Codex Business usage drains 5–10× faster than Plus under identical GPT-5.5 task

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

• #23607
• #22157
• #23297
• #23457

Powered by Codex Action

This seems related to #14593, but my case is specifically a Plus-vs-Business metering discrepancy under identical conditions.

Would help a lot if identical-task comparisons produced a per-run usage receipt rather than only account-level bars.

Something like: model, reasoning level, admitted context size, tool-call count, compactions, subagents, and terminal stop reason for each run. Then support can separate a real plan-tier metering bug from "same prompt, different runtime path" without asking users to infer from dashboards.

@Keesan12 here is my results

As you requested, I conducted an experiment that substantiates my concerns. I endeavored to present the findings as transparent and representative as possible.

I consolidated my personal (ChatGPT Plus) and business accounts into a single environment. Notably, I conducted the test without altering the codebase on my personal account. Subsequently, I logged into my business account on the same application and device, copying the identical prompt.

The following rules were adhered to:

• Same machine
• Same codebase version
• Same repository/codebase
• Same prompt
• Same model: GPT-5.5
• Same reasoning level: high
• Same local/cloud mode: local
• Fresh session

The tests were the initial tasks I performed upon waking up. Consequently, they were executed approximately 15 hours after inactivity in Codex. Despite this, when I opened Codex in either account, the 5-hour limit was already below 100%. It was 99% in my personal account and 96% in my business account, which is unexpected in itself. These levels were prior to sending any prompts.

I commenced from the personal account, and the initial state is as follows:

<img width="1280" height="959" alt="Image" src="https://github.com/user-attachments/assets/b5bee989-bccb-432a-aaa0-9769a5c39891" />

After some time, I received a final response.

<img width="1269" height="960" alt="Image" src="https://github.com/user-attachments/assets/806ba62b-9407-4dfc-bf19-82ff54c15a59" />

As evident, the 5H limit has been reduced from 99% to 81%. Consequently, _it took 18%_ to complete this prompt.

Upon logging out and logging into the Business account, despite the fact that both the 5H and 7D limits were sufficient for my work, I encountered a warning indicating that I had exceeded the available messages. This message persisted, which could be considered a bug.

<img width="1269" height="960" alt="Image" src="https://github.com/user-attachments/assets/bd13a28e-4ebf-4080-9bc3-11c7053742a6" />

Subsequently, I replicated the identical prompt from my personal account. Please find the initial state below:

<img width="1269" height="960" alt="Image" src="https://github.com/user-attachments/assets/b53bf940-d910-46d9-81ac-eec0640577df" />

After some time, I have obtained the desired result and have exhausted the remaining 5-hour limit at the 46% level. Consequently, this task has _consumed 50%(!!!)_ of the 5-hour limit.

<img width="1269" height="960" alt="Image" src="https://github.com/user-attachments/assets/04702ba4-83b2-4ff0-8199-f151fc4cc211" />

Below are the outcomes of my experiment:

• In this scenario, we observed that the initial 5H limit at the beginning of the work was already below 100%.
• An irrelevant warning message was displayed on the business account.
• The business account utilized 2.8 times more tokens (18% compared to 50% of the 5H limit) to complete the same task.

Although 2.8 is not the highest difference relative to what I typically encounter, it is challenging to establish a clear comparison every time.

I should emphasize that this issue has not recently emerged. The wired token usage was observed from the day I began using my Business account. Therefore, I believe the money spent on the subscription did not provide the expected service, as indicated on the product page where it states, «Business ChatGPT & Codex includes Everything in ChatGPT Plus and Business Codex plans.» However, my experiment clearly shows that it includes at least three times less.

***

After conducting the test, I refrained from accessing my Business account. I logged out with 46% of my 5-hour limit remaining.

After an hour, I logged back into my Business account and was surprised to find that the 5-hour limit had been fully utilized showing 0% left.

It is important to note that no requests were submitted since I left my account with 46% and discovered it with 0% later.

<img width="1269" height="960" alt="Image" src="https://github.com/user-attachments/assets/41da2ebe-9037-441f-8cc7-92889aff2f0a" />

This is just ridiculous.

Update: possible explanation found for part of the discrepancy

I found that the Business account had Codex Speed set to Fast.

The UI says:

• Standard: default speed
• Fast: 1.5× speed, increased usage

There is also a Codex changelog / public discussion indicating that eligible ChatGPT plans may now default to the Fast service tier unless explicitly opted out.

This may explain part of my observed Plus vs Business discrepancy, because my Business account may have been using Fast while Plus was using Standard.

However, I still think this is a serious transparency/UX problem:

1. The public pricing page suggests Plus and Business have comparable Codex included usage.
2. Business can silently/default run in a higher-usage mode.
3. The main task UI does not make this sufficiently obvious.
4. In my test, Business consumed 50% vs Plus 18%, around 2.8× more.
5. After I stopped using Codex and logged out with 46% remaining, the Business account later showed 0% remaining without further requests.

I will repeat the A/B test with both accounts explicitly set to Standard and post the result.

Update: possible explanation found for part of the discrepancy I found that the Business account had Codex Speed set to Fast. The UI says: Standard: default speed Fast: 1.5× speed, increased usage There is also a Codex changelog / public discussion indicating that eligible ChatGPT plans may now default to the Fast service tier unless explicitly opted out. This may explain part of my observed Plus vs Business discrepancy, because my Business account may have been using Fast while Plus was using Standard. However, I still think this is a serious transparency/UX problem: 1. The public pricing page suggests Plus and Business have comparable Codex included usage. 2. Business can silently/default run in a higher-usage mode. 3. The main task UI does not make this sufficiently obvious. 4. In my test, Business consumed 50% vs Plus 18%, around 2.8× more. 5. After I stopped using Codex and logged out with 46% remaining, the Business account later showed 0% remaining without further requests. I will repeat the A/B test with both accounts explicitly set to Standard and post the result.

I checked it as I have the same problem and mine was set to 'fast'... I didn't set it.

Adding another data point that may be related to the unexpected Fast-speed explanation in this thread.

Environment:

• Codex Desktop Microsoft Store package: OpenAI.Codex_26.601.2237.0_x64__2p2nqsd0c76g0
• Package version: 26.601.2237.0
• Platform: Windows, Microsoft Store installation
• Model/intelligence setting observed: GPT-5.5 High

Observed behavior:

• After updating Codex Desktop on 2026-06-04, the Speed setting was silently set to Fast for GPT-5.5 High.
• I did not opt into Fast.
• Because Fast is described as an increased-usage mode, a silent switch to Fast can make quota drain look like a metering/account discrepancy even if the immediate cause is that the client selected a higher-usage speed tier without explicit consent.

Expected behavior:

• Codex Desktop should preserve the existing Speed preference through updates.
• If Fast is selected automatically because of account eligibility, plan defaults, or migration behavior, the app should make that explicit and require confirmation before switching to an increased-usage mode.
• The main UI should make the effective Speed obvious before a turn is sent.