OpenAI Quietly Cuts Your Codex Usage by 50%

0:00$200 a month is what I'm paying, and it isn't for a Netflix subscription or a gym I never go to. It's for an AI coding agent that, for a few glorious weeks, made me feel like the most productive engineer I've ever been in my life. And then one morning I open Codex, look at my usage bar, and it's sitting at 12% 3 days into my week. And I'm sitting there

0:21wondering, did I actually use that much? So I scroll back and check, and I do the math in my head even though I already know the answer won't be a good one. And that's the allowance as it stands today, because this Friday it gets cut in half, which is not what I pictured when I assumed the $200 subscription meant getting the same $200 subscription I signed up for. OpenAI is cutting Codex

0:42usage by 50% this Friday, October 30th. And the way they're doing it is genuinely one of the messiest product stories I have seen from a major AI company, and there have been a few. It comes down to the timing of it, the resets they ran in the weeks leading up to it, and what they chose not to say publicly. So let's go through exactly what's happening, who it actually hurts, and what the real number is, because it is not 50% for everyone. For some people it's 24%, for some people it's zero, and for a specific group of power users, it's closer to 85%. But before we get

1:17into the numbers, I want to give a bit of context, because if you are not a daily Codex user, you might be wondering how we even got here. Codex, the way OpenAI uses the name here, is not just an autocomplete tool. It's a full agentic coding system. You give it a task, like refactor this module, find this bug, or write tests for this function, and it goes off on its own. It reads your code base, writes the code, runs the tests, and iterates, so there's

1:45no babysitting. You just check back in. It's less a co-pilot and more a junior engineer who never sleeps, never complains about your legacy spaghetti code, and actually reads the ticket before starting. And if you have worked with junior engineers, that is not always a given. And the $200 ChatGPT Pro plan, which OpenAI calls Pro 200, gave access to this with what they billed as a 20x allowance, meaning 20 times what

2:10the basic plus subscriber gets. For power users building real things, that was not just a feature. It felt like a professional grade tool at a professional grade price. The first week I seriously leaned into Codex for production work, there was this refactor I'd been putting off for 2 weeks. It was one of those jobs where it's obvious exactly what needs to happen and nobody wants to sit down and do it. I set it up as a Codex task, went to a meeting, came back, and it had done 80% of it correctly, and I sat there for a full minute just staring at the diff. And two

2:43thoughts hit at the same time. And the first was, am I going to have a job in 5 years? And the second was that I genuinely did not care right now, because this was incredible. That's the headspace a lot of us have been in all October, and it's why what lands on October 30th hits the way it does. So, let's get precise about what's changing, because the details really matter here. Starting this Friday, October 30th, existing Pro 200 subscribers, specifically anyone who had an active plan at some point between September 22nd and the morning of September 29th, drop from a 20x

3:17allowance down to 10x, which is half. And alongside that, GPT-6 Pro messages drop from 200 a week to 100. And those are the two concrete changes. Now, OpenAI did announce this, and the announcement was not buried. Thibault Sottiaux, who runs Codex at OpenAI, posted about it on X, and the tweet got 17 million views, and grandfathered subscribers got an email. So, the

3:44information was out there. But the part that genuinely got under my skin is OpenAI's official help page, which is supposed to be the canonical source of truth for what a $200 subscription includes. It says, "New plans get a lower usage allowance to reflect our increasingly efficient models." And that is the entire statement. There are no numbers and nothing like you're going from 20x to 10x or your allowance is being halved, just the word lower. And everything concrete, meaning the actual multiplier, the size of the cut, and the fact that it was specifically half, came from a

4:20tweet and an email. I have worked in product before, so the way these announcements get written is familiar, where someone sat down, drafted that help page, reviewed it, approved it, and published it. And the decision to not include the actual number in official documentation is a choice. Whether it was cynical or just bad communication, I can't say. But when you are a paying customer and have to find out the size of a change to your subscription from a viral tweet rather than the company's own support page, something is off. Because tweets are searchable for a while and then they are not. Since tweets get lost, deleted, and

