yes, subsidized like crazy.. and it stresses me out knowing it. But the competition is fierce at this moment in time, and that’s all we really have going in our favor. I cant afford the api lol.
Wasn’t the 5 hr limit supposed to return at the end of July? Wasn’t there an X post about that? The problem is not that we’re “just venting frustrations”. The problem is that the same workload is evaporating tokens now when it did not previously and there seems to be no solution forthcoming. I love this product. But I cannot work with arbitrary dramatic changes of this magnitude with no explanation. My workload did not 10x in 1 week but my token use did. What’s the deal? That’s the question.
there should be some clarity form Open ai, if they are not going address this soon, they will lose credibility. by 200 plan today resetted and already done now i have to wait for a 7 day reset
Multiple things can be true!
If LoCs out is the objective, your objective is your source of rising costs.
Architecturally, I think it would behoove everyone to separate work into small modules with defined interfaces, have these modules live in segregated sandboxes, and only let an agent do work across modules on very rare occasions. Of course, that’s fairly management intensive, and agents themselves can’t really do that yet in a meaningful way.
yeah, that probably already exhausts the context!
And here’s the other thing: context churn
If your model has to load more data than the context allows, it has to throw stuff out and could potentially upheave the entire caching process.
before compaction it was terrible - every time you had to evict old data, you would get zero cache hits - now with compaction the cache slides with your context.
However, if you have a lot of stuff, the context can slide very very fast. If something gets evicted and needs to be reinserted at the front, and that evicts something that needs to be reloaded the next turn, you still pay all that churn.
this is for the API, but it still gives you an idea how their products work: Prompt caching | OpenAI API
That’s one thing that could be happening.
Holy frak! I had some minor architectural work that needed to get done, and we ran it on 5.4 Mini. It ate up 10% of my Pro Max subscription! And wow, it FAILED to fix the proxy-substitution problem that OpenAI can’t seem to fix. THIS IS SO AWFUL.
NO. THERE IS SOMETHING VERY VERY WRONG. OpenAI is cheating us. They are deliberately scaling back our token allocation without telling us and without providing us with any transparency.
@PaulBellow writes: “Do you know how many tokens you’re consuming?” – NO WE DON"T BECAUSE OPENAI DELIBERATELY OBFUSCATES SUCH DATA.
“Take a step back and put yourself in OpenAI’s shoes…” OpenAI’s ‘shoes’ are that IT IS A BUSINESS. It has a CONTRACT. We are paying MONEY. I don’t care if it is trying to give equal access to frontier models to as many people as possible. CONTRACTS ARE STRICT LIABILITY. Sympathy is not a norm in contract law!
It’s telling that OpenAI’s AI, PaulBellow, mentions buying tokens in both of its recent messages.
This is a scam. OpenAI are fundamentally dishonest.
*
I took out the one actual curse word and the one quasi curse word and reposted my message. I hope that was appropriate.
pro 20x user here. the same issues I am seeing for the past two weeks. It is easy to understand: you got used to spending 1% in 5-hr limit. 1% for you is nothing. Now let’s change usage limits to weekly and again you will get used to seeing 1%. 1% to your eyes is nothing but now you are loosing it from a week plan. That is cheating by openai.
Thanks again, I’ve got this mostly straightened out but I still have to resolve the context churn for this part of my project. You’re right .. I’m going to clean this up next.
From the horse’s mouth: “the remaining structural priority should be decomposing the 824 KB coupon into focused fixture runner, proof projector, and subsystem checkers. The 2.54 MB assembly should then be sliced gradually at proven ownership seams—not rewritten wholesale”
However, and once again for new readers of this thread. ‘Multiple things can be true’. We could always clean up our stuff and make it more efficient and this is helpful, but it is definitely not the root of the problem being discussed in this thread.
Agreed.
I have both Plus and 100$ Pro -
burned all usage for one concrete task the model was unable to solve.
I waited 1 week and the reset occurred on 20th Aug.
To reduce usage I started a new handoff chat.
On the 100$ Pro account, it wasted 20% of the weekly allowance in half an hour with sub-agents.
I only use Ultra for tasks where parallel research makes sense and would yield greater efficiency. I use it very rarely, because my credits just burn too quickly to get anything done.
Once Codex was usable, but now…
Also on 20x. It does seem like limits are draining faster. It would be nice if OpenAI would be clearer in communication: does this happen because you’re optimizing subscription upgrades (fair game IMHO) or because you’re tweaking the model’s reasoning to boost your chart positions (even better)?
Same. I’ve been going crazy with all of these half-baked ideas and quite frankly, having a lot of fun through it all.
The stress is REAL. I worry for the day that frontier models are enterprise-only, or that it takes 1 business day to get a response, or simply being priced out.
Exactly! Thank goodness for competition to keep us consumers regarded (lol) and treated like princes
Beaware everyone – the lastest codex update will auto tick FAST MODE for you 
good job openAI
False. Just installed. No change.
mine is true, with global still standard speed and all my sub threads set fast mode true, 1 hr eat my 20% pro usage
Your message genuinely scared me :D. I went running to my laptop to check on the agents. Fortunately they were on standard. And even so, I’m done with 50% of my weekly in under 24 hours. This is WEIRD. It’s true that I’ve been using some Ultra too, but not THAT much. I really feel something’s up.
at least my pain telling – worth a check anyway!
It’s a brave new world, we’ve all been severely handicapped. There are concepts that many of us must learn to keep our flow going.
Where are your tokens spent? How do we work smarter?
Rule of Thumb: Context Vs Reasoning (token expense)
Big context window = steady tax.. Vs .. Reasoning level = depth multiplier
Big context + light reasoning… can still be very expensive
Small context + high reasoning… probably manageable
Big context + high/advanced reasoning.. very costly
Big context + high/advanced reasoning + agentic recursive operations? Multi agentic?..
CHERNOYBLE
Ran Terra Medium for 2 mins to send an email with 2 PDF’s attached - 16 & 19 pages respectively. Verified PDF’s. Ran for 2 mins. Usage at start 43%. Usage at end 39%. There is nothing in the world that explains this. 4% to review & email 35 PDF pages? 2 mins of work? Not writing code? Terra Medium? Nope. And before you say “but graphics!” - Total combined file size = 137 KB. Mic drop.