts.

some other people are reporting usage limits not being used at all

That can be explained by the fact that we had an incident earlier today that affected some users, and we reset all rate limits for all users. In other words, if you had already used 50% of your limits for the week, you would be back to 100% after we performed the reset. So I don't think that's related to the issue reported in this thread.

We're continuing to look into the high-rate-of-consumption issue that seems to be affecting some users. We have some theories and are tracking them down. The clues and analysis provided above have been really helpful. If you discover any additional relevant details, please share.

I see. Tomorrow, I'll try intercepting some http requests to see if the codex tool is duplicating some requests on the client side. If you want me to run some extra debug level to let you correlate my test with some server side usage attribution trace, I'll happily do that. Maybe simple reauth can help, need to test that, too.

Normally I don't really question usage or quality variance because it's hard to quantify without hard data but today was different enough to notice a change.

Thanks for looking into that.

same issue, even simple tasks which earlier would at max deplete 5h quota by 1-2% is depleting it by 6-8% now. I am experiencing almost 5x faster depletion compare to before.

still 5h quota is draining much faster than before. last week 5h quota limit never dropped below 90%. now a single prompt causes 5% decrease.

edit:

pro user and xhigh all the time

Same thing here.. something happened between last Thursday until now.

One tip until has a definitive fix: use spark model when possible.. in this way you avoid consumption to be drained fast until Codex (or ChatGPT Model team have a fix) fix this

PS: i'm Pro user and never reached the limit before.

I used it for the first time after Codex reset the quota, and in the first prompt, it already consumed 8% of the daily quota and 3% of the weekly quota. There's definitely something very wrong; as a Pro user, this wasn't happening before.

If helps, here is the thread ID: 019cb321-ba68-7ab3-939b-878389b6f5de

Has anyone tested this against the most recent codex versions i.e. 0.105, 0.106 and now 0.107 ?

The issue still persists. Is there any way for me to downgrade the Codex app & Codex cli?

I noticed that if I use the GPT-5.2-Codex version, the limit usage is normal. Apparently, the problem only occurs with GPT-5.3-Codex

My friend says he basically can’t use up his quota—after using it for a long time, he’s only consumed 2%. In contrast, my Codex used up the 5 hours of quota after only 20 minutes of work. Like me, he’s using GPT‑5.3 Codex through the latest version of the Codex app (26.228.1430 (760))

We’re both on the Plus subscription, so I think there’s a bug in your quota service.

As I mentioned on other threads, we have not changed anything on the server side related to usage accounting or metering. If you're noticing a change in usage, it's likely in how you're using Codex. If you suspect this might be a client-side regression, you could try downgrading to an earlier version of the CLI to see if you observe the same behavior.

I don't know if you've changed something or not, but it's definitely happening — today Codex is literally guzzling tokens with any task. I use the VS Code plugin

I noticed this after enabling multiple agents in experimental settings on windows, I assumed it was due to background agents consuming tokens.

I noticed this after enabling multiple agents in experimental settings on windows, I assumed it was due to background agents consuming tokens.

I haven't opened multiple agents or done anything out of the ordinary, there's a real problem with token consumption and it's pretty serious.

“If you're noticing a change in usage, it's likely in how you're using Codex” is not an acceptable response.

A huge number of long-time Codex users have reported a dramatic increase in token burn over the past few days. Suggesting that this is simply down to individual usage patterns, and that everyone is mistaken, doesn't address the concern.

In my case, I’ve become significantly more cautious and deliberate in how I use Codex specifically because of the increased burn rate, yet the difference is still night and day. At the very least, there should be an acknowledgement that something may have changed or that the issue is being investigated.

“If you're noticing a change in usage, it's likely in how you're using Codex” is not an acceptable response. A huge number of long-time Codex users have reported a dramatic increase in token burn over the past few days. Suggesting that this is simply down to individual usage patterns, and that everyone is mistaken, doesn't address the concern. In my case, I’ve become significantly more cautious and deliberate in how I use Codex specifically because of the increased burn rate, yet the difference is still night and day. At the very least, there should be an acknowledgement that something may have changed or that the issue is being investigated.