4:57buried, while documentation is supposed to be what stays. And the documentation, in this case, said nothing at all. Now we get to the part of the story that completely changed how I understood the cut. And it's the piece that most coverage has glossed over. On October 4th, Sutiyo posted something that, at the time, sounded like a fun community

5:18engagement move. He said that over the next 28 days, each day OpenAI would either ship one clear improvement to Codex or ship a full reset. A full reset means everyone's usage meter goes back to zero, so whatever had been burned through that week comes back. And someone in the community did the thing that online communities do and built a tracker, a site called codexresets.com that logs every reset from Sutiyo's posts. And in the first 23 days of that run, October 5th through October 27th, it shows 12 reset days, which is 12 resets on a meter that is supposed to last 7 days. When we put that into

5:56actual math, 23 days is about 3.3 weeks, so the plan itself handed out 3.3 weekly allowances during that window, and the 12 resets handed out up to 12 more. So, a heavy user who drained the meter before every single reset was effectively getting up to 4.7 times the plan allowance during October. Now, and this is important, most people did not drain it every single time. So, 4.7x is a ceiling, not an average. But, we don't

6:26need the ceiling to see the problem. Take someone who is running GPT-6 Astra, the flagship model, at a pace that would normally drain their week in about 2 days. At 20x allowance, they hit the wall on day two. But, in October, with a reset dropping every 1.9 days, they barely ever hit the wall at all. So, they were on Codex almost every day,

6:48draining and refilling over and over. And over the next 4 days, two things happened. This Friday, October 30th, the allowance halves, so that same user goes from lasting 2 days to lasting one, and then on Sunday, November 1st, the 28-day run ends and the near daily resets end with it. Now, that user goes from Codex almost every day to Codex one day a week. And that is not 50%, that's a seventh, and nothing OpenAI published prepared anyone for that specific reality. Because the resets were never part of the plan, they were goodwill gestures, community-building moves, and celebrations, never documented as a feature and never sold as part of the

7:32$200, so nobody has to say anything when they stop. So far, this has leaned into the frustration angle, which I think is warranted. But, I want to do something a lot of the content around this story hasn't really done, and give OpenAI's argument its proper weight, because it is more sophisticated than the outrage makes it look. In his post, Sutskever says the new plan will net out at half the dollar in API spend compared to the old Pro $200 plan, and that's a specific claim. He then argues you'll still get more work done than if you were on the Pro $200 subscription 1 month ago, and

8:06his reasons are concrete. The first is OpenAI's commitment to never bring the 5-hour window back on Pro, which, for anyone who used Codex before they removed it, was a genuinely annoying constraint against burning through too much usage in a short burst. Now, you can dump your entire weekly allowance into one afternoon if your workflow calls for it, and that is a real, meaningful change. And he also points to the fact that OpenAI has been aggressively cutting API prices, where the week before Dev Day, they launched GPT-6 Soul and GPT-6 Luna at half the price of their predecessors. So, the argument he is making is that the

8:46dollars halved and the model prices halved, so you break even or come out ahead. And the thing I had to sit with is that for some models, that is completely true. Which brings us to the actual math. I know, math in a YouTube video, but bear with me, because this is the part that really determines whether this cut matters for your specific situation. Codex bills in credits, and OpenAI publishes a rate card that lists the credits per million tokens for every model it supports. And when we pull up that rate card and the API price list side by side, every single number on the rate card is exactly 25 times the

9:24corresponding API price. That is the conversion factor. It means your allowance is essentially a budget denominated in API dollars spent at list price. Sutskever even said this directly, calling the cut half the dollar in API spend. So, your allowance is a gift card loaded with API dollars, and the gift card got cut in half. But, how far half a gift card takes you depends entirely on what you are buying with it, and different models cost very different amounts. Let's start with GPT-6 Astra, the flagship. It is the simplest case, and unfortunately, the harshest one. Astra's price hasn't moved. It is still $10 per million input tokens and $50 per million output

10:05tokens. So, half the dollars buys exactly half the tokens, and there is no model price drop softening this, and no recent discount to point to. So, if you've been running Astra because Soul wasn't capable enough for your work, this is a clean, straight 50% cut, where every word of the headline is true for

