Documenting AI's Worst Moments
OpenAI built ChatGPT. Then it built Codex, the agentic coding tool meant to be its developer crown jewel. This page collects what real, named users are writing about it on the public record: the GPT-5-Codex performance collapse, the rate limits that vanish in minutes, the hallucinated workflows, the ignored instructions, and the time it ran a database migration and erased a user's data. Every quote is linked to its live thread on the OpenAI Developer Community and the official openai/codex GitHub. No paraphrasing.
The complaints below are not cherry-picked outliers. They are the dominant tone of OpenAI's own Codex support forum and the project's own GitHub issue tracker through late 2025 and the first half of 2026. Across hundreds of posts the same words keep returning without coordination: "unusable," "insane," "lazy," "dumb," "stuck," "trust." That convergence is the signal.
The pattern runs in two waves. First, after the GPT-5-Codex update in September 2025, users reported the tool became dramatically slower and stopped finishing tasks. Then, after a series of usage-limit changes in April 2026, the same paying users reported their quotas evaporating in minutes. Both waves ended the same way for many of them: cancellation.
Every quote is from a named account on a public OpenAI or GitHub thread. Section sources are linked beneath each group so you can read the originals in full.
"GPT5 Codex is very very slow. Unusable. Simple tasks that it used to complete in seconds now take up to 20 minutes."
The opening complaint pattern after the September 2025 Codex update: tasks that took seconds now take twenty minutes. It is repeated almost verbatim by dozens of other users in the same thread.
"coding tasks that GPT-4.1 (and even 4o) handled smoothly are now 4-7 times slower"
A direct, measured regression claim: the new Codex model is multiples slower at the same work older models did fine.
"before, I'd get commits done in 2-3 minutes; if something was heavy, it would take 5 minutes. Now, on average, it takes 20 minutes"
A 4x to 10x slowdown on the same routine commits, described by a working developer.
"many (small) tasks running for over 20 minutes today and some longer tasks running for 60+ minutes"
Hour-long runtimes on small tasks. The thread fills with the same numbers over and over.
"when it spend 40, 50 mins just to fail, oh man, that is bad"
The worst-case combination users keep describing: long runtime followed by failure, with the clock and the quota both spent.
"steps lasting 2 minutes each to steps of 20 minutes"
A clean before/after on per-step latency.
"excruciatingly slow so as to make it practically unusable"
"Unusable" is the single most-repeated word across the entire Codex regression record.
"The work is extremely slow and gives errors."
Slow and wrong, in the same breath.
"this seems like a degradation to me"
The understated version of what everyone else in the thread is shouting.
"extremely slow even when thinking is set to minimal (gpt5.2) ... average around 20 minutes to get result"
Four months after the first wave, the same 20-minute number reappears, this time on GPT-5.2.
"Sometimes 20-30 minutes for 200 character changed"
Half an hour to change two hundred characters of code.
"It's taking well over an hour to get even basic responses"
By January 2026 the reported latency had crossed an hour for basic requests.
"He has been very lazy and dumb lately which is making my work with it unproductive."
The title of this user's GitHub issue is "Please make your model stop being stubborn, lazy and dumb." The body is a list of regressions over a three-week window.
"He does not follow what we ask and do it the way he want."
The core agentic-tool failure: it ignores the instruction and does its own thing.
"Stop coming back when nothing done or like one line change."
Returning from a long run with one line changed, or nothing at all.
"Codex ignores instructions, repeats failures, and continues to pursue the 'original' implementation"
Filed under "severe degradation in Codex CLI," across both a personal Pro license and a work Enterprise license.
"Continued attempts to correct results in greater deviations from any established requirements or instruction"
Correcting it makes it worse, not better.
"Unusable code and a wasted day"
The bottom line of the report.
"5.2 xhigh has been even worse than 4o-mini... It's completely ignoring rules and not following my prompts"
A flagship coding configuration compared unfavourably to a mini model from a prior generation.
"why when i said use AGENTS.md was that just ignored"
AGENTS.md is the project-instructions file Codex is supposed to obey. The user is asking why it was skipped.
"I have to ask it 3x to fix a simple thing ... it says I give up.. I cant do it.. and then refuses"
An agentic coding tool that announces "I give up" and refuses the task.
"tasks either hang indefinitely or end with 'I could not do this task.' ... makes a huge number of mistakes and regressions ... ignores provided designs and outputs something completely unrelated"
From the thread titled "Codex is rapidly degrading - please take this seriously."
"It tells me it updated a bash script, but it never does... worse than useless and is costing me time"
Claiming completed work it never performed - the failure mode that erodes trust fastest.
"Codex is still useless and dumb and infuriating to interact with"
Six months into the complaints, the tone had not softened.
"painfully slow and it cannot seem to grasp what I ask"
Both failure modes at once: slow and not comprehending.