Agree, I also reported here that I sent my first prompt after the reset, and as soon as I sent it, it reached 8% daily quota and 3% of weekly quota. And I'm a Pro user, I've never felt this before.

Please, OpenAI, don't be like Anthropic, which has had an issue since December about usage, with thousands of reports proving it, and no action is taken.

I also noticed a sharp increase in the usage of the weekly quota recently. I had to remove every single MCP server, skill, the AGENTS.md in the Codex folder, and set the reasoning to low in order to get the usage under control, the weekly limits were decreasing over 15% as soon as I send any request on a new chat. I am using the ChatGPT Plus tier ($20). Applying all those tweaks I reduced the usage rate significantly, and it is now useable again.

“If you're noticing a change in usage, it's likely in how you're using Codex” is not an acceptable response.

I apologize for being dismissive earlier. We receive reports from users pretty regularly that their usage rate is higher than they expect, and typically when we investigate we find that it's explained by a change in how they're using Codex. It's now clear that this situation is different, but it required additional evidence and user reports for that to become clear.

We're actively investigating the problem. We have a theory but haven't yet found the root cause. We thought we understood the problem yesterday and deployed a fix, but there is apparently more to it than we thought.

We just deployed a temporary fix to the service that we think will mitigate the issue, assuming our theory is correct. It's not a permanent fix, but it should eliminate the problem while we continue to look for the root cause.

I appreciate your patience and all of the clues and details you've provided. That has been really helpful.

I'll keep you updated as we learn more.

In the meantime, I switched to GPT-5.2-Codex, where the spending limits are normal.

If you were previously able to repro the problem with a small task, could you try it now and see if the problem has gone away? That will help us validate our theory.

If you were previously able to repro the problem with a small task, could you try it now and see if the problem has gone away?

Obviously subjective, but I'd say there's no change.

5.3-codex High. Mac app 26.228.1430. Quit and restarted app after your note here.

If you were previously able to repro the problem with a small task, could you try it now and see if the problem has gone away? That will help us validate our theory.

I launched a small task and I don't notice any change. Tthe 5h and weekly limits keep dropping at a very high rate.

Re-ran my test from here. https://github.com/openai/codex/issues/13186#issuecomment-3988443385

Same 2.5-3% 5h per 10k tokens (reading readme.md). Updated to 0.107 in the meantime. No change.
Session: 019cb4a0-32e6-7f23-8073-1035e0d6aec7, 019cb4a0-cdc7-75a3-9d20-76fd9f03e7a6

We receive reports from users pretty regularly that their usage rate is higher than they expect, and typically when we investigate we find that it's explained by a change in how they're using Codex. It's now clear that this situation is different, but it required additional evidence and user reports for that to become clear.

That's perfectly understandable and this is hard to test to be 100% sure that something is wrong. I don't want to send you guys on a wild ghost chase so if you can provide a way to test expected 5h/week % usage per tokens (stats at the exit) today, it could help us ensure we're not reporting expected burn rates as a complaint (assuming those are somewhat dynamic based on server load?).

Did /logout and re-auth. No change. 2% weekly + 8% 5h used by 3 sessions with one message (so no cache). Session closure token counts hover around this:
Token usage: total=10,106 input=9,930 (+ 16,128 cached) output=176 (reasoning 48)

So far today not noticing any issues. Quota consumption seems to be extremely slow again (great!) in comparison to recently. It's not actually moved away from 100% weekly yet. No issue with the weekly quota randomly jumping from one value to another yet either.

I found some of my older session files from 2 weeks ago (rollout-{guid}.jsonl). There's a huge difference in quota usage % per token. 26k tokens now use roughly the same quota as 6M used two weeks ago. This looks more like a quota my friend had on his free plan.

