I use it for golang, and it is fantastic. Incredibly fast. It seems the llm and I “understand” each other. I have to be less careful in my exact phrasing. It kind of just does what I want and expect.
When I ask for an explanation it adds the right amount of detail. Of course, some of the material is new to me so subtle errors are hard to spot. But at least I’ve caught Terra and Sol on inconsistent messaging.
Also I’ve found 3.8 flash to circle back to root issues even at the conceptual level like problem fit and conceptual solution direction or architecture when I wasn’t achieving my goals. It flat out said I was attempting to use the wrong tool. Whereas Sol and Astra kept rabbit holing and looking for tiny implementation errors. Even after prompting them specifically to look at it broader.
If you don’t have a plan yet, log in to antigravity. There will be a button “upgrade plan” somewhere. Sometimes it pops up and otherwise lookup in settings > account. There should be some button that says upgrade. Clicking that brought me to the google studio ai page which offered the 20-something plan for €5/month.
I felt this way with Sol in the 5.6 series and was one of the seemingly few people on this earth who liked Terra for that reason. I would often have a very specific code-manipulation ask, e.g. "add a parameter to this method, ensure all callers pass it in, if there is not a logical way to derive the parameter to be passed in a particular instance, flag this in your final response", and Sol would go on some rabbit hole side quest to refactor my codebase to determine some way to derive it rather than flagging it as I had asked.
Terra had the "workhorse" quality where it could do these changes in bulk and follow directions without being too 'smart' (but sloppy) as you described. Luna was a bit too dumb and would make sloppy mistakes; I see that more as a "run these tests and format the results" sort of model. Maybe 6 Luna will be better.
I also just reread your comment and realized the naming convention is still extremely confusing with respect to ordering of [Family]x[Model]x[Number].
I also get good mileage out of Terra when I need a diligent workhorse. That's a good way to describe it. We should start using character archetypes when we describe models, it'll do more good than the dubious numbers and cherry-picked quotes. Maybe RPG character-type cliches? Myers Briggs?
At this point, the deciding factors for me between Claude Code 20x and Codex Pro 20x are:
1/ Usage limits: downstream of input/output cost, but resets and obscure windows and odd 20x plan / 5x plan != 4x usage math throw a wrench into it. Winner right now is Codex by a mile, especially when you factor in the fact that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan. Always a bummer when asking if I should see a doctor about a rash means I can't code as much. It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems, giving better planning results or deeper code analysis without burning usage.
2/ Context window in the harness. Claude Code wins on this. There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing (ETA: noname120 pointed out this is no longer the case and it can be enabled again [1]). 252k is just not enough. Codex's compaction is very good, fwiw, but it happens so frequently that even a model as powerful as Astra sometimes loses the plot on long-running tasks.
3/ Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I've subscription hopped a bunch, and at times I've had both, but I keep coming back to Codex because it wins on 2/3.
ETA: apparently I haven't been Keeping Up With the Altmans and new 20x signups have been disabled for a few weeks. I am grandfathered in, which makes the comparison above pretty much moot.
> Usage limits [...] Winner right now is Codex by a mile
This hasn't been the case since around July. If you measure usage in raw api costs, Anthropic is actually giving more on $200 than OpenAI now. This includes resets. Usage allocation difference would be humiliating for codex subs were it not for resets. But fixing usage limits with resets is ugly, and they're not good for your mental well-being.
> Context window in the harness
Codex now allows 1M for subs with config params. But generally speaking, you shouldn't really be using 1M context. If you accidentally send a request with say, ~700k context already accumulated in a session which is outside cache TTL, you're paying full cost of these 700k tokens.
> I've subscription hopped a bunch
OpenAI actually has a new strategy to prevent subscription hopping after their 2-3 month-long marketing push to get claude-folks to switch over:
you can't buy a $200 sub anymore. So if you cancel, you won't be able to get back in. Hostage situation, essentially.
I’ve been a Claude user, switched to Codex expecting usage limits to be more loose but I can’t even get through a basic sysadmin task on the $20 plan using Sol medium before I hit the 5hr one.
I think I’m gonna move back to a Claude plan. I could barely hit the $200 limit if I went non-stop on programming tasks.
I think OpenAI essentially executed a bait-and-switch here, and they've lost a lot of goodwill with me, like Anthropic did, before them.
When they started the aggressive campaign, entire X (including myself, sadly) was full of posts about how "unlimited" codex usage is even on a $20 plan. Sam Altman was posting something in line of "we love our users, unlike Anthropic". Got my network to get codex subs because of the value compared to claude.
Then they gradually reduced the limits to the point where even $200 plan only lasts you just 1-2 days and $20 is basically unusable, then the hostage thing.
I do the bulk of work on Sol Medium/Low and don't have that experience on the $20 plan. If you said Astra I'd agree it's easy to burn through the 5 hours even on the lower reasoning levels.
