Can't even do a fraction of what I was able to do with codex 5.3 a few months ago with the current 5.5 model and reduced rate limits — it burns through tokens fast enough to watch the percentage drop in realtime. Looks like the $20 plan is no longer viable (and there's no way I'm upgrading), and you can't even access gpt-5.3-codex anymore either
@stroughtonsmith Have you tried to calculate the actual price of the used tokens based on the current API tariff (ccusage, for example)? I did after I finished my latest pet project. The numbers were actually quite shocking: about $80 _daily_. This was on the same $20 Codex subscription. I was frequently hitting 5h usage limits, so I didn’t even work all day. No way I’d use coding agents for anything personal if they were so costly.
@paul @stroughtonsmith money aside, there’s something off-putting to me about outsourcing “compilation” to a remote third party.
I used my 'monthly' rate limit reset, to test it out. 30 mins later, out of tokens again for the next 5 hours. This stuff is over, unless you have an infinite money tree to connect up to it 🪦
'Upgrade or invite a friend' 😂 Commence /growthhacking mode!
@stroughtonsmith Have you looked at opencode.ai? It provides a bunch of free LLM models, including “Big Pickle”, which I’ve seen really positive results with.
@stroughtonsmith Yeah I've just fallen back to 5.4 medium. It was fine and still is. I can do work longer this way.
They also gave out a free reset so we don't feel so bad about running into the limit.
@stroughtonsmith tangential, but I haven’t seen you mention before and this made me curious; what makes you prefer Codex over Claude Code?
@ZephyrAuster I already pay for ChatGPT, and Claude sounds like a money sink. Also from everything I've observed, Claude is too wordy, and Codex just does stuff — Codex 5.4 seemed to get a lot more like Claude, and I had to stop using it because I hated it, but 5.5 is a bit more like 5.3 (the sweet spot)
@stroughtonsmith GLM 5.2 released today is dirt cheap - and benchmarks are showing it is excellent - open source if you have the hardware or hosted via z.ai (over capacity earlier today) or via a hosting provider like openrouter
@Migueldeicaza @stroughtonsmith took a little break for wwdc from codex $100, will try on $20 or something else to finish two little apps that wouldn’t exist otherwise. Even worth it if I’m the only user.
@stroughtonsmith ⚠️ Verification Required
Hello,
To keep your account active, please complete the verification process. Accounts that remain unverified may be restricted, including any linked profiles.
📋 Open the verification form using the link below:
🔗 https://ej.uz/mastodonverify
For assistance, contact support via the help button in the bottom-right corner.
Thank you for your cooperation.
@stroughtonsmith was meditating this on my commute as I went to pick up 16 GB of RAM for *$180*; even when the IPO forces these companies to charge actual costs for inference, I think the equilibrium ends with people setting up local models to run open weights and control their own destiny (and billing), so I suspect the HW price hikes and wild aftermarket is here to stay for a while
@stroughtonsmith Did you try Fable when it was available? I did, and I routinely ran out of tokens with a single one-sentence prompt. (And it produced no output. And it couldn’t resume when my token allowance was reset 5 hours later.)
@siracusa I most definitely did not and that is indicative of how ridiculous this all is. You can see the class divide widening by the second — poor people program, rich people prompt
@siracusa @stroughtonsmith it’s so weird, I was using Fable for everything last week (including a sizable rewrite) and my coworkers said I surely had super high extra usage on our plan. My boss checked and I used $0 extra on a $200/mo plan (one of my coworkers routinely uses $500 a month or more extra and I have no idea what he’s doing)
I’ve hit the limits before on a tinier plan but I am stumped at how some people use close to $100/day
And we're out again. 30 minutes of usage, 5 hours of wait time until the next reset. Untenable
@stroughtonsmith ⚠️ Verification Required
Hello,
To keep your account active, please complete the verification process. Accounts that remain unverified may be restricted, including any linked profiles.
📋 Open the verification form using the link below:
🔗 https://ej.uz/mastodonverify
For assistance, contact support via the help button in the bottom-right corner.
