While I agree with you that this is likely AI assisted, I think this may be changing now.
People speak in the manner of what they consume. If you consume a lot of claudish, you will eventually start talking claudish too. And I've already noticed people talking claudish in real life.
The entire reason why hardware prices are so absurd right now is because manufacturers across the board are doing everything to prevent that crash.
They all collectively chose NOT to increase supply with increased demand. So if the bubble pops, they just go back to previous prices without oversupply driving the prices to rock bottom.
I've developed several plugins for different harnesses, and I needed some upstream change for most of them.
The deciding factor for me whether or not I will work on the feature of the plugin is whether I (or rather, my agent) can look in upstream source and evaluate if it can be done with minimal upstream change, which I then contribute. And generally, even if no upstream change is needed, agents work so much better when they can read the code.
So why not just make Claude code open source? Considering also that source code was leaked once 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.
> 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.
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.
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.
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.
Don't forget that Opus 5 was tracking fable on many benchmarks, yet it was borderline unusable for any coding work. My Claude sub usage has been 100% fable, 0% opus 5.
Benchmarks often don't survive contact with reality.
That's not my experience at all. Opus is an extremely capable coder on high or xhigh effort. It can read academic papers, implement algorithms from the description in the paper alone and reproduce results without breaking a sweat. This is remarkable because it is pure reasoning on unseen material; in some cases the paper was just published and there wasn't an implementation to learn from in the training data.
did you actually verify that it's output in those scenarios is good ? in my experience opus has been a disappointment and constantly trailing behind actually solving hard problems versus the OpenAI models.
I'll say that both have terrible writing style though.
> did you actually verify that it's output in those scenarios is good ?
Yes, in the sense that it reproduced results in the paper or known solutions obtained by other methods. In fact, Opus is very good at checking it's own work in my experience.
Opus is fine at coding (for correctness), but horrible at talking about code. I don't really see the defect rate going down when using Astra or Fable 5.1, but they are just more coherent in both how they explaing code/architecture/choices, and how they actually code the thing. With Opus, I'm using smaller models to delete the vast majority of comments and 'clean up' correct code that is too weird.
Thing is, I'm still reading the majority of generated code, and I have colleagues who'll laugh at me if my PRs are a shit show. I fear what vibe coders are pushing to the servers of myriads of start ups, and pity the poor people who'll have to clean it up in a year or two.
Mines pretty much inverted - my colleagues and I noticed almost zero difference between the quality of code in Opus vs Fable. Occasionally I'll switch to Fable for an arduous debugging task but that's about it.
> For users with cybersecurity use cases that may be blocked by our cyber safeguards, we
recommend accessing our models with reduced cyber blocking classifiers via our Cyber
Verification Program. Claude Opus 5.5 will be available through this program in the near
future.
Anthropic has used "in the near future" for Mythos-class models too, but CVP is still Opus 5 only.
Why even have the program designed for trusted access to cyber capabilities if you're not providing access to cyber capable models via the program?
It seems like OpenAI's takeaway after Sora is to not stretch themselves thin and focus on what's important.
Based on what's publically available, they're focusing on hacking uncontesting orgs using misconfigured sandboxes and math puzzles.
Your statement is essentially unfalsifiable. We can't possibly discuss whatever Sam Altman is doing in his private office room, nor should we assume OpenAI is working on anything other than what has some public traces.
You can't work on anything in an organization as large as, and with as many eyes on as OpenAI/Anthropic without it becoming publicly known very quickly.
While I agree with you that this is likely AI assisted, I think this may be changing now.
People speak in the manner of what they consume. If you consume a lot of claudish, you will eventually start talking claudish too. And I've already noticed people talking claudish in real life.
reply