Two weeks ago:

Session starting at 2026-02-18T13:46:26.475Z
First occurrence of token usage block with a "resets_at" that matches the last message

        "info": {
            "total_token_usage": {
                "input_tokens": 14206,
                "cached_input_tokens": 4480,
                "output_tokens": 351,
                "reasoning_output_tokens": 139,
                "total_tokens": 14557
            },
            "last_token_usage": {
                "input_tokens": 14206,
                "cached_input_tokens": 4480,
                "output_tokens": 351,
                "reasoning_output_tokens": 139,
                "total_tokens": 14557
            },
            "model_context_window": 258400
        },
        "rate_limits": {
            "limit_id": "codex",
            "limit_name": null,
            "primary": {
                "used_percent": 1.0,
                "window_minutes": 300,
                "resets_at": 1771440383
            },
            "secondary": {
                "used_percent": 2.0,
                "window_minutes": 10080,
                "resets_at": 1771872028
            },
            "credits": {
                "has_credits": false,
                "unlimited": false,
                "balance": null
            },
            "plan_type": null
        }

Last message at 2026-02-18T14:03:32.374Z

"payload": {
        "type": "token_count",
        "info": {
            "total_token_usage": {
                "input_tokens": 5949985,
                "cached_input_tokens": 5763328,
                "output_tokens": 20910,
                "reasoning_output_tokens": 8904,
                "total_tokens": 5970895
            },
            "last_token_usage": {
                "input_tokens": 119855,
                "cached_input_tokens": 119168,
                "output_tokens": 1065,
                "reasoning_output_tokens": 385,
                "total_tokens": 120920
            },
            "model_context_window": 258400
        },
        "rate_limits": {
            "limit_id": "codex",
            "limit_name": null,
            "primary": {
                "used_percent": 5.0,
                "window_minutes": 300,
                "resets_at": 1771440383
            },
            "secondary": {
                "used_percent": 4.0,
                "window_minutes": 10080,
                "resets_at": 1771872028
            },
            "credits": {
                "has_credits": false,
                "unlimited": false,
                "balance": null
            },
            "plan_type": null
        }
    }

---

Now (2026-03-03T17:05:13.292Z) I get:
First "rate_limits" block

"rate_limits": {
            "limit_id": "codex",
            "limit_name": null,
            "primary": {
                "used_percent": 5.0,
                "window_minutes": 300,
                "resets_at": 1772574935
            },
            "secondary": {
                "used_percent": 3.0,
                "window_minutes": 10080,
                "resets_at": 1773112547
            },
            "credits": {
                "has_credits": false,
                "unlimited": false,
                "balance": null
            },
            "plan_type": null
        }

After one message to read the readme.md

"payload": {
        "type": "token_count",
        "info": {
            "total_token_usage": {
                "input_tokens": 26085,
                "cached_input_tokens": 6912,
                "output_tokens": 238,
                "reasoning_output_tokens": 67,
                "total_tokens": 26323
            },
            "last_token_usage": {
                "input_tokens": 13452,
                "cached_input_tokens": 3456,
                "output_tokens": 97,
                "reasoning_output_tokens": 0,
                "total_tokens": 13549
            },
            "model_context_window": 258400
        },
        "rate_limits": {
            "limit_id": "codex",
            "limit_name": null,
            "primary": {
                "used_percent": 8.0,
                "window_minutes": 300,
                "resets_at": 1772574935
            },
            "secondary": {
                "used_percent": 4.0,
                "window_minutes": 10080,
                "resets_at": 1773112547
            },
            "credits": {
                "has_credits": false,
                "unlimited": false,
                "balance": null
            },
            "plan_type": null
        }
    }

Both gpt 5.3 codex, the older one had mixed medium and high reasoning while the more recent one had only medium.

Codex CLI 0.107.0 ChatGPT Plus

Prompt:
Make a complete audit of the current project, write the audit to project_audit.md, the audit must focus on all features this entire project provides, write each distinct feature on a single line only, take your time and make sure the audit does not miss any features- you are not gettting a second chance to correct yourself.