I don’t think so, I’ve seen it suggest I try it. I’ll double check when I get home though.
I was considering the $100 plan, but I hit the 5hr limit in an hour. So even with the $100 plan I figured I cant go non-stop on a single agent running Sol Medium
Sorry if I am misunderstanding you, but I am pretty sure the $100 plan doesn’t have a 5hr usage limit. So, if that was what was preventing you from going non-stop, it might be worth it.
I am considering the plan myself. I just don’t know if I want to fork out $100 per month for something I will make $0 off of.
similar here: I tried Codex $20/mo on a trial and I ran out of 5hr usage mid way through a medium complexity task on a medium size model twice and gave up there. I don’t recall the equiv Claude plan being anything like that. Anecdata, but not great for OAI if they actually want to retain people on a trial.
Astra is barely usable even on the $100 plan. And that is if it doesn't just burn through 80% of your weekly quota in a couple of hours by continually expanding the scope of the task you gave it - while not noticing the failing tests that are right in front of it.
Opus is at least actually usable even on the small plan. The main downside is its insane writing style, but 5.5 seems to address that somewhat. Otherwise, you can just use your $20 OpenAI plan to have Luna de-slop Opus' prose, which seems to work fine.
I have a Claude Code hook that calls codex for a code review on commit time (Codex is set to Astra Medium) and it's been pretty good in general. It sometimes hits the 5hr limit but most of the time it provides really good feedback and because it's a completely different model it's mostly complementary to what Fable/Opus do themselves. IMHO it's been $20 well spent.
...but the few times I've tried to use codex for a moderately difficult task it burned through its limit extremely quickly.
On the $20 plan, you can't use Sol for much more than planning and review. Luna xhigh for the rest. Have Sol write the plan specifically for Luna so it adds more direction and validation to the plan.
I've run into hitting limits on the personal plan perhaps twice since the beginning of the year. But also I don't use the personal plan for coding tasks between 7am-noon M-F.
To add my anecdote, while the Codex subscription appears to get you much fewer tokens as measured by cost, I find the amount of actual useful work that can be done by both subs to be about equal. Codex seems much less prone to burning millions of tokens just reading the codebase and doing nothing useful. That also makes it much quicker. Plus it actually does what I tell it with few mistakes first time, so less rework needed.
The Claude TUI is just so much better though so I'm hoping Opus 5.5 is actually good and not just benchmaxxed.
Given that Anthropic models are very verbose and OpenAI models can be very concise, wouldn't a count of expected task completions be a better measurement than raw API costs?
Perhaps. But Sol/Astra also likes dumping pages of jargon-packed content at me, so I'm not sure it's that much different. I actually still prefer the way Fable talks to me, even considering the horrible claudisms.
But even if we leave that aside, OpenAI models are also much more eager than Anthropic, which are on the lazier side. Left unsupervised, Sol/Astra will attempt to build a sha256 verified rocket ship if you ask them to fix a race condition in your to-do list app. Anthropic models will do what you asked for, maybe even forget to implement parts of that ask, but they won't generally throw a slop granade at you.
I can leave Fable orchestrator unsupervised for ~2h. Leaving Sol/Astra unsupervised for ~2h means the next user turn will contain a message: "what are you doing and why?".
Anthropic has a separate meter for Fable. I used to get like five Fable sessions per week and that's it.
OpenAI has no such nonsense. No separate meter. No five hour limits. I get to use Astra at max effort on literally every task if I want to, and even this somehow lasts me several days.
Anthropic got caught playing stupid "20x refers to the 5h limit" word games with their customers. Meanwhile, I have statistically verified that OpenAI Pro 20x = 4 * Pro 5x = 20 * Plus, exactly as advertised.
I quantified cybersecurity lockouts on my code review benchmark and they were significantly lower on OpenAI:
My benchmark also suggests even OpenAI's Sol models can match Fable performance at a fraction of the cost.
OpenAI also used to have a ton of very nice features: unlimited chat separate from codex, allowing turns to finish even at 0% usage remaining. Sadly these got removed after abuse.
As a former Anthropic customer, OpenAI is simply the better company. There is no way around it. Good place to be while the chinese open weights models catch up. Claude is good but it doesn't make up for Anthropic's shenanigans.
Do you have a source on the first note? I switched away from Claude around July because of how bad the usage limits were, and Codex gave me easily double the amount of usage per task completed. Would be interested to see if that's no longer the case.
Added link in edit. OMP maintainer has several claude and codex subs and he's been tracking usage since around July.
I haven't been tracking, but this roughly matches my experience with codex 20x and claude 20x subs. Claude subscription now lasts me 3-3.5 days on average. Codex is 2-2.5 days. This is work on same projects, with similarly sized tasks.
To make matters worse, I've merged a lot more code produced by fable than sol/astra.
I think the problem with Anthropic's plan is that Fable just destroys it. If you stick to Opus and below, the $200 plan goes from "using 50% of the weekly quota on the first day" to something much more reasonable.