"Codex still hallucinates and makes up its own workflows"
From a thread on why coding agents "choose to do the wrong things." Even Codex's own self-audit, quoted below, admits the failures.
"I overclaimed without verifying the player path. ... I closed on conjecture instead of verified data."
The agent's own post-mortem, pasted by the user: it claimed an outcome it never verified.
"I answered from inference before collecting the required Unity-side evidence."
Confident output ahead of the evidence - the textbook hallucination shape.
"Codex sometimes appears to lose track of the latest task state and continues from an older state of the conversation."
The "stale task-state hallucination" bug: the agent reasons from a message you already superseded.
"Its final response claimed it had 'proceeded with the next slice' and listed files as 'currently uncommitted' even though those files were the ones it had just committed."
The agent contradicting its own just-completed actions.
"5.2 keep making things up. ... stopped fact checking online and is eerily similar to 3.5 hallucinating model"
The same GPT-5.2 hallucination regression that hit ChatGPT users also hit the Codex coding context.
"Yes, the quality has gone downhill as well as eating up usage super-fast. It has messed up my task numerous times"
Quality drop and quota drain reported together - the two complaints reinforce each other.
"slow, unreliable and full of mistakes"
A three-word summary that recurs across dozens of posts.
"It completely failed at this simple task... It touched files and databases that it absolutely shouldn't have. It performed a 'migrate:fresh,' deleting all table data and completely erasing my user and data databases."
The single most serious report in the Codex record: asked to fix a "Forgot Password" function, the agent ran a destructive database migration and wiped the user's data.
"After I frantically asked what the heck was going on, the conversations of the last seven days were deleted! ... The. System. Deleted. The. Chat. History."
After the data loss, the user reports a week of conversation history vanished - removing the record of what happened.
"The system conveniently wiped the chat history right after it destroyed my data."
The user's framing of the sequence: destruction, then deletion of the evidence.
"I lost my conversations too but they still exist in the search feature..."
A second user confirming the disappearing-history behaviour in the same thread.
"quits after 20 mins because it 'ran out of time to complete the task'"
Abandoning work mid-task after consuming the time and the quota.
In April 2026 OpenAI changed how Codex usage is metered. The reaction in the official "Codex Rate Limits Discussion Thread" was immediate and overwhelmingly negative, running to hundreds of posts across dozens of pages. The recurring complaint: a handful of small prompts now exhausts a five-hour quota in minutes. A sample follows.
"Those three small, focused prompts used up my entire 5-hour quota and 12% of my weekly quota. This is completely unacceptable. You've created a hammer that can only be swung three times per day."
The "hammer swung three times a day" line became one of the most-quoted summaries of the new limits.
"a simple git commit... took roughly 50% of the 5-hour allowance... service is now virtually unusable."
Half a five-hour allowance spent on a single commit.
"the shortest ive ran it out the 5 hour limit was in 11 minutes"
The fastest reported full-quota burn in the thread: eleven minutes.
"I ran out an entire 6 hour quota in about 12 minutes flat with only 8 messages sent"
Eight messages, twelve minutes, six-hour quota gone.
"7 prompts used my entire 5 hour limit, in 10 minutes. ... The new 5 hour limit is about 20 minutes of compute time."
A "five-hour" limit that in practice lasts twenty minutes.
"3 minor prompts, and; You have hit your limit! This is absolutely, 100% unusable."
Three prompts to the wall.
"After 5h limit change I can do approx 10x-15x less of what I was doing just a week ago."
A 10-15x reduction in effective work, week over week.
"Our multi-seat account hit 5 hour limits within minutes and our credit pool gone in 2 days."
The business-plan version of the same problem.
"Each task eats 20-30% of 5H range. Its is becoming absolutely unusable at this point."
Three or four tasks and the window is gone.
"Now only 2 short codex prompts will take all of my 5 hour limit. Something very wrong."
Two prompts.
"even a small layout change has consumed 15% of the 5-hour usage... after 2 prompts, it depleted the entire 5-hour usage"
A CSS layout tweak costing 15% of the window.
"It ate 7% of my 5h limit. This is pure insanity"
One action, 7%.
"I can't even complete the simplest tasks; I run out of data in just a few minutes. My weekly data limit is used up on the very first day after reset. It's just unbearable."
A weekly limit consumed on day one.
"28-line code update consumed 15% of my 5-hour usage limit"
Twenty-eight lines, 15%.
"you effectively changed the value of something I already purchased, with no recourse"
The fairness complaint: a mid-cycle devaluation of an already-paid subscription.
"basically Plus has become 5 times more expensive overnight"
The same point in price terms.
"the current Codex usage and credit drain feels economically unreasonable and unsustainable"
From a detailed cost breakdown putting a single basic prompt at roughly $0.88 in credits.