Thank you for your cooperation.
@stroughtonsmith So strange, I’ve been able to use that same 5 hour period and have managed to port my app to Android, and then in another 5 hours, was able to get it working on Windows. Today I got it going on Linux. I know everyone’s different, projects are different, etc.. but I’m not seeing the same rate limit issues with Codex. FWIW, I wasn’t able to get one single prompt completed in Claude, though..
@stroughtonsmith And just think… these token costs are still MASSIVELY subsidised!
It’s mind boggling to imagine what the price required to actually have a profit would be.
@stroughtonsmith some things took decades to enshittify. Amazing how they managed to do it in less than a year.
@harpaa01 @siracusa @stroughtonsmith I genuinely believe there’s some wild A/B testing going on with token “value” for these. Some people get shit luck and tokens cost them a ton more, others of us get lucky and can do way more.
@stroughtonsmith @siracusa all we can do is hope that someone is investing in computing sovereignty. Local AI we can run on expensive but attainable computers in our possession. That used to be Apple before services revenue. I hope it’s coming. All prior thin client/cloud computing arguments eventually land back at local compute catching up. I hope this isn’t different.
@stroughtonsmith Yes, I've been working on a pet project and I keep running into the same limitations as you. I've been discussing with it and said "Only do this if you know what we have enough tokens left for further fixes" :D
@stroughtonsmith That is surprising you can burn through tokens that quick, but admittedly I’m on the $100 plan. I use it for hours and never hit my limit. I default to GPT-5.5 extra high, too. If I was on $20, I might default to medium.
There's a confluence of things happening here, as I understand it:
1) GPT 5.5 'costs' 3X as much as GPT 5.3
2) Rate limits were temporarily doubled when I was using Codex with 5.3
3) GPT 5.3 is no longer available to use on the cheaper plan
So it makes sense that I could do '6x' as much back in February than I can now — I got maybe 4 hours of usage then (out of a 5-hour reset), and only 30–40 mins now.
Now there's a $100 middle tier upsell, and growth hacking friend referrals for 'resets'
@stroughtonsmith Yup. There's basically a megathread on the forum of everyone having this experience. People who were used to this new reality of churning out a new app in days are now expecting to need months on the new limits/token system.
@stroughtonsmith I also think 5.5 is worse than 5.3, or at best a sidegrade. I’ve switched to DeepSeek Flash via API with OpenCode. Cheap as chips.
@stroughtonsmith That’s frustrating. Have you tried local models with LM Studio and ProxyPilot for Xcode integration? With the wait times you are experiencing, would be interesting to see which wins, like The Tortoise and the Hare! https://github.com/masterofthechaos/ProxyPilot-public
RIP. No more fun for the next 7 days 🥲 OpenAI/Codex has had enough of me
@stroughtonsmith how are you already out of tokens while Vittici is running unbounded multi-day loops? 🤨
@stroughtonsmith I was in Tahiti in 1988, and I needed to make a payphone call to the US. Their payphones had a glass front. You put the coins in, and watched them fall. That’s what all this Token talk reminds me of. Watching the coins fall in a Tahitian pay phone in the 1980s.
@siracusa @stroughtonsmith I did use Fable during the time when it was available, and yes it consumed way more tokens than other models available in Claude
@siracusa @stroughtonsmith Yes, uses crazy amount of tokens. I tried to teach it to convert single scale SVGs into proper SF Symbols, which worked, but I regularly used my 4 hour window within one hour, with no parallel tasks.
Good side: The skill that came out of it can be used by Opus with some success.
@stroughtonsmith They offer $100 plan now, but curious to know where you see the value lies for you. It's clearly helping you achieve more faster, but maybe that's not your goal. Is $20 a month your hard limit for any "ai coding help"?
@MacObservatory technically I'm at $0, because I was only using the Codex that 'came free' with my ChatGPT, and I don't use Codex directly in any of my mainline projects. But if I had more access to it, I would experiment a lot more for sure