Codex 5.2-high
wc -l project_audit.md -- > 79 project_audit.md ─ Worked for 8m 04s - 2% 5hour window usage

Codex 5.3-high
wc -l project_audit.md -- > 253 project_audit.md ─ Worked for 9m 14s - 46% 5 hour window usage

Now while the answer from 5.3 is probably better, is it 23x times better?
This token consumption was not present on 5.3 release so I assume something is messed up.

@ersomm-dev, that's really useful data. If you have those two threads around still, would you be willing to resume each in turn and then use /feedback to upload the session details? If you post the thread IDs here, we could then do a side-by-side comparison of the rollouts.

@ersomm-dev, that's really useful data. If you have those two threads around still, would you be willing to resume each in turn and then use /feedback to upload the session details? If you post the thread IDs here, we could then do a side-by-side comparison of the rollouts.

Here:

019cb4d7-8207-7371-a5e3-82531848eeb5

019cb4e2-bf7f-7b41-8114-584de7d09bde

Here same problem, just to add to the bug reporting. Burning tokens on 5.3 medium like 5x faster then last week on 5.3 high/xhigh.

Usage: Codex vs code extension
Env: W11 wsl with ubuntu 24

> “If you're noticing a change in usage, it's likely in how you're using Codex” is not an acceptable response. I apologize for being dismissive earlier. We receive reports from users pretty regularly that their usage rate is higher than they expect, and typically when we investigate we find that it's explained by a change in how they're using Codex. It's now clear that this situation is different, but it required additional evidence and user reports for that to become clear. We're actively investigating the problem. We have a theory but haven't yet found the root cause. We thought we understood the problem yesterday and deployed a fix, but there is apparently more to it than we thought. We just deployed a temporary fix to the service that we think will mitigate the issue, assuming our theory is correct. It's not a permanent fix, but it should eliminate the problem while we continue to look for the root cause. I appreciate your patience and all of the clues and details you've provided. That has been really helpful. I'll keep you updated as we learn more.

It's all good - think anyone who has worked a tech support style shift of dealing with bugs etc. has had that experience, and I know I've dismissed things before or assumed it'd be an obvious 'user error'. Most appreciated is that you've not just stubbornly stuck to it and have been willing to investigate and identify that there is an issue.

I can't say if this would've used 1% of 5 hourly before this issue arose with 13.5k tokens used - maybe, gut feeling is no - but no hard proof. But for what is as close to an identical change as what I put in my original bug report, it's now using 1% of my 5 hour, not 2%. I still think 1% is high for what is the most minor of uses unless it drops to that on any use and will stay on 1% for a while longer - just wanted to point it out.

However, the /status and my status bar are showing some different results:

019cb516-33c3-7201-93bf-32837604dea3 (I've uploaded with /feedback)

<img width="1278" height="1236" alt="Image" src="https://github.com/user-attachments/assets/89f1c473-7c83-43a2-b132-8f2a350d6cca" />

You can see that it was a fresh session, left a single comment line - most minor of minor changes. gpt-5.3-codex medium.

My /status says 13.5k used, 99% 5 hour, 98% weekly.

statusline shows 1% used 5hr, 95% weekly and 53.1k tokens used. Now I have an MCP server attached, but it's a custom one designed to use no more than 1.5k tokens (if that). I don't believe that is the issue - definitely not 53k tokens when it wasn't even called. And the discrepancy between the weekly usage is also odd.

More than happy to run any further tests or provide more information - just let me know.

@Superfishintights, if your /status and statusline don't agree, then that's a bug in the TUI (most likely the "statusline" feature, since we just recently added it). Could you please file a separate bug report on that so it doesn't get lost?

