@Ric_RTP sorry but the math is supposed to work like this: not sure about salaries in Stockholm but most fired engineers cost much more than the $2000 per month. And the ones that remain do 5x the work they did prior to using AI with a $2000 per month budget for tokens
@BitcoinSapiens You can't make an abundance of private seaside houses with perfect views. So if money goes away, how do we determine who gets the best view?
@openclaw@OpenAI that's two bans back to back. 5.7 days the first time. got turned back on then banned immediately. was only refactoring some code which is exactly what @steipete suggested. @sama what gives?
@openclaw upgraded to @OpenAI business and they banned me twice after one afternoon's worth of work. nearly 5.7 and 5.9 days ban for each time. Impossible to get their support to tell me why or turn it back on after endless back and forth
@steipete help - I watched your interview with @lexfridman and heard @sama telling us to go to openai for @openclaw models. but I just got banned AGAIN for 6 more days for doing nothing but trying to vibe code an app for my private use. @OpenAI support gave me the runaround
Just moved our team to ChatGPT Business to trial Codex (switching from Anthropic)…
Within 24 hours: locked out for 5.7 days
“Usage limit reached (team plan)”
We’re NOT using the API.
This is Codex via ChatGPT OAuth.
Total usage ~600k tokens.
The “trigger”?
A simple background worker/heartbeat checking a todo list.
So what does “unlimited” actually mean?
If any real dev workflow (even light automation) trips abuse systems, that’s a pretty big gap between marketing and reality.
@OpenAI@sama — is this expected, or a false positive?
Hard to justify switching if this is how Business + Codex behaves in practice.
It's as if they have done nothing based on that last reply from @OpenAI . Meanwhile we did our own research and confirmed we had not use nearly any of the ChatGPT model as we had just turned it on and were starting to find ways to work it into our workflows.
Here's our full GPT-5.4 usage history:
Mar 15: 1 successful task (write coding tests) — first ever run
Mar 15: 1 bug sweep spawn
Mar 17: 1 connection test (single message)
Mar 19 morning: 1 spawn (died silently, respawned on Sonnet)
Mar 19 evening: ~3 spawns for coding test workers → rate limited after the first one
Mar 20 (today): 1 spawn → immediately rate limited
That's maybe 7-8 total requests across 5 days. Their message said 33-168 messages per 5-hour window. We're not even close to 33 in a 5-day window.
Status: account still locked.
After giving @OpenAI that massive list of technical detail that they wanted, the come back with a boilerplate answer that makes zero sense. It is as if we haven't provided any information and they have done zero investigation.
Hello,
I understand you’ve provided detailed logs and technical context regarding the rate limit behavior you observed while using Codex through OpenClaw, and I truly appreciate the thorough information. I am here to help clarify.
The number of Codex messages you can send depends on several factors, including the model used, the size and complexity of your coding tasks, and whether tasks are executed locally or in the cloud. For example, small scripts or routine functions may only consume a small portion of your allowance, while larger codebases, long-running tasks, or extended sessions that require more context can use significantly more per message.
For ChatGPT Business plans, local Codex message limits over a 5-hour window typically range between approximately 33–168 messages, depending on these factors. This means that even a small number of requests can trigger limits if they are computationally intensive or require maintaining large context.
Codex usage limits are not strictly tied to request count alone—they are dynamically calculated based on workload complexity and execution patterns. In cases where requests are clustered (as seen in your timestamps), this can contribute to rate limiting even if the total number of requests appears low.
Additionally, based on the setup you described (using OpenClaw as an intermediary layer), it’s possible that some of the behavior you’re encountering may be influenced by how requests are being managed, queued, or retried within that environment. This can sometimes result in patterns that trigger limits even when the apparent usage seems minimal.
For future reference, you may review Codex Pricing on the OpenAI developer website for a full breakdown of limits and rates and Using Codex with your ChatGPT plan
If you have further questions, please feel free to reach out.
Best,
[Redacted]
OpenAI Support
Just so you know what you're up against with @OpenAI support after freezing your business account for using @openclaw - after 16 hours they replied with this extensive task list to keep us busy:
Hello,
I understand the context you’ve shared around your OpenClaw setup and how the usage pattern is intended to be lightweight, and I appreciate the detailed explanation along with the screen recordings. I am here to help move this investigation forward.
To proceed with a deeper review, I’ll need a few additional technical details from your side:
1 Workspace / Org ID (explicit workspace_id or org ID to look up runtime entitlements)
2-Request_id(s) and exact timestamps for the blocked requests (oreqs/DR links or router request IDs)
3-Server logs / DR links or Datadog traces for the blocked requests
4-Billing/account subscription confirmation (billing account ID) from the workspace admin
5-Clarification on whether OpenClaw used an older authentication token (include token/session ID)
6-The exact location where the message appears (you mentioned terminal—please confirm if it appears anywhere else such as UI or API), along with any HAR files if available
7-Whether other users in your workspace are experiencing the same issue
Once I have these details, I’ll be able to have a more in-depth investigation.
Best,
[Redacted]
OpenAI Support
The big catch: The largest model Microsoft has actually released is 2 billion parameters, not 100 billion. The 100B claim comes from benchmark demonstrations with dummy/test models, not a trained, usable 100B model you can download and run. The framework can theoretically handle that scale, and Microsoft demonstrated it with benchmark scripts, but nobody has actually trained a 100B ternary model and released it.
Also FYI
@OpenAI
if you want to beat
@AnthropicAI
in the race to win businesses, you can't kill someone's account for 5.7 days with no warning. You want these systems to be mission critical and downtime like that will send your customers elsewhere.
@openclaw@sama Also FYI @OpenAI if you want to beat @AnthropicAI in the race to win businesses, you can't kill someone's account for 5.7 days with no warning. You need to understand these systems will eventually be mission critical and downtime like that will send your customers elsewhere.
FYI @openclaw folks - is anyone able to use openclaw on an OpenAI business plan?
@sama literally went on stage and pitched the Codex OAuth integration as the way to use GPT through agents like OpenClaw. The whole point is background automation. If a single cron job reading a markdown file 24 times trips their abuse detection, then every OpenClaw user on a team plan is one heartbeat configuration away from a 5.7-day lockout.
Just moved our team to ChatGPT Business to trial Codex (switching from Anthropic)…
Within 24 hours: locked out for 5.7 days
“Usage limit reached (team plan)”
We’re NOT using the API.
This is Codex via ChatGPT OAuth.
Total usage ~600k tokens.
The “trigger”?
A simple background worker/heartbeat checking a todo list.
So what does “unlimited” actually mean?
If any real dev workflow (even light automation) trips abuse systems, that’s a pretty big gap between marketing and reality.
@OpenAI@sama — is this expected, or a false positive?
Hard to justify switching if this is how Business + Codex behaves in practice.