10:23you, and there is no clever work around. And something makes the Astra math even more stark, because Semi Analysis did a month of actual agent work and measured the token distribution. What they found was that 96.6% of the tokens in a typical coding session are cash reads, which is your code base being pulled back in on every turn. The actual code being written, the output you care about, is just 0.3% of the tokens. So, when you are running a big project through Astra, you are mostly paying for your own code to be reread constantly, and now you get half

10:57as many rereads. Now, when we look at GPT-6 6.1 Soul, that is where Soty O's math actually holds up. The everyday Codex model a few months ago was GPT-5.6 Soul, and its cached input cost 40 cents per million tokens. Then GPT-6 Soul launched and cut that to 20 cents, and then GPT-6.1 Soul at Dev Day cut it again to 10 cents, which is a quarter of what it was. And as we saw, 96.6% of the tokens are cash reads, so that single price line drives almost everything about the cost of a typical session. When we run the math without waiting, a million tokens of agent work

11:37that cost about 57 cents on GPT 5.6 sole costs about 19 cents on GPT 6.1 sole. And half the dollars at 19 cents per million buys about 1.5 times the tokens the full dollars bought at 57 cents. Against August, which is the baseline Sotio uses in his post, you are actually ahead on sole with more work for less money. So, the claim holds up. Now, that is a pretty flattering comparison point, but it is not dishonest. If we move the baseline to the week right before the change and compare against the last week of September, when GPT 6 sole was the model, half the dollars buys about 76%

12:17of the tokens. That's a 24% cut, which is still a real cut, but not 50%. It is the same plan and the same change with two different answers, depending entirely on which month it's held against. And 1 month ago is the month that makes it look like a gift. And that's the month Sotio chose. Then we get to GPT 6.1 Tera, the one that changed everything 4 days before the cut. On October 26th, 4 days before October 30th, OpenAI brought back Tera, and the pricing is exactly half of GPT 6.1 sole on every single line. At $1

12:51input, $5 output, and 5 cents cached. Which means on Tera, the half the allowance buys the same number of tokens that the full sole allowance did. So, the cut, in token terms, is zero. And OpenAI launched a model that makes their 50% cut feel like no cut at all 4 days before the 50% cut. I don't know whether to call that convenient or clever, maybe

13:15both, but the timing is something. But the catch, as always, is capability. When we check Artificial Analysis' benchmarks, Tera at high effort scores 42 and sole at high scores 50. And Terra can reach 51 at max effort, but at that point it costs more per task than Soul at medium. So, Terra is genuinely 2x on the tasks it can handle, and whether your specific work fits that profile is something only your own diffs will tell you. So far, we've been talking about tokens though, though. And tokens are actually the wrong way to think about this for most people, because what you really want is finished work, completed

13:52tasks, and shipped diffs. Artificial Analysis publishes cost per task numbers for every model at every effort setting. And when we look at those numbers, the story gets even more interesting. GPT-6 Astra at high effort scores 55 on their benchmark and costs about $3.24 per task. And GPT-6.1 Soul at max effort also scores 55, which is the same benchmark score and the same quality output on these tasks. But Soul gets there for about $1.27 per task, 39 cents on the dollar compared to Astra. So, with the same budget, Soul finishes roughly 2.5 times as many tasks. Now, if we cut the budget in half, Soul at max still finishes

14:36about 1.3 times what the full Astra budget handled. So, if you move off Astra to Soul, you can actually come out ahead of where you were before the cut in terms of completed work. And it gets even wider when we bring Terra in. Astra at medium scores 51 at about $2.73 per task, and Terra at max also scores 51, but at about 96 cents per task. And half the allowance on Terra finishes about 1.4 times what the full Astra allowance handled. So, the argument starts to look genuinely solid. But, and I want to be really honest about this, Astra at max effort scores 57, and no GPT-6.1 model