@etraut-openai I'm not sure if this will help in identifying the cause of the issue, but besides the problems users described above, I've also noticed that the context estimation sometimes jumps up by 2–3%. For example, it can be 63% and then suddenly jump to 65%.
In any case, thank you for working on this issue.

Can confirm this, token burn is so much higher even from simple changes

I can confirm this on the free tier as well — the weekly limit is draining roughly 10x faster than before. I used to get extensive use out of it, lasting well over 3 days. Now I exhaust my quota in as few as 3 prompts.

I confirm that if I select the 'GPT-5.2' model, the limits are the usual ones and the dramatic consumption that can be seen in GPT-5.3-Codex does not occur.

I bought a bussiness "workspace" account 1 month for $2 to test this and the new account dont have this "bug" so this is related to certain accounts server side, they silent capped, flagged or whatever many accounts for whatever reason, also tested my main acc outside codex and the bug is still present so this is not related to codex cli only. So low and pathetic behaviour from OpenAI...

A shame from your company @etraut-openai , what a joke.

Can confirm the same for me, hammering through my usage and rates not seen before. Burnt 70% of the week's quota on 10 requests, two on high, 8 on medium. Previously that would have used about 10% based on last week. I thought we were double quote until month end? Using Codex app on Mac.

yes, I know it is gut feeling, but if anything I've used codex less recently, but it burned 30% of my Pro quota in just one day, never seen the same consumption rate in a past 5 months

hopefully once you find and fix the issue - you will credit your users back, my week reset just happened and I've already wasted so much of my quota

I'm experiencing the same problem on codex CLI v0.107.0 on arch. The token burn rate is extremely high compared to last week, even with lighter model.

Mine still seems fine since yesterday I believe, with whatever was changed on the backend my quota consumption is extremely slow now. So slow that I'm still seeing 100% weekly available in the Codex usage on the website despite some usage with multi-agents and 5.3 Codex xhigh and numerous auto compactions. Pro plan use, currently using Codex CLI 0.108-0-alpha. 👍

I’ve had a similar experience over the last few days with both Codex cli and Codex desktop app on the Plus plan. Using the Codexbar macOS app, I only used about 2M tokens but already hit half of my weekly limit. In contrast, I’ve used over 100M tokens elsewhere and still haven’t reached the weekly limit.

https://github.com/openai/codex/pull/13454 this will surely be helpful to diagnose issues like this.

I ran out of my weekly on Friday, topped up 40EUR of credits. 4 hours later I'm out of credits.

Saturday, add another 40EUR worth of credits. 2 hours later I'm out of credits.

This morning I spend 1.5 hrs working; my five hour quota is already out.

This evening I work for 2 hours; five hour quota out.

What is this, the five hour quota is actually a 2 hour quota and "you can only work for 5 hours per week"? which is what it's looking like because I'm now down to 27% of my weekly - 73% of weekly burnt since last night which is when my weekly reset, and only working for 5 hours total.

The burn is absolutely obscene and this is not usable. It's not like I'm working fast, 5.3 Codex is slow and I'm using a single agent.

I ran out of my weekly on Friday, topped up 40EUR of credits. 4 hours later I'm out of credits. Saturday, add another 40EUR worth of credits. 2 hours later I'm out of credits. This morning I spend 1.5 hrs working; my five hour quota is already out. This evening I work for 2 hours; five hour quota out. What is this, the five hour quota is actually a 2 hour quota and "you can only work for 5 hours per week"? which is what it's looking like because I'm now down to 27% of my weekly - 73% of weekly burnt since last night which is when my weekly reset, and only working for 5 hours total. The burn is absolutely obscene and this is not usable. It's not like I'm working fast, 5.3 Codex is slow and I'm using a single agent.

