Comment by eggbrain

My experience is that the GPT-family of models are very smart and figure out bugs, edge cases a bit better, but it produces code that is much less mergable – if you review the code, it introduces a lot more useless/inappropriate heavy abstractions and wrapper functions, compared to the Claude-family models which introduces the right amount of straightforward human-style code.

I can recognize so much of the GPT/Codex generated code long after it gets merged (not by me).

Additionally, the time spent on every agent turn on GPT 5.5 is much longer compared to Claude Opus 4.8, which means iterating on the code takes a lot more patience, and there's a lot more nitpicks to pick when actually using GPT 5.5 to do software engineering.

Feels like GPT-style models are more geared on doing one-shot software vibing (and handling the vibe coded mixture) compared to Claude's focus on actual software maintenance. I got a GPT Pro sub for free and wanted to cancel my Claude subscription so much, but I still keep reaching Claude models a lot more. Frustrating.

Codex IME is just smarter, I think it shows given both anecdotes but also how OpenAI has always been at the front of programming competitions and math problems.

But Claude models seem to be better at long term problems or more ambiguous problems.

I'm curious as to what the primary benefit here. Are there secret improvements in training? There hasn't been much in fundamental model architecture, I don't think. What about harnesses? I wonder what's pushing the AI. It seems like harnesses is the main thing pushing AI ever since CoT.

I guess enjoy it while it lasts? OpenAI won't be able to subsidize that forever either.

100% I constantly get errors and timeouts on single responses in Claude, and certainly hit limits all the time. Codex rarely. In fact, I bought a second $200 Codex plan because the quotas seemed fair and I didnt have constant issues. Claude is so great at a lot of things, but unfortunately Anthropic beats you away with a stick every chance they get.

I've only ever had the $20 month claude plan but last night took the time to setup opencode + openrouter paying for deepseek + glm. Previous experience, while extremely awkward, I'd hit my limit within one or two chat replies and it'd take me like 4 limit cycles to complete my task. Now I'm able to complete an equivalent task entire task for less than $2 in two cycles (ask -> revise).

I'm doing basic web development here utilizing animejs. Nothing too complicated (mostly saving time doing the scaffolding, still write the bulk of animations manually).

Truly believe that American companies are going to get completely curb stomped by China due to greed, ineptitude, and violating the social contract.

I have been using both codex and Claude in my day to day, trying to not get to attached to one. I want to be able to work with any provider in case one of them does something bad.

I feel like Codex made a big push to run everything on your laptop. With Claude, I get 4 cpu's, a fair amount of ram and 30gb for every one of my dumb ideas for free in the cloud containers. Codex used to be similar, but last time I tried it just kept pushing me to run it locally on my laptop, which I really did not want to do with 20 requests going at once. That's the main advantage for me at the moment.

I've found Codex to be the better subscription for OpenClaw, because the limits are indeed very generous. However, I've found more and more that Claude Routines/Scheduled agents can replace all the tasks I use OpenClaw for, so I've been slowly switching over to Claude Code. Aside from OpenClaw, I don't find a lot of value in Codex as a harness on it's own.

I have trouble justifying gpt after that gross stuff with the war department.

Though the day is coming when there’s no distinguishing, I’m sure.

I do slightly prefer 5.5 for complex work but Claude quota usage has gotten infinitely better since the dark days a few months back - has gone from being infuriating to something I pretty much don’t have to worry about with it as a daily driver. (In fact, hitting GPT weekly quotas is more annoying now). Understand if people are still scarred by the issues + poor comms around them, though.

Do you use a token service like open router or just subscribe to / unsubscribe from various models sequentially?

Wait till you kick the tires of Qwen Coder.

[flagged]

they have a graph with cost comparison between the models. This model is just a little over the other models as cost. The graph is logarithmic :)

I'm also on the $100 max plan. I let Fable rip on a complicated issue involving hot-reloading modules in a GUI app built with Racket, it's fixed a couple issues over the last hour, and I've used about 17% of my session (not weekly) limit.

That’s odd, I used it on a pretty complex refactoring task and it worked for 22 mins and used only 15% of my 5-hour limit. I’m on the $200 Max plan though.

The CLI when you select it says it has 2x the usage as opus. Not sure if that matches what you are seeing.

I do wonder if you switched models mid-session, you would have lost all your cache. Reloading the context into cache can really eat through your usage.

I too am on the $100 plan and I second this.

