massa

@robinebers opus 4.6 is severely misaligned

when I say Claude Opus 4.6 is behind, that's what I mean this model is truly the laziest of the bunch Codex would never

claude is severely misaligned

@VictorTaelin opus 4.6 is EVIL x.com/tycnio/status/…

Seems like there is a bug on Anthropic's API, affecting all models, where the request will stream the final visible token quickly, then just hang for ~5 more seconds before sending message_stop and actually closing. Repro: gist.github.com/VictorTaelin/6…

opus 4.6 is fucked #issue-4111985825" target="_blank" rel="nofollow noopener">github.com/anthropics/cla…

@thsottiaux @steipete @Dimillian

The image speaks for itself. For Quality: Codex & ChatGPT > Claude Code I asked each to review plans. `which is likely to get me to the goal? and the goal is an extremely high-end interface... not buildable by regular LLM training distrubition... see for yourself

@gabriel1 yeh that does not actually work.

i find great success with standalone "cleanup prompts" to my prs. you can stuff every rule you got into agents.md, but beautiful code is secondary until it's the only focus for example: "simplify all code so it's extremely easy to consume, remove not strictly necessary code"

okay this solves a lot of problems and is actually amazing. thank you @thsottiaux

@JaredOfAI @trq212 thats a YOU problem id do the same at 106 . i dont blame him

@trq212 My Claude code 1m opus 6.4 rejected to create more sub-agent in our conversation after creating ~120 sub agents with 64% used context window , is this a feature or bug?

we need a better word than vibe coding man, Claude can create the most beautiful things

@JaredOfAI @trq212 loolllll

@trq212 Why is it okay, though, that they ruin the codebase most of the time?

@jorgemanru both will poison yur codebase

My very subjective perception from the last weeks: - Claude was ahead of Codex. - Codex suddenly became as good as Claude, sometimes better and faster. - Overnight, Claude is substantially ahead of Codex again, both in speed and performance.

@rezoundous they are both terrible

Claude was getting stuck on this problem for 1 hour, and Codex just 1 shotted it. I could've sworn it would've been the other way round not too long ago.

How is it that no one is addressing the escalating situation involving Codex and Claude? They are actively poisoning code bases & wasting ur time. Insane. I let them code today on their own for 10 minutes & they fucked up an entire weeks work. Thank god for git.

Codex wastes so much time and effort and creates so much unnecessary code for "migrations" and "backwards compatibility" and "regression testing" and "exception catching" when it's unnecessary and adds a lot of code debt. I tried using only codex with 5.4 for a new feature because so many people were posting on X about how "real coders use Codex not Claude!" and I just find it full of slop, bloat, useless "safety" checks, loose typing, and just bad coding practices. I don't know, maybe gpt 5.4 can write a CUDA kernel better than Opus 4.6 or something, but really I find codex nigh unusable. Claude Opus just gives me only what I ask for, and it's much better at using cli tools.

@big_duca these agents claude code and chatgpt suck

We have AGI (for coding). And yet so much software is still so damn buggy. (including my own startup) Why?

@thsottiaux ok but gpt 5.4 fails all my benchmarks. pls git --reset first

Working at OpenAI is fun because questioning everything and taking risks is part of the culture. Within Codex, the team asks itself how we could make it an order of magnitude better every few months and then sets most things aside to go and do it across the entire stack. Some examples were the Codex App and our first deployment of Cerebras inference with WebSockets. We are now well under way on the next bet and it’s making even our best engineers nervous as it’s at the edge of what’s possible today.

@iannuttall the models been off since its been out its a refression

anybody else felt like gpt 5.4 has been a bit "off" today? just feel like I am fighting with it on things it had no issue with before like building my chrome extensions for local vs prod, working with conductor ports, etc

@thsottiaux @DavidOndrej1 gpt 5.4 is a regression though tibo .... whats up with that

@DavidOndrej1 Smoking incredible code I would walk a mile for code from GPT-5.4

GPT 5.4 *is not* better than Opus 4.6 i have no idea what people are smoking

@NickADobos Also with subagents: - Codex: yeah, I’ll spawn agents in batches and let you know when everything’s done - Me (40 min later): did they finish? - Codex: you’re right, I should have spawned the agents. I will do it now