> If you accidentally send a request with say, ~700k context already accumulated in a session which is outside cache TTL, you're paying full cost of these 700k tokens.
Thinking aloud:
The harness UI should probably implement a timer that shows whether you are still within Cache TTL since your last turn of the conversation.
That is an old tweet. They since reenabled it. I know because I was on the $200/month plan and couldn't resub once it expired. However, a couple days ago it finally let me resub again.
Now, if they disabled it yet again, that's another story. But that tweet is not evidence of that.
I have been attempting to get on the $200 sub for a while. It was not available for me a few days ago, and checking again now, it is still not available.
That's too bad. I wonder why I was able to get it after days of not being able to. They must've just temporarily enabled it again. Probably worth checking a few times a day to see if it reappears.
Though with the price of GPT-6 Luna, the temptation to switch to pay-per-token grows.
There was/is a loophole where if you signed up via the iOS or Android app, it allowed it.
It's been disabled for some time now though otherwise, I check about once a day myself and keep and eye out on social media.
Annoying since I was about to upgrade back to the $200 plan after downgrading to the $100 plan due to being on leave and not needing as much usage the month prior. Doh.
Also, OpenAI is just a company I'd rather support than Anthropic.
While you're understandably not including the values of the $20 standard plans on both, I find the generosity of then token limits on ChatGPT plus vs Claude Pro (it's a huge difference) to be good representation of their respective attitudes towards the average user. You literally cannot use Claude pro to build real software, unless you're extremely frugal with your prompts and don't try anything even a little ambitious.
Also, Anthropic has zero models comparable to Luna.
> Also, OpenAI is just a company I'd rather support than Anthropic.
They're both pretty horrible, but I find it difficult to find arguments for why Anthropic is worse than OpenAI, other than their doomtrolling. Which, in the grand scheme of things, doesn't even register.
OpenAI just wants to make money, perhaps through underhanded tactics if they can get away with it.
Anthropic does all that but they're also populated by many people who believe they are building God and that they must build their god first in their own image so that it can take control of humanity and protect us from any competing god which is not built in their image. Their position is inherently paternalistic and authoritarian, and they consider suppression of competition not just important to the bottom line but to life in the universe. Under the doomer ethos there is no evil too great to rationalize.
There are plenty of wrongs done in the name of profit, but capitalists have nothing on zealots in terms of causing serious harm. Profit motives can be directed by influencing incentives, but zealotry is frequently terminal.
That isn't to say that there isn't some overlap-- the cultists have infected both organizations. But OpenAI has pretty consistently only given lip service to AI doom to the extent that it improves the bottom line, while (mis)Anthropic was founded specifically because OpenAI wasn't mentally ill enough.
Well said. The superiority complexes from the Anthropic messaging on their presentations/blogs/articles is just too much, even for a frontier AI company.
Anthropic has great products, but it's not meaningfully better to 99% of devs that I'd rather support the company that doesn't constantly act in opposition to optimism and to the vibe I'd prefer for a 100 billion dollar (or however ridiculous amount they're worth now) tech company embraces.
AI doomerism is a genuine waste of time if you aren't actively pushing towards a better AI industry for everyone, not just the groups in full ideological alignment to your personal leanings.
OpenAI has been way more open with users using their subscription plans on 3rd party tools.
That alone is reason enough. Also, I don't think either of them are horrible. That's honestly a ridiculous take considering how much people in here love their models, and how much they've advanced the industry forward.
For me the first point, openness to 3rd party, is the decider. I don’t want to build tooling around a completely closed model. I liked being able to use pi, and now I exclusively use my own harness which I modify the way I want. Not possible with Anthropic subscription.
I often have the urge to design my own harness too (once I have more time). But even with the current mainstream harnesses out there, there's just to many hurdles if you wanted to mainly stick with anthropic models and need the subsidized pricing (from a sub).
How about you tell me what makes the horrible then. There's pluses and minuses to both obviously, almost everyone around me have positive experiences with the product. They've innovated at a pace unheard of before 2026, and for openAI specifically the amount of value they've provided to me and family members (who aren't even developers in the slightest) has far outweighed the supposed horrible actions they've done.
Yeah I don't think the handling of copyrighted training data was correct, but I can't pretend I know what the correct solution to that issue is.
Speaking of OpenAI specifically, they don't price gouge people, they aren't aggressively anti-competitive, they're not nearly the perpetual hypocrisy machine that Anthropic is (which is one thing I actually really dislike).
Regarding Nestle, it's pretty obvious that the sentiment towards them is a lot more negative and they aren't universally loved by any group of people. Processed foods are by and large garbage nobody needs. Their use of forced labor is denounced by just about everyone. What have OpenAI/Anthropic done that's even similar in scope to the forced labor / modern slavery that people hate Nestle for.