I had it analyze a project I was working on with Opus 4.8, and it blew through 23% of my session limit in one go. Does not portend well for my budget.

Yes, and this is also why I haven’t yet tried the new “dynamic workflows” which spawn hundreds of agents that happily eat through your token limits.

What is your effort level?

They didn't even reset credits for this lol

I just asked Fable for a complete code review of my lone lisp project. Started out strong. Launched Fable agents, then spent like 10 minutes thinking... And then got interrupted by a switch to Opus 4.8.

> Fable 5's safety measures flagged this message for cybersecurity or biology topics.

> They may flag safe, normal content as well.

> These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them.

Here are the results of the agentic code review session:

  ┌──────────────────────────┬───────────────┬────────────────┐
  │          Agent           │ Fable 5 turns │ Opus 4.8 turns │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ values                   │ 134           │ 0              │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ data-intrinsics          │ 104           │ 0              │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ tools-tests-build        │ 81            │ 0              │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ core-intrinsics (failed) │ 25            │ 0              │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ system-memory            │ 44            │ 20             │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ reader-modules           │ 104           │ 25             │
  ├──────────────────────────┼───────────────┼────────────────┤
  │ linux-startup            │ 95            │ 15             │
  └──────────────────────────┴───────────────┴────────────────┘

This 40 minute session cost me 16% of my weekly usage. A simple code review of the most critical areas of my project got flagged as a cybersecurity risk. It really made me not want to try it again.

I had a similar experience. I wanted to test it by asking it to summarise a scientific OMICs-related paper. It gave a warning about me potentially developing a bio-weapon or something like that. And switched back to Opus 4.8.

We just blocked it at our org for this reason. They will "retain agent request and output data associated with this model, regardless of you Cursor Privacy Mode setting."

What does "zero data retention" mean? What kind of data does it need to unlock?

You really misunderstand what AI-doom people are worried about if you think this is anywhere near the top (or middle, or bottom) of the list of concerns.

I don't think offering a product under a certain set of terms obligates a company to maintain that offering forever. The bait and switch is certainly annoying but seeing as they're very upfront about it you can't say you weren't warned. Don't like it? Don't use it.

Yup - who cares about x-risk or red lines for domestic mass surveillance anyways? I draw my red lines at prioritizing profitable customers when heavily resource constrained. That's the true definition of evilness!

I wouldn't call Anthropic ethical. But between Anthropic and OpenAI, Anthropic is the more ethical one

Why would you have ethics when you could get that IPO money instead?

It's unethical to price it in a way not everyone can afford?

It smells like an architecture-related issue to me. They wanted to release the model asap, but they're still implementing the fine-grained controls to constrain the model to non-subscription users.

They said they would release it back into subscriptions as capacity allows in the future. If they don't, people are going to point back at it and rake them over the coals.

The bar is just too low.

More ethical in some areas, actively user hostile in others

More of a free trial to those authenticated and qualified with existing payment. Subscription billing is going away for sure though eventually based on the economics. Token “all you can eat” is a capital furnace otherwise.

(I’m highly confident open models will eventually achieve a similar performance benchmark with distillation over time)

I think they are just prioritizing enterprise customers, because this is were historically they made most money.

That's not how it works. They don't need revenue, they need addicts.

Specifically they need businesses that fired people and adapted their business to the products, so when the unsubsidized costs hit the businesses are forced to eat the true costs.

Yes they can't afford to give the products for free, but what is essentially happening with AI services is economic dumping, keep costs artificially low to get people to fire everybody, and then Jack the rates once they have Monopoly control

That's a big problem for all of the AI companies. Most people don't find the technology compelling, accurate, or ethical enough to pay for a subscription.

Why wouldn't Anthropic just wait until people start subscribing, do some kind of marketing push, or obtain some kind of other sustainable revenue stream, before they go IPO? I wonder if they see the writing on the wall with all of this and want to cash out as quickly as possible?

I just assume Opus is constantly nerfed based on capacity. I was exclusively Claude for a long time, but the inconsistency in quality, constant outages, and slow downs were too hard to work with.

I just use dumb and fast models now. I'm more engaged. I think that the higher the quality of the model, the more you tend to vibe with it, and then the more hallucinations you then miss. I'm not sure which is more productive, but I definitely burn out faster the more I vibe. At some point you're spending your time on forums, discord, or youtube instead of engaged with what you're building. Or you yak shave about your tooling and end up creating the 600th multi-agent gastown harness and blowing thousands of dollars on tokens to create it only to discover it's too expense to actually use.