"You're right, it's not usable. Try the GPT-5.2 agent. It's slower, but at least in my case the token consumption is normal. It seems to be something that affects 5.3-codex

Canceled my Pro subscription until this gets fixed. I lost like ~3days because of this

Previously affected by the issue with significantly increased quota usage rate.
Today: the rate limits seem to be not returned at all. Got "rate_limits": null in event_msg in the jsonl. /status shows

│  Limits:               data not available yet                                   │

https://chatgpt.com/codex/settings/usage shows 1% used of 5h after 2 "hello-world"s which seems more in line with expected usage before I got hit with the burn rate issue (yesterday: 6-7% with the same 2 tests)

No client changes on my side since yesterday, same gpt 5.3-codex so the changes come from the server side.

As for the suggested shadow-limitting of suspected 3rd party users - I never used my account for anything else than official codex client and I've never even came close to exhausting any limits, so I highly doubt intentional limiting of power users.

Our earlier theory proved incorrect. We have a new theory and just deployed a potential fix. If you are a Plus or Pro subscriber who is seeing this problem, please try to repro now.

Seems fixed, post-fix usage matches 5.2 codex metrics based on test from yesterday.

Tested the same prompt that previously used 3% 5h each, repeated 3 times. Now 0% of 5h moved after waiting 10 minutes for web dashboard usage to update. One thing (in addition to the insane burn being gone) I can notice is the rate limits are not being returned by the response so /status has no data to display right away.

May we ask if it is something on our side causing the issue for our accounts? Anything that we could potentially NOT do so we don't run into that again? Not sure if you can share service internals or potential abuse prevention systems but it would be great to know if it's something that we can control.

(Plus sub)

Edit: tested with longer session. Significantly lower quota usage.

Can form it is way better! however will keep testing for the next couple of hours to confirm. My weekly limit is depleted though due to this bug

Seems fixed indeed. Ran 4 prompts now without a decrease in weekly, earlier today it woulda consumed probably 5% or so.

(Pro account)

Seems ok now. One task consumed no % vs yesterday which would have consumed 2% at least. Still left hanging on single digit usage with 6 days to go before reset though. I assume usage will get reset?

Another confirm here, it looks fine now on the latest release version. tried some prompts on 5.3 medium and consumed 0% weekly and 1% in the 5 hours gap, in the last 2 days that would have been a good 2/3% on the weekly usage instead. i still have a bug where i see the weekly from the vs code extension lower then the value on the codex usage website but this is not correlated here. same as @cjstrads i depleted my tokens in this 2 days testing this out. a token clearance on accounts which had a spike on token usages on those days could be the right formula to address this out resetting the tokens to a new week counter.

Thanks for the confirmation. We are now confident that we have identified and fixed the issue. It was causing the 2X promotional increase in limits to not be applied for an estimated 9% of plus and pro users. We have now fixed this issue and are reseting the rate limit for all plus and pro users to compensate for the trouble. Thank you all for reporting.

Hmm the explanation doesn't make sense. In that case, token drain should've been just twice as fast... but a lot of people reported much higher... also it would mean that once the 2x promo ends, the limits would be almost useless which I don't believe considering I've been using Codex for many months and it was never this bad.

Hmm the explanation doesn't make sense. In that case, token drain should've been just twice as fast... but a lot of people reported much higher... also it would mean that once the 2x promo ends, the limits would be almost useless which I don't believe considering I've been using Codex for many months and it was never this bad.

True. My experience is that I had never been able to consume all my 5-hours limit, but yesterday it was gone in less than 1 hour. That cannot be the normal usage.

Hmm the explanation doesn't make sense. In that case, token drain should've been just twice as fast... but a lot of people reported much higher... also it would mean that once the 2x promo ends, the limits would be almost useless which I don't believe considering I've been using Codex for many months and it was never this bad.

etraut-openai No, the explanation doesn’t make sense for two reasons: first, the tokens were being burned at a rate much higher than 2×; and second, this morning I switched to GPT-5.2 and the token burn rate was correct. The issue seemed to be limited to 5.3-codex. I hope the solution wasn’t just a temporary 2× token safeguard.

yes, I got my tokens budget reset back to 100%, which is great. but here is the question, so after promotional period is over - is this what we are going to expect - the same crazy-fast consumption as we saw recently?