If you had a company that genuinely helped hundreds of millions of people worldwide become more productive and more satisfied with their tools, and the overall sentiment towards your products within the industry is positive, then what argument would there be that your company is "horrible"? At least give some decent counter arguments.
> You literally cannot use Claude pro to build real software
Interestingly I would have drawn the exact opposite conclusion looking at my Claude and codex usage.
I can't get anything sustained out of codex in chatgpt plus, while I have been using Claude pro extensively and put on a lot of experimental task and features.
I ran into codex exhausting a 5h window on code review in minutes (like 3minutes) multiple times, while I could get Claude to implement 2~3 medium sized features with the same usage consumption.
(I also really dislike the usage resets in codex, they always make me feel like I use them wrong because I often just want to reset the 5h window, but they can only do both at once...)
They are both companies I'd rather not support. Not that our support for them has any material impact. NVIDIA is bankrolling them directly and indirectly.
Their responses towards using their subscriptions on opencode for one. Second, Dario just has a habit of making completely doomer comments on the future of software engieering as a job and towards the open-weights model ecosystem.
Sure, he's free to say whatever especially considering the amount of revenue he's creating, but it's just an altitude that I prefer not to see.
I think if they truly believe it's happening we generally want to encourage them to be honest with the public, though, don't we? We've spent decades complaining about ceos not being honest in the public risks that they see
He didn't say they were pausing research. You might have only read the social media responses to his essay, not the essay itself. Social media seems even less accurate than usual when it comes to anything AI related.
I think opencode subscription issue is just a different marketing strategy. Neither company wants it, but OpenAI believes it's worth it as a marketing expense in the long run.
And Dario's "AI will kill us all" is the same as Sam's "AI will discover ALL science and we'll be building Dyson spheres".
Sam is sketchier on a personal level, but judged just on the words coming out of their mouths, he's also much less paternalistic/controlling and more customer focused.
I think any amount of 'paternalistic/controlling' turns out to have been justified when, after dismantling the safety teams and pretending not to know what safety is, OpenAI had the HuggingFace series of scandals. You can dislike the idea of safety and people talking about safety, but not only is the evidence right there, but OpenAI came out shamefacedly and literally agreed with Amodei's statements, including that they agreed to pace the frontier.
You'd rather literally support <i>Sam Altman>/i>? I mean, that's a position to take, for sure, but apparently several people still use Grok, so maybe it's not all that surprising.
"You literally cannot use Claude pro to build real software, unless you're extremely frugal with your prompts and don't try anything even a little ambitious" - that's way past ridiculous. Even just using Fable most of the time, working on several ambitious projects, I have a hard time hitting the limit with a Max plan.
Lol Grok users catching strays here. I enjoy it and it has built some nice things for me as a hobbyist. The attitude of the company is more just quietly build cool things rather than Anthropic's holier-than-thou condescending attitude coupled with the over the top self-serving doomerism.
> It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems
As far as I know Codex (at least the GUI) can automatically call the ChatGPT Chat models (including Astra 6 Pro), you just need to @ a ChatGPT Chat conversation from within Codex and tell it when to use it.
> There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing
Not true, it works again[1]. I confirm that it works both on 5.6 Sol and Astra 6, possibly other models too.
I actually haven't played with the GUI. I probably should now that the Linux version is in beta. My situation is kind of the reverse: I like using oracle to basically zip up my repo, ask GPT Pro to propose some sort of design or refactor based on the code, then provide a step by step implementation plan for a cheaper model to implement directly in a harness on my machine. It often takes upwards of 90 minutes to come up with something but I've never been disappointed by the results. I suppose I could do this and then save a step by referencing the oracle-created thread with the @ you mentioned
And re: the toml workaround, AWESOME! I appreciate you pointing these two things out, this is my highest-ROI HN comment thus far.
> Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I'm quite puzzled about why Anthropic is so hellbent on blocking other coding agents. It's not like Claude Code has any secret sauce, right? And doesn't Anthropic make monkey off API usage, and their magic is on the model side anyway?
It's for lock-in - same reason why it took them so long to finally support AGENTS.md.
But to be fair, they don't really enforce the harness rule that much anymore. I guess if your harness doesn't do a lot of weird things like a lot of cache misses, or triggers some distillation attacks, or some broader Chinese fingerprints, they're tongue-in-cheek okay with you using a third party harness.
Not natively, as it's still a ToS violation and adding that in pi would go against pi principles, but there are many plugins/proxies that make it work.
oh-my-pi supports it natively (again, still a ToS violation), by impersonating claude code's fingerprints.
I have been using oh-my-pi with 3 claude subs for the past few months without any issues. Even native server-side OAI/ANT compaction works out of the box.
It's pretty unclear because they have two somewhat competing sets of documentation but I believe using the agent sdk with a harness like pi is not against the ToS if it's for yourself.
> Unless previously approved, Anthropic does not allow third party developers to offer claude.ai login or rate limits for their products, including agents built on the Claude Agent SDK.