15:15in Artificial Analysis' current index reaches 57. The gap is two benchmark points, and right now there is no cheaper model to switch to that closes it. And the people who need those top two points are the ones doing the most complex work, the hardest problems, the tasks that push them toward a $200 subscription in the first place. They're exactly the power users OpenAI built this tier for, and they're the ones this cut hurts most with the least recourse. Now, to soften the blow, OpenAI gave grandfathered subscribers a one-time credit grant of 62,500 credits, which has been reported as worth $2,500 in API value, and based on

15:55the 25-to-1 conversion factor, that is exactly right. It sounds decent until we do the full math on what the cut actually takes away. Semi analysis meter the new Pro 200 at about $2,084 a month of GPT-6.1 sold for a typical heavy workload, and the old plan held double that. So, the allowance cut costs

16:17roughly $2,084 a month in usage value. And from October 30th to December 31st, when the grant expires, is about 2 months, which comes to roughly $4,300 of allowance value removed from the plan. The $2,500 grant covers about 58% of that gap, and then it's gone. In practical terms, that's about 13 billion tokens of GPT-6.1 sold at that agent mix, or under 2 billion tokens of Astra, and it is meaningful. It buys time to test Tera, to recalibrate your model defaults, and to figure out what you actually need going forward, but it is a cushion, not a refund. And the expiry date pushes you to make that

16:57decision by December 31st, whether you are ready or not. One subscriber summed it up on X this week in a way I thought was very honest. They said they had 12% of their allowance left with 5 days until renewal, and once that resets and their 62,500 credits were gone, they were shifting my Pro plan to Claude Max, in their words. And that sentiment has come

17:18up more than once in threads this week. So, OpenAI may have handed Anthropic some of their most productive users, and we'll see how that plays out over first quarter. Now, when we look at the shape of the pricing as it stands, it tells us something important. Plus is 1x for $20, Pro 100 is 5x for $100, and Pro 200 is 10x for $200. And when we divide each of those out, every single plan buys

17:44exactly the same allowance per dollar. The old Pro 200 was the only tier where a dollar bought twice as much as anywhere else, and from Friday, that is gone for everyone. The 20x plan did not disappear, it just got repriced as Pro 500 at $500 a month. Except, SemiAnalysis measured it and found it holds only about 21% more Astra usage than the old $200 plan. So, it is 2.5 times the price for roughly a fifth more usage. And its headline feature, Ultra Fast mode, burns the allowance eight times faster than standard, which is not flattering math. What's interesting here is that Anthropic's Claude Max has been

18:23working this way for a while, where Claude Max 20x holds about 10 times what Claude Pro holds over a month. And with this restructure, OpenAI has basically moved to where Anthropic already was, with a similar multiplier structure. The difference is the base, since Claude Pro holds considerably more monthly usage value than ChatGPT Plus does, though the open token cost is higher, so there's

18:46not quite as wide a gap as it looks. None of that makes the cut feel better if you are the person experiencing it. But it does suggest this is not OpenAI getting greedier in isolation. It's the whole industry converging on the same model, where there is a baseline and a multiplier, and the multiplier no longer means a disproportionate deal at any

19:07tier. I want to take a breath here, because there is a bigger pattern behind all of this that's worth naming. What we are seeing with this cut is a fundamental tension in subscription-based AI products. Because when a company charges a flat monthly fee for access to a capability that has a real measurable cost per use, it is always managing a distribution. Since most subscribers don't come anywhere close to the limits, while a small percentage of heavy users hit them constantly. And historically, the heavy users carry a kind of implicit deal, where they generate the use cases, the case studies, and the social proof, and they become the demos, and in

19:47exchange, the pricing leans slightly in their favor. On Pro 200, that deal runs out on Thursday, October 29th. And the way it's ending is technically announced, but practically opaque. Because the real change is not just the allowance number. It's also the end of the 28-day run, on top of a reset habit that never appeared in any documentation, that has been running since September of last year, and that a lot of people have unknowingly built their workflows around. And the part that gets me is not the price change, because prices change, and that is normal for the industry. It's the resets. Nobody can argue that the resets