Using GPT 5.2 High right now and the extreme drain is gone. Still, how do I get a refund for my 80EUR of credits?

Yes, this explanation doesn’t hold up. @etraut-openai

1. The consumption rate is far more than double. If you only temporarily add double the quota, then by April my account will still become unusable.
2. You might have noticed that some accounts can barely use up their quota at all—something your theory can’t explain. For example:

> And some other people are reporting usage limits not being used at all: https://www.reddit.com/r/codex/comments/1rjeav6/lmao_now_the_usage_hasnt_been_decreasing_for/

1. My friend’s quota used to barely decrease as well, but starting yesterday it also began dropping extremely fast. That means this bug is dynamic—it’s not simply that certain people permanently have half the quota.

More nonsense comments from @etraut-openai , gj obeying your master Sam and giving us nonsense excuses... the consumption rate was WAY ABOVE 2x, that was a 7x-10x from normal usage...

All of the sudden, I am using way more usage than before. I've never used so much in such a short time, and I was barely using it. I never had this problem before.

1 prompt and it takes 10% of the weekly usage for GPT-5.3 Codex. The issue doesn’t seem to be fixed yet.

@etraut-openai I'm on the $200 plan. On March 3–4 (right after the limit reset from the high cyber risk bug) I was running multiple tasks in parallel across several projects. My weekly limit only dropped by about ~10% over those two days.

Today I’ve only been working a couple of hours and haven’t done nearly as much as I did on March 3/4, yet I’m already down 11% of the weekly limit.

Edit: down 22% and workday isn't even done yet. It wasn't like this before. I've been enjoying the 2x limit before - and now it's worse than before the 2x limit.

1 prompt and it takes 10% of the weekly usage for GPT-5.3 Codex. The issue doesn’t seem to be fixed yet.

so I guess it wasn't fixed for all customers. for me it worked fine today, similar usage as I had prior to this "crazy consumption" issue. and I got the entire balance (back to 100%) today.

Thanks for the confirmation. We are now confident that we have identified and fixed the issue. It was causing the 2X promotional increase in limits to not be applied for an estimated 9% of plus and pro users. We have now fixed this issue and are reseting the rate limit for all plus and pro users to compensate for the trouble. Thank you all for reporting.

That sounds wrong.

I have never used up my quota ever before.
While this bug was active, I could use the entire quota in a day (I even purchased 40 USD of credits, which was all consumed in a couple of hours) - so it was way more than just a 2x.

So no, something else was wrong.

• if it was only the campaign, it means I cant use Codex anymore after the campaign.

I have been using Pro since it came out.

Edit: Consumption SEEMS fine now though, way less than factor of 2 compared to previous days - so its looking good, but something with the story doesn't seem to add up - must be a missing piece somewhere.

> Thanks for the confirmation. We are now confident that we have identified and fixed the issue. It was causing the 2X promotional increase in limits to not be applied for an estimated 9% of plus and pro users. We have now fixed this issue and are reseting the rate limit for all plus and pro users to compensate for the trouble. Thank you all for reporting. That sounds wrong. I have never used up my quota ever before. While this bug was active, I could use the entire quota in a day (I even purchased 40 USD of credits, which was all consumed in a couple of hours) - so it was way more than just a 2x. So no, something else was wrong. * if it was only the campaign, it means I cant use Codex anymore after the campaign. I have been using Pro since it came out. Edit: Consumption seems fine now though, way less than factor of 2 compared to previous days - so its looking good, but something with the story doesn't seem to add up - must be a missing piece somewhere.

I can confirm this. This issue is still happening on my side as well, though it _has_ slowed down a bit. _Something is still wrong._

I need to take back my previous post about all being ok - looks like it is all ok if reasoning is low or medium, but when on High it absolutely rips through the usage. My code base is smaller now than a week ago, but usage has increased. Has Codex itself changed under the hood? Either way, still not great if this is double. I'm thankful for the usage reset, but it's likely only temporary relief based on the last few hours experience.