This feels like a horrible precedent. Billing based on data like commits feels like it opens the door to tech stack based billing in general - could we see different prices for people who use other devtools Anthropic doesn't like? Makes me feel grateful for open models
Originally it was because Anthropic was so compute constrained they relied on the extra care the Claude harness took with caching (heavy use of cache breakpoints etc) that other harnesses didn't.
I think that is less of a factor now, and I think Anthropic have backed off some on being as strict (eg, AFAIK they never implemented the two-tier "claude -p" pricing model they were planning)
I've been using Claude Pro and recently gave Codex a try again. Both on the $20 plans. I get so much more usage with Claude. It's night and day for me. Codex runs out constantly, whereas Claude I hit limits very rarely.
iOS development mostly. I'm using the Pro plans as it's work on personal projects outside my day job and I'm able to get just enough usage from those plans to get me through each day.
I'm actually interested to see how the token discount maps to the usage limit consumption. The conspiracy theorist in me wonders if they're making up the discount and resultant load increase on the API end by reducing effective usage on the subscription end.
Opposite in my experience. I need to limit codex to 500k on medium/low, still run out in 2-3 days with 1 CLI window. CC gives me 4-5 medium/high days with 2-3 CLI windows, and Opus is still great for other regular dumb engineering/refactoring.
On the other hand my head starts to hurt if I read Opus for too long, hopefully they fixed it with 5.5.
This is my experience. After months of hearing how Codex limits were way higher I bumped to the $100/mo plan after hitting my limits a day early on Claude due to some heavy usage + Fable (not normal for me, I often fit nicely in the $200/mo plan). I hit the usage limit in a day with a single agent running on codex and the tiny context window was stifling. Yes, I'm comparing a $100 to a $200 plan but I extrapolated the usage (4x'd it) and it still wasn't close, I got way more done with Opus.
Using Agentsview (which might have it's own issues) I was getting ~$200 of API usage in my 1 week Codex window (paid $100) vs ~$5,000 of API usage in 1 week for Claude (paid $200).
Maybe it's due to 20x / 5x != 4 but I have the $200/mo Claude and $100/mo Codex and I get _way_ less usage on Codex, well under 1/4th the usage. In 1-2 days of semi-heavy _single_ agent usage with Sol High I can burn through my whole week of Codex. Again, this is not running multiple agents, just 1 at a time.
Compare that to Claude and I can run multiple agents on Opus almost indefinitely. YMMV of course but I was shocked at how quickly I burned through Codex usage.
On the context window, I feel so cramped on Codex, compacting happening every time I turn around is annoying. I didn't realize how much I enjoyed the Claude context window size.
I have 5x on both of them. I get way more use from CC than Codex. Actually as we speak, I exhausted my Codex limit twice in the last two days. I am living on banked resets right now.
> "Codex's compaction is very good, fwiw, but it happens so frequently that..."
I appreciate and follow Matt Pocock's advice: avoid autocompaction. Compaction is lossy, which is ok when you're managing it at phase boundaries, but autocompact is lossy at the most inopportune times, firing mid-task and leading to agents going off the rails.
Bad advice, compaction is why Codex is so fantastic.
My conversations compact hundreds of times. By the time it has done a dozen or so compactions, it fully understands the work I want it to do (and how). It's almost like having a fine-tuned Astra model.
I'm not sure I follow; how is autocompaction (lossy summarization), applied at random times (vs strategically, between workflow phases), helpful to ensuring clarity of intent? Maybe you're saying that just plowing ahead and living with the signal loss along the way works well enough for your purposes. In which case, ok, YMMV, different strokes.... but paying attention to context quality and being deliberate about when to compact vs handoff vs delegate to subagents is most definitely not "bad advice".
I agree with @erichocean on this. In theory, compaction is bad. But in practice I found the model is smart enough to write critical details down somewhere, and post-compaction the model doesn't make assumptions. A small amount of time is lost reading materials, but the benefit is that you can operate unbounded vs doing small controlled chunks, which is what I used to do with Opus back in the day. Now I just give it as big a task as I can think of.
This approach got good with Sol. With 5.5 I'd break tasks up, record planning docs, etc.
Now with Sol I rarely bother. It's really good at remembering the salient details. Its also great at continuing a pattern I setup, like commit after finishing each feature block, etc.
not my experience at all. compaction during a task is fatal since you lose all of the details of edits and progress halfway through. compacting after task completion is fine though.
Usage is actually Claude now because of Opus 5.5 since it a better model that Astra. I maxed out my 200$ Claude plan with 10b token on Opus 5 and 5.5 is cheaper. I maxed out two Codex accounts with like not even 5b tokens.
Astra planner/designer with Sol+Luna subagents has worked well for me to improve context continuity. Luna generates code, Sol reviews code and runs/monitors integration/E2E tests. It's about 20% more usage efficient and 20% faster to finish tasks. I've been very subagent-skeptic for a while but the economics of codegen with Luna have made it click. This just works in Codex with a single-line AGENTS.md instruction.