20:24were sold to you, because they were never on any plan page. But when something runs for months, when a community builds a tracking site for it, and when people structure their work week around it, it starts to feel like part of the product, even if it technically is not. And when it ends with no notice, you don't just lose the resets, you lose the mental model you

20:44had built about how the tool worked. results are fascinating in this context. He offered Pro users 2x the usage in exchange for bringing the 5-hour window back, and most said no, because they preferred keeping the freedom to spend the whole allowance at once over doubling its size. So, what Pro 200 is actually selling, it turns out, is not the number on the label. It's the right to work without constraints on your schedule and blast through a week of Codex on a Saturday afternoon when inspiration hits. And that is a real value, but it is a very different product than the one people thought they were buying when they saw

21:21the 20X. Okay, that's enough analysis. Let me tell you what to actually do with this. If you were grandfathered and haven't touched anything yet, the first thing to understand is that there is nothing to bank. The old allowance does not carry over. So, whatever doesn't get used by Thursday, October 29th, is gone and Friday starts from wherever you are. Second, change your default model right now. If you have been running Astra out of habit because it was the best and you had the budget for it, that default is costing significantly more per task than it needs to. So, make GPT-6.1 Soul your everyday model and reserve Astra

21:58specifically for the tasks where Soul demonstrably fails in a way that affects your output. And that does not mean Astra feels smoother or Astra gives better vibes. It means Soul tried this, it broke, and Astra fixed it. And that is the bar. Third, actually give Terra a week, not a day. I know the benchmark scores are not as high, but benchmarks are someone else's tasks, not yours. So, run your actual workload through Terra and see what breaks. You might find that 70 or 80% of your Codex usage runs perfectly fine on Terra, which means you have effectively doubled your allowance for that work, and the stuff that needs Soul or Astra

22:37will make itself known. Fourth, spend the 62,500 credits with intention. They expire December 31st regardless of whether they have been touched. Use them on the high-stakes, genuinely complex work that actually needs the flagship model, and don't let them drain passively in the background during routine work that Terra could handle. And fifth, most importantly, judge the plan by the first clean week with no resets. Everything before that will was inflated. The first week of November, after the 28-day run ends, is the actual product and what you are paying for. So, evaluate it on that, not on the October experience where the meter reset every other day. And if that

23:17first clean week shows you hitting the wall on real, meaningful work, meaning work that can't be simplified or shifted to a cheaper model, then compare Claude Max 20X against Pro 500 before deciding anything. They are different products on different models, but they're in the same price neighborhood, and the per task math might land very differently than expected. I started this video with that 12% moment, that feeling of opening a tool you have built your week around and realizing the ground is shifting without anyone really telling you. And I want to end honestly, because that feeling is valid, and you are not overreacting. Paying $200 a month for a

23:54professional tool means making a real commitment to it, and that deserves clear communication when the tool changes. What frustrates me is not that OpenAI made a business decision. Businesses make business decisions, where prices change, models evolve, and allowances get restructured. And all of that is normal. But what frustrates me is the gap between how much is actually changing and how little of it is clearly documented anywhere the average

24:19professional user would think to look. Because that 12% moment I described is not really about the 12%. It is about the fact that I didn't see it coming, that I had made plans and commitments and workflows around the tool, and the terms are changing underneath them. And the help page that is supposed to prevent exactly that feeling just says lower. For what it's worth, I am staying on Pro 200 for now, shifting to Soul as my default, putting Tera through its paces this week, and being a lot more deliberate about when I actually reach for Astra. Because maybe by December I'll have a clear picture of what I

24:52actually need. Maybe I'll upgrade or move elsewhere, but I am going to let the math and the diffs make that call, not the frustration. And that's my recommendation for all of you, too. So, don't make any major decisions this week. Run the first clean week and do the model math for your own workload. And come back and tell me in the comments how it went because I genuinely want to know whether the Soul and Terra shift makes up the gap for most people or whether the Astro-dependent crowd is

25:19as stuck as the number suggest. If this breakdown was useful, subscribe because I am going deep on AI tooling and the real economics behind these products every week. This kind of pricing complexity is only going to get more intense from here, so someone has to do the math. See you in the next one.