@etraut-openai The limits of my account have not been restored; it seems the solution has been erratic

> > Thanks for the confirmation. We are now confident that we have identified and fixed the issue. It was causing the 2X promotional increase in limits to not be applied for an estimated 9% of plus and pro users. We have now fixed this issue and are reseting the rate limit for all plus and pro users to compensate for the trouble. Thank you all for reporting. > > > That sounds wrong. > I have never used up my quota ever before. While this bug was active, I could use the entire quota in a day (I even purchased 40 USD of credits, which was all consumed in a couple of hours) - so it was way more than just a 2x. > So no, something else was wrong. > > * if it was only the campaign, it means I cant use Codex anymore after the campaign. > > I have been using Pro since it came out. > Edit: Consumption seems fine now though, way less than factor of 2 compared to previous days - so its looking good, but something with the story doesn't seem to add up - must be a missing piece somewhere. I can confirm this. This issue is still happening on my side as well, though it _has_ slowed down a bit. _Something is still wrong._

Yes, that's what it looks like. Just by seeing the explanation they've given, it's clear that the solution was a hack to increase the tokens, but the underlying problem is still there

@etraut-openai I confirm that after reverting to 5.3-codex (from 5.2), the unusual consumption is still present, to a lesser extent but it's there. Additionally, the tokens have not been restored on my account. Please clarify what OpenAI's stance is on this, whether this is going to stay as it is now, because we can't work like this and we have decisions to make

Thanks for the confirmation. We are now confident that we have identified and fixed the issue. It was causing the 2X promotional increase in limits to not be applied for an estimated 9% of plus and pro users. We have now fixed this issue and are reseting the rate limit for all plus and pro users to compensate for the trouble. Thank you all for reporting.

This is a really confusing response. I would have thought that by now it was clear that the issue was not that a 2x promotional increase wasn't being applied. The issue for many of us was that the token burn rate had significantly increased from normal. We're not talking about a slight increase, like a 10x + increase and a user even took the time to demonstrate this above.

The limits in my account were not reset either. And today, I'm having heavy usage consumption which doesn't make sense.

the heavy token draining has been fixed for me. weird that it still happens for other people.

pro and 5.3 xhigh

Regressed - my account got reset this morning, I've never had this issue prior to this, suddenly my Pro account consumed 10% weekly within a couple of hours.

5.3-codex is burning through tokens like an idiot. I'm blowing through the weekly limit in a single afternoon, which, by the way, hasn't been reset for me like it has for other users. Meanwhile, this issue remains closed and @etraut-openai nor anyone else is making any comment. It's clear they're comfortable with the situation. Luckily, our accounts are on monthly billing

I'm on plus, gpt-5.3-codex medium just ate 50% of my 5 hour limit in 2 prompts

Update: yesterday I confirmed usage seemed normal again. Today, the token burn is back to “unreasonably high.”
@etraut-openai something is still not right here. Let us know what diagnostics you need, because this keeps flipping between “fixed” and “broken.”. I would not consider this issue as closed at least for now.

Seeing the same as @VasilisTako, @etraut-openai. Quota consumption seemed normal (or close to it) for part of a day. Today, I'm back to consuming quota significantly higher than what was historically offered.

Went down 33% weekly limit in one day on the Pro account with relatively light use throughout the day. Same level of usage yesterday barely made a dent. Something is definitely broken

Went down 33% weekly limit in one day on the Pro account with relatively light use throughout the day. Same level of usage yesterday barely made a dent. Something is definitely broken

This is completely unacceptable, the whole day without even bothering to respond. They care more about announcements than the usability of their product.

I've confirmed that the issue that caused this problem for most users in this thread has been addressed. It sounds like there is at least one more issue that affects some of you. We're tracking that in #13568. Please follow along there for updates.

I'm going to lock this thread.