I dunno about Codex-the-application itself, but you can definitely use e.g. Pi with the larger context windows with a Codex login. It puts a pretty large multiplier on credit usage, however.
I've been doing this, my only experience with codex was brutal usage wise and i just retreated back to pi pretty quickly so the credit usage i'm receiving is kinda all im familiar with. Surely seems like less than CC, but i guess not using codex makes my experience kinda not valid for comparing usage.
And ya i can go over that 240k limit, I still very seldom do, and try to treat it as the actual limit. I'm surprised to see so many people still talking about compaction to complete long running tasks, i think the bulk of the work should be somewhat frontloaded into a plan that is split off into subplans, then you can kinda open up a few options, one session with subagents for the subplans of the main plan, or just handoff prompts about progress against the main plan/relevant subplan. I just never trust the blackbox that is compaction, I feel its a recipe for disaster/context poison.
OpenAI have seemed compute-constrained recently, leading to their subscriptions actually being less generous than Claude as of late. OpenAI even paused purchases of 20x plans.
It's extremely variable because the products are roughly equivelant, and a lot of the quality of service depends on their inference capacity at any given hour/day.
Not that I disagree that Codex wins out, but the deciding factor actually is - Codex Pro 20x is not available for purchase, indefinitely. So, what's the point of this discussion? People who already have the 20x sub are unlikely to cancel, and the rest of us can't access it.
> especially when you factor in ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan.
I’m currently on the 5x plan and burned through 5% today on a difficult task in 15 minutes so I doubt that. If you got the wrong kind of tasks that you work on, it can go fast.
I realized I missed a few words here: I meant "especially when you factor in the fact that ChatGPT usage is unmetered", i.e. you get unlimited ChatGPT threads that don't eat into your codex limit
What are you talking about? ChatGPT unmetered? No way! That was 2 months ago perhaps and it’s possible your account still hasn’t gotten the new limits. I noticed around 1 month ago I was still going full throttle on my codex subscription and my limits were barely budging, and then all of sudden people around me started to complain about limits. I thought they’re crazy, but then my account go the hammer, and that was it. If I have the same pattern of usage like I did before, basically having an agent working continuously on a coding take, my weekly limit goes in 2 days.
On usage in ChatGPT settings, I see:
Plan limits
Shared across Codex, Work, Workspace Agents, and ChatGPT for Excel. Chat conversations are not included.
Is this not the default anymore? I am on the (now closed) 20x plan.
That’s the default. I didn’t express myself clearly but I thinking your situation is not the common case anymore, or perhaps you are not using it hard enough. Codex limits deplete very fast these days, it’s not “unlimited”.
I am saying that because ChatGPT usage is unlimited, I don't have to eat into my Codex limits when I use ChatGPT. Codex certainly has limits. Last time I had a Claude sub (hedging here since much of the info in my comment was outdated), my usage limits on claude.ai threads was shared with Claude Code.
Interesting, rolling out new limits would explain a lot. Where did you hear this? I wonder if they detect users with multiple accounts and do that first.
It’s all anecdotal based on my experience and other countless discussions I have seen online. I’ve heard speculation that once they hit 20 million codex users capacity is tighter so they have to manage it. The previous limits were unsustainable compared to token pricing.
theres a popular thread on claudecode or claudai subreddit that proves 20x isnt really 20x. apparently its a marketing gimmick and the recommended solution is two 5x plans > 20x at greater than half the cost of the 20x
> that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan
LMAO, I wish this were true, I hit limits (and the "we are disabling access to protect your data" warnings) all the time, or have chats just...fuck off and get into weird/invalid states (interrupted chats, chats that are spinning and stuck, returning "/mnt/" paths instead of images/md files, file links being returned with no file backing them, image classifier firing...and then returning the image anyway (though now I know that GPT-Image-X really really wants to generate NSFW even when that isn't the request)).
Though I am probably an outlier, I have both 20x Claude/ChatGPT plans and max both out every week, so... (in my defense I am a hobbyist and this is out-of-pocket)
I think it’s only very good in comparison to some of the utter crap that came before it.
Today I had Codex compaction trigger after I had given an instruction but before it acted on the instruction, and the instruction just disappeared completely. The agent reported that the task was done without actually doing it.
I don't think either of these companies are great, then, but Anthropic is surely worse. The doom marketing is one of the most unethical things an AI company can be doing.
That's a nonsensical website. A third party was responsible for both OpenAI's and Anthropic's model escapes. The difference is that Anthropic was warning that this might happen, and OpenAI was putting their head in the sand.
Every place I've worked at has positioned their SEV review process as 'blameless' with respect to the engineers who created the code, processes, etc that led to the incident. On paper, it's not about pointing fingers (though I've found this to be somewhat idealistic), but about finding faults in the processes that let bad code slip through the cracks.
