[dmazin]

I got sick of the inconsistency caused by Anthropic tinkering with Claude Code and had canceled my 20x. My plan was to switch to Codex so I could use it in Pi.

I am specifically talking about switching because of the harness, not model quality. Anyone else match my experience?

I wonder how many other people recently did the same. It would be prudent of Anthropic to let people use Pro/Max OAuth tokens with other harnesses I think. Even though I get why they want to own the eyeballs.

[redrove]

I’ve been using Codex Pro since they lobotomized Opus 4.6. Codex is so much better, GPT 5.4 xhigh fast is definitely the smartest and fastest model available.

For a while there I had both Opus 4.6 and Codex access and I frequently pitted them against each other, I never once saw Opus come out ahead. Opus was good as a reviewer though, but as an implementer it just felt lazy compared to 5.4 xhigh.

One feature that I haven’t seen discussed that much is how codex has auto-review on tool runs. No longer are you a slave to all or nothing confirmations or endless bugging, it’s such a bad pattern.

Even in a week of heavy duty work and personal use I still haven’t been able to exhaust the usage on the $200 plan.

I’ll probably change my mind when (not IF) OpenAI rug pull, but for spring ‘26, codex is definitely the better deal.

[walthamstow]

I also made the switch to OpenAI, the $20 plan, I dunno about "so much better" but it's more or less the same, which is great!

The models and tools levelling out is great for users because the cost of switching is basically nil. I'm reading people ITT saying they signed up for a year - big mistake. A year is a decade right now.

[redrove]

I underscored using xhigh + fast mode when saying it’s so much better.

Now with Opus 4.7 of course the “burden” of adjusting reasoning effort has been taken away from you even at the API level.

In my experience people don’t change the thinking level at all.

[sitkack]

What issues did you consider about sending your code base to OpenAI?

[walthamstow]

None mate. Code is cheap, it's not worth anything any more, especially not my little personal projects

[Scotchy]

Any alternative to Claude Design ? Tried Figma with Opus 4.6 but it doesn't come close in my experience.

Codex is abysmal for UI design imo.

[dgb23]

It really depends on what you‘re trying to do and what your skillset is.

But if you go information architecture first and have that codified in some way (espescially if you already have the templates), then you can nudge any agent to go straight into CSS and it will produce something reasonable.

[makingstuffs]

Have you tried stitch.withgoogle.com?

[freedomben]

Thanks for the tip! Hadn't seen that, but definitely giving it a try.

[gbalduzzi]

I created some decent prototypes with stitch but I don't know how it compares to claude design

[freedomben]

stitch.withgoogle.com

[joelmanner]

I've been using paper.design and it's been working well for me via mcp on claude code

[StrangeSound]

Google Stitch

[tommica]

I left anthropic a while ago because of the similar shenanigans they had earlier. I went with opencode & zen.

I still have their subscription, but am using pi now, mainly because something happened that made my opencode sessions unusable (cannot continue them, just blanks out, I assume something in the sqlite is fucked), and I cannot be bothered to debug it.

For what I use the agents, the Chinese models are enough

[hboon]

Doesn't using pi be against their terms of use about having to go through Claude Code cli for all Max plan usage? (I had use Droid with Max previously, it was a great combo).

[the_mitsuhiko]

It's unclear right now. The current stance is that using pi or other coding harnesses eats into extra usage and that is the behavior one sees today. We have added a hint to pi now that warns you when you use an anthropic sub.

[hboon]

Thanks for the great work.

[tommica]

Probably - it was that kind of confusion that resulted in me switching providers.

Plus I like being able to switch a model.

[resonious]

I also cancelled my 20x and switched to Codex. At this point even the Codex CLI seems to perform better than Claude Code... And so far I'm on the OpenAI Pro plan and haven't even needed to upgrade to their $100/mo plan. I'm getting more value for almost 10x cheaper.

[hboon]

I switched to Droid+Opus (with Claude Max) many months ago and it was my favorite combo.

Had to stop because they don't like us proxying requests anymore.

[serial_dev]

My experience is the opposite of this thread's consensus. Context: Full time SWE, working on large and messy codebase. Not working on crazy automations, working on fixing bugs, troubleshooting crashes, implementing features.

Anthropic models write much better code, they are easy to follow, reasonable and very close to what I would done if I had the time... OpenAI's on the other hand generate extremely complex solutions to the simplest problems.

I was so disappointed by non-Anthropic models, that for a couple of weeks I only used Anthropic models, but based on this thread, I'll go back and give it another try. It's good to go back and try things again every couple of weeks.