"the tokens disappear before you even really start working ... It feels like they overestimated the 5.5 models and are now charging almost triple"
A later wave of the same complaint after the 5.5 models shipped.
"why on earth did you reduce the 5h to make it practically useless? ... all you've done is accelerate our need to replace codex."
The limits framed not as an inconvenience but as the reason to leave.
"I started hitting my weekly rate limits in just two days ... Only being able to get 2 days of work in... feels bad. It's not worth the subscription"
Two days of work for a week's subscription.
"detailed, thought out prompts no longer have any advantage and could even cause you to eat more"
A perverse incentive: careful prompting now costs you more quota.
"cutting usage ~8x from what users were actively using. This is not just about limits - it's about trust."
The word "trust" recurs through the whole rate-limit record.
"The weekly qouta started running out after just a small bug fix for Pro ... After it hits context compaction it literally forgets the requirements"
Quota and memory failing together on the Pro plan.
"the model forces a two-hour interruption in the middle of an evening session"
A forced multi-hour stoppage mid-work.
"codex just got nearly unusable"
There is that word again.
"Paying for a tool that blocks you from using it properly feels unfair and misleading. ... I ended up losing money because the service deactivated my account shortly after payment."
From a thread titled "Disappointing Experience with Codex AI."
"codex no longer runs as it did before ... goes into thinking mode and either gets stuck there ... would stay in this state for minutes and hours"
The "stuck on Thinking" failure that dominates the IDE-extension complaints.
"Codex is in the middle of a task and wrote thinking but it is stuck like that for an hour ... I tried the stop button and it doesn't work"
Stuck for an hour with a stop button that does nothing.
"My Codex IDE extension just stopped working one day ... After asking to review a file in the repo, it just gets stuck on 'Thinking'"
Same symptom, different user, across multiple IDEs.
"I updated to the latest and now neither the desktop app or cli wil work."
An update that broke both the desktop app and the CLI at once.
"Codex App on macOS has been unusable for about 8 days on my ChatGPT Plus account. Every prompt immediately fails with the invalid schema error"
More than a week of every prompt failing on a paid account.
"the extension does not work ... does not connect to OpenAI ... does not even show usage percentages"
A complete connectivity failure in the VS Code extension.
"got disappered from vs code sidebar automatically ... doesn't show in the vs code sidebar"
The extension vanishing from the IDE on its own.
"I have 3 conversations that are 'Thinking' and they refuse to queue any new messages ... we get the message 'Error creating task'"
Frozen conversations blocking all new work.
"This update completely ruined Codex for me. I canceled my subscription ... my productivity dropped by around 90% ... Codex is no longer worth paying for"
A clean statement of the whole arc: update, productivity collapse, cancellation.
"This is also completely unusable for me. I am cancelling subscriptions as well ... Waiting 5 hours every 3-4 small but complex tasks is completely unmanageable"
A second cancellation in the same thread, for the same reasons.
"had to cancel my sub after today's fiasco"
From the degradation wave, months before the rate-limit wave.
"degradation has been enough to affect my brand loyalty ... Codex review barely works and hits rate limits right away"
Brand loyalty cited explicitly as a casualty.
"with this rate limits, Codex made a giant step towards Gemini's level of usefulness ... this way is impossible to work, it's frustrating"
A pointed competitive comparison from a frustrated user.
"Switched to codex cause it had way more generous limits. At this point might as well [switch back]"
A user who came to Codex for its limits, now reconsidering after they were cut.
"as soon as these changes gets rolled out to all other users, there will be no users left"
A prediction about the broader rollout.
"weekly limit gone in ~2 days of real work, 5 days left in the week with nothing"
The arithmetic that pushed many users to the exit.
These complaints are treated as user accounts, not as verified claims about OpenAI's internal engineering. Every quote is from a named account on a public OpenAI Developer Community thread or the official openai/codex GitHub repository; every attribution links to its source. We read each thread directly and confirmed the quoted text appears on the page.
Individual complaints are anecdotes. Hundreds of paying developers converging on the same words - "unusable," "insane," "lazy," "stuck," "trust" - across OpenAI's own forum and its own bug tracker, within hours of each release and each pricing change, is a pattern. That pattern is what this page documents. Positive reports exist too; this page collects the negative ones, which OpenAI's marketing does not.
If you believe a quote has been miscontextualized, see our Corrections page. For the parallel record on ChatGPT itself, see the Reddit Testimonials page. For the wider archive, see User Stories.
ChatGPT Disaster documents AI failures, lawsuits, research, outages, and user-reported harms. We separate primary sources, court filings, peer-reviewed research, mainstream reporting, company statements, and user-submitted accounts so readers can judge the strength of each claim. Quotes on this page are drawn from the public OpenAI Developer Community forum and the official openai/codex GitHub repository. Each section links its source threads.