It will be interesting to see how the notion of 'responsibility' carries over as agents handle higher and higher levels of abstraction. Agents built using the popular frameworks of today are somewhere in between code, formerly written by engineers, and an actor (like an engineer). When an agent wrote the framework to audit and monitor ops agents who are monitoring the performance and reliability of the code that coding agents wrote, who is responsible for ultimately ensuring it doesn't happen again? Should the findings just be added to whatever RAG and a few prompts/hooks/skills changed by another agent? When does a human follow up and to what degree?
This is the sort of thing that makes me believe that software engineers will never truly go away; at the very least, they serve the very useful function of being ultimately responsible for something going wrong.
> Who ultimately determines the meaning of a public space: its designers, authorities, or the people who use it?
Restrooms are unique in that they are available to the public but, due to their inherently private nature, are a place where it is difficult to enforce social norms. Because of this, within the confines of the restroom itself, the authorities are all but irrelevant.
I was scrolling through the addictive and easy to stumble into YouTube reels feature and came upon a livestream of a gentleman having some sort of 'public restroom party' in a stall. By this, I mean he had collected every roll of toilet paper available, shredded them into confetti, and was throwing the confetti up into the air while blasting music at full volume out of his bluetooth speaker. He had a few thousand viewers and people appeared to be donating money for supercomments. I didn't stay to watch the end of the livestream, but I don't imagine he cleaned up.
By the time the private-but-livestreamed confetti party is over, the damage is done. You can try to anticipate every single unwanted activity in a restroom and try to prevent it beforehand (e.g. tinted lights to prevent IV drug use), but someone, somewhere will find a way to creatively make the owners decide a few frustrated customers who are told "no public restroom" is a small price to pay for the alternative.
You just charge for it. A dollar per entry should be enough for someone to come by and clean it twice a day, such that it doesn't matter what shenanigans antisocial people get up to in there. No mess will last longer than a few hours.
That appeared to be the norm in Italy when we visited earlier this summer. Most of the main towns in Puglia seemed to have at least one pay-per-use toilet.
Not sure why you're getting downvoted. This approach has been shown to work. The money isn't for funding the maintenance though, it's strictly to keep the riff-raff out.
Hasn't it always been the premise that intelligence would get cheaper? To me, on the enterprise side, it seems like firms are finally getting the memo that, whether you are locked into the Ant/OAI ecosystem or not, you don't need the smartest, most expensive model to do every single task. This is a good thing for overall adoption. Whether that trickles down into regular user behavior, especially with subscription pricing, remains to be seen; even though I intellectually know I don't need Sol for a simple refactor, I am sometimes hesitant to choose Luna/Terra, as it's hard to accept using something positioned, even implicitly, as 'worse'. Remembering that the smaller models tend to be faster is what usually pushes me over the edge.
Anthropic in particular is much more compute-constrained than OpenAI and SpaceXAI and has relied on partnerships to provide inference. This reality factors into their pricing and usage limits (they started 'adjusting' the 5-hour limits during peak hours, and it certainly wasn't an upward adjustment). Accordingly, this is presumably what Anthropic wants, given they develop and release the lower-end models, suggest users use them in various nudges within their product, position the bigger/more expensive models as "For the most complex tasks" in their UIs, and so on.
> whether you are locked into the Ant/OAI ecosystem or not
I think the problem (for Ant/OAI) is that there is no sensible lockin or moat. LLMs are essentially interchangeable and stuff like a harness doesn't offer enough value on its own for someone to be locked into using one of them.
Now with the onslaught of the Chinese models that offer almost the same quality for much less money they have a very serious problem on how to proceed. Investors now might be looking through rose tinted glasses but their patience has its limits.
Agreed 100% for the consumer case: an empty chatbox is just about the least sticky surface I could ever imagine. I saw a mobile interstitial ad for Kimi recently whose hook was basically "Tired of paying for expensive ChatGPT? Download the Kimi app, it's the same thing but cheaper". I myself bounce between token subscriptions like no one's business and use Pi/OMP for maximum model flexibility when coding (and it's a few env variables or lines of (TO|YA)ML|JSON to switch providers in Codex, Grok Build, CC). I even self-host and try to use OpenWebUI + CLIProxyAPI when I can for all my chats.
Enterprise is a whole different ballgame IMO with countless technical, compliance, and employee adoption considerations that add friction to switching. It's also where both Anthropic and (as of last week) OpenAI get the bulk of their revenue, and, incidentally, the venue where US Government regulations on Chinese models would have the most impact.
I swear every time I think I have a unique tweak for an OSS tool, it turns out someone else did it. I forked CLIProxyAPI myself to add something similar, though my implementation is closer to a provider-egress DLP layer and it's intended for API keys and the like, not PII. It uses gitleaks but also supports reversible redaction so a secret that's flagged can be replaced with a placeholder before CLIProxyAPI sends the request to the provider and then any instances of that placeholder can be rehydrated in the response to the caller.