Of course, I was annoyed that they lobotomized 4.6, the difference was day and night, and Anthropic is certainly not a company I trust. In my opinion, it shows their willingness to rugpull, so I'm looking at other approaches. Since 4.7, things went back to normal, things you'd expect to work just work.

[yokoprime]

I feel like Opus 4.7 vs GPT 5.4 is pretty much just flavor variants, the big difference is in the harness. I like the Claude Code CLI better than the Codex CLI, it just clicks with how I like to interact with agents. The codex app on the other hand is better than the Claude app in code view, so if I had to stick to an app it would be codex all the way.

[athrowaway3z]

I've been on pi for a few months now, build a custom tmux plugin so i can use nested pi and mix and match codex / claude instances.

pi has been the better harness out of all the ones i tried, first and third party.

Ever since the Anthropic block i've just canceled all my claude subs. Used to be codex was a bit worse, now they're practically equal. Claude is slightly better at directing other agents but the difference is too minor and not worth the money.

Claude usage limits / costs are absurd.

Any 'principles' people praise anthropic for are not that relevant to me anyways because i'm not a US citizen.

[ai-tamer]

(Disclosure: I work on tamer, an OSS supervisor for coding agents — biased.) Add one more to the count. The OAuth-across-harnesses idea would help, but it doesn't fix the shape of the problem. "Harness" has always felt off to me. Exoskeleton is closer — Claude Code, Codex, opencode wrap the model and augment it from the inside. What's missing is a layer above that's explicitly not an exoskeleton: a thin supervisor. A master that watches and guides, nothing more. It just relays I/O and hands approval back to the human.

[uvu]

Same, I am from 5x plan and cancel and switched to codex as I want to use Pi.

[KronisLV]

> I wonder how many other people recently did the same.

Some negative signal for better overall view on things: I'm still with Anthropic and will probably stay with them for the foreseeable future.

I think after DoD/DoW shenanigans (which in of itself felt like a reasonable take on the part of Anthrpic) they got a bunch of visibility and new users, so them hitting some scaling limits is pretty much inevitable - so some service disruption is inevitable. Couple this with the tokenizer changes and seeming decrease in model performance (adaptive thinking etc.), and lots of people will be rightfully pissed off, alongside increased downtime (doesn't matter that much for me, definitely does matter for anything time-sensitive).

At the same time, in practice I've only seen it do stupid things across 8 million tokens about 5 times (confusing user/assistant roles, not reading files that should be obvious for a given use case, and picking trivially wrong/stupid solutions when planning things), alongside another 4 times that tests/my ProjectLint tool caught that I would have missed. The error rate is still arguably lower than mine, though I work in a very well known and represented domain (webdev with a bunch of DevOps and also some ML stuff, and integration with various APIs etc.).

At the same time, the 85 EUR they gave to me for free has been enough to weather the instability in regards to pricing changes and peak usage. They've fixed most of the issues I had with Claude Code (notably performance), and the sub-agent support is great and it's way better than OpenCode in my experience. They also keep shipping new features that are pretty nice, like Dispatch and Routines and Design, those features also seem nice and not like something completely misdirected, so that's nice. The Opus 4.7 model quality with high reasoning is actually pretty nice as well and works better than most of the other models I've tried (OpenAI ones are good, I just prefer Claude phrasing/language/approaches/the overall vibe, not even sure what I'd call it exactly, all the stuff in addition to the technical capabilities).

At the same time, if they mess too much with the 100 USD tier, I bet I could go to OpenAI or try out the GLM 5.1 subscription without too many issues. For now they're replacing all the other providers for me. Oh also I find the subscription vs API token-based payment approach annoying, but I guess that's how they make their money.

[benjx88]

Because the Harness is the Moat and key IP not the Models themselves that is the why! now for both OpenAI and Anthropic with all their money raised and the compute they acquire and have in the books of course no one can easily replicate, whom can afford all those datacenters and Nvidia GPUs interconnected is why OpenAI throws you a bone and gives you an Open Source SDK Harness but not the one they actually use for ChatGPT. But now both of them have to deliver and do all the bull-shet they said this models can do... truth is they cannot. So now the bubbles burst and we will see what happens. We all have to buy iPhones or MacBooks so that makes sense, we all use Chrome or Google Search, Instagram, TikTok.

All these models and agents are shortcuts for all of us to be lazy and play games and watch YouTube or Netflix because we use them to work-less, well the party will be over soon.