It's possible that they will transition to usage credits but why not take them at their word? To date they have continued to offer better and better models to their subscription plans.

HN needs to take a chill pill. Could it be that Mythos is expensive and they just want to give people a taste of it? I mean the alternative is not offering it at all?

Those already landed! Oh, you weren't talking about 4.8?

I don't think they'll phase out subscriptions ever, their whole play has been to drive demand from the bottom up. Get engineers hooked on building with claude at home, then get them to demand the ability to use it at work, and bend over their employer with no lube.

They'll probably tighten the quotas to reign in whales though.

They almost certainly already make a fuckload more money off API pricing than they do subscriptions, even if there might be more total subscription users. So offering subscriptions even at some loss is probably going to continue. Honestly, I'd be surprised if they even lost money on most subs; there are definitely Token Whales out there who mess up all the accounting up, though.

Realistically I think Anthropic just has insane demand but finite capacity to run models, and Fable will just make them more money if they dedicate it to API pricing. I suspect the goal here is something like: get individual engineers/PMs on their personal plans to taste Fable and then go to their meetings and say "Yes doubling the price of every single input/output token is a good idea, boss".

I certainly hope not. PAYG is not predictable enough for smaller companies or individuals. Where I work (non-tech company), PAYG would never fly. We aren't big enough for that. Of course, you can set usage budgets, but there's a pretty big difference between $200/user/month vs. the equivalent PAYG usage being closer to $1,000/user/month, if you currently use the subscription plan to its limits each week.

Going PAYG only will effectively take these tools away from a huge amount of people and accelerate the push for local LLMs.

OTOH, accelerating the push for local LLMs would also be fine with me.

I doubt it, given the importance of those subscriptions for building and maintaining market awareness.

The AI landscape is changing rapidly, and with Apple announcing the option to change the AI backend, and potential requirements enable AI choices as well, similar to EU browser choice requirements (this is more reading tea leaves than any actual requirements I am aware of). The new OS changes coming to support Googlebook, and deep Copilot/AI integration into Windows will make maintaining user facing subscriptions essential for independent model developers like OpenAI, Anthropic, and Mistal to remain relevant longer term.

If the don't maintain that relevance there is increasing likelihood that they will get consumed by other companies whether it's Apple, Microsoft or Google to form a foundation for their OS, or other cloud providers.

I switched back to Sonnet. It replies faster so I work faster. Also cheaper. But I really like the speed. I have to be more specific with what I want. Also I stop it more often than Opus. These new models will be awesome, but they need to increase the speed.

K2.6 on Cerebras is basically a preview of the future. We'll eventually get similar performance locally with Tenstorrent hardware.

Agreed, everything since 4.6 has been worse

Just like how Elon forced FSD in Tesla to be subscription-only (he was incentivized to do so).

It could be a much bigger MoE model

Thankfully this, like most other tech, will get cheaper through the years.

I think the goal is "private citizens: subscriptions; corporations: per-token billing." It's getting people addicted to LLMs on cheap subscriptions so that they can then force companies to pay for expensive inference.

I don't see how it won't be. They lose insane amounts of money on subscription plans. I'm sure they still lose money on usage-based billing, but probably not as much.

Models are getting better, but there's a negative change in terms of "productivity" per dollar. Yeah, I can throw 5 sub-agents at the problem, but the cost is getting significantly higher. And yes, I can crank out the solution much faster, but again, at some point that cost will be hard to justify. And it doesn't matter if the cost is subsidized by a provider, if it's paid by your company, or from your pocket. We are slowly reaching a point where the cost will be too high to justify the gains.

This is probably the end of 'use the best model no matter the price'

The pricing can be a bit deceptive though. A good model can deliver the same results in fewer tokens.

Kind of like billing a programmer by the hour.

Why wouldn't it be? How much would you pay a scientist at this point to think about a problem for you and give you a solution?

If it's that big of a problem to you, you're free to just... not use the freebie?

The point of SaaS was that the marginal cost (of supporting another user) was low. That does not apply to LLMs.

Where is this text coming from?

[edit] -- I see that this comes from the system card -- dang merged the comments from the other discussion so that explains the confusion.

The AI circular infinite money glitch won't last forever. I hope.

If you have good expertise in a domain and access to cheaper models, you may still be more skilled than someone without expertise but a lot of money to bruteforce the problems using SOTA LLMs.

Not with Codex

Really, you have never heard of shareware or trial periods?