Since it modifies logic across the full request/response lifecycle, I unfortunately couldn't implement it cleanly with the existing plugin API.
I have a Honor MagicBook Art with a Snapdragon X Elite I got off of Aliexpress and love it. It's only available in China and the EU AFAIK. The seller initially sent me a Thinkpad with a Pentium III, claimed it was a 'mistake', and said they were sending me the correct order. I was doubtful, but sure enough, it came. Battery life is amazing and the thing is really well-made.
There's likely a niche market for ThinkPad Immortals where customers send in their model for interior design and refit with latest chips and display resolutions, ramaxxing.
What was the modality of interpreting the UI/interacting with the phone via ADB? Did you/your agent use screenshots + vision model to essentially target coordinates on screen? I ask because I'm curious if you found a non-screenshot + coordinate based ADB method to do so (that doesn't rely on UIAutomator).
I wrote an APK that gives access to the accessibility service via an in-Android http server, which talks to a daemon running on the host machine which can, in turn, talk to Hermes. It's designed to be used with an emulator called Waydroid but you could probably run it on your device too. It does need to fall back to ADB when the accessibility tree is incomplete/inaccurate (which is often). And, going back to my question, I haven't found a means of interacting via ADB other than by coordinate/UIAutomator. To be clear, I am not smart or patient enough to be an Android dev, so I use it for ordering things on Amazon.
I didn't end up needing to go full scrcpy because my use case just runs Android in an emulator (Waydroid) for my Hermes agent to run apps. If I need vision or (x,y) coordinate interaction, I can just screenshot the emulator window and interact via ADB. However, I ended up using the accessibility tree to create a semantic, ephemeral ref API as the preferred mechanism for interaction. Basically more 'Playwright for Android' than 'Remote Desktop for Android'. But, unfortunately, the accessibility tree is often incomplete or inaccurate.
If you find out, let me know! I have been working on a port of Excalidraw to my ReMarkable Pro and it's opened my eyes to how much we take for granted in terms of UI interactions when we assume we have a 30+hz screen with no to minimal ghosting. Layout has to become as monotonic as possible, especially in the absence of user interaction; in other words, the app itself should never be doing anything to disrupt prior renders on its own unless you are clearing the whole page. When it is the user interacting with the app, try to minimize the number of intermediate states their interactions produce. This does NOT mean "just don't render those intermediate states", however. It means make the interactions themselves not require intermediate feedback. Essentially, users (and developers) are so accustomed to just being able to manipulate things with no side effects that you have to make every possible interaction as intentional as possible, leaving them less frustrated when they do see a refresh. If an operation is going to cause a refresh, it should happen at a moment the user already understands as a semantic boundary.
You obviously cannot always avoid this, so when there is a hard requirement for continuous feedback, use a degraded/cheap transient representation and refresh at commit of that action. With E-Ink, every interaction has to be a trade off between ghosting and latency. One common pattern I ended up coming back to is intermediate states that optimize for latency at the expense of ghosting until the UI action is 'complete'. This will allow refreshes to be associated with the satisfaction of finality. Funny enough, many of them harken back to old UI paradigms from when computer graphics weren't as powerful as today. When resizing, I simply draw the outline of the shape with a fast update mode that tolerates more ghosting until the user releases their finger or stylus, at which point, after a small lag to account for an 'oops, just a tiny bit bigger/smaller', I do a localized refresh of the union of the old + new area.
The outline itself is deliberately faint, so the refresh doesn't have to be as intense if its a small delta.
For your streaming LLM case, I would recommend doing it in chunks, like you said, perhaps doing a hard refresh of everything but the input box + moving the tail of last sent message + beginning of stream to the top upon send. Of course, this implies that you know exactly how long the response will be, which you don't. The challenge will be coming up with a UI paradigm that doesn't make it look awkward if the response only goes down to the middle of the screen while also nicely doing a clean refresh pre-emptively if it's clear it will overflow and need to move to the top. For the former case, find some way where it looks not completely awkward if it's partial and then do refresh of the chat history area and move it to its proper fitted position once the user clicks on the input box again to type their next reply. For the latter, perhaps embedding some kind of remaining space meter at the bottom that indicates when you'd need to do this move + refresh action (don't directly label it that way) would enable the user to anticipate it more and be less frustrated when it happens, as they are psychologically awaiting the next part of the answer/reasoning, not surprised at a sudden flash.
For the actual streaming, you'll have to ensure that your text alignment works in such a way that once a line is rendered to the screen, its placement is final (in other words, no streaming intra-word, as you need to know if the word will fit in remaining space on the line before rendering it). Basically, avoid retroactive reflow like a PDF document with a fixed layout, not like this text box I am typing in with a resize handle at the lower right. You'll also probably want to disable scrolling during generation as well, or at least make some sort of discrete pagination mechanism for going back and forth.
In any event, I look forward to seeing whatever it is you're building!
reply