> With Codex (5.3), the framing is an interactive collaborator: you steer it mid-execution, stay in the loop, course-correct as it works.

> With Opus 4.6, the emphasis is the opposite: a more autonomous, agentic, thoughtful system that plans deeply, runs longer, and asks less of the human.

Ain't the UX is the exact opposite? Codex thinks much longer before gives you back the answer.

I'm personally 100% convinced (assuming prices stay reasonable) that the Codex approach is here to stay.

Having a human in the loop eliminates all the problems that LLMs have and continously reviewing small'ish chunks of code works really well from my experience.

It saves so much time having Codex do all the plumbing so you can focus on the actual "core" part of a feature.

LLMs still (and I doubt that changes) can't think and generalize. If I tell Codex to implement 3 features he won't stop and find a general solution that unifies them unless explicitly told to. This makes it kinda pointless for the "full autonomy" approach since effecitly code quality and abstractions completely go down the drain over time. That's fine if it's just prototyping or "throwaway" scripts but for bigger codebases where longevity matters it's a dealbreaker.

I think it's the opposite. Especially considering Codex started out as a web app that offers very little interactivity: you are supposed to drop a request and let it run automatously in a containerized environment; you can then follow up on it via chat --- no interactive code editing.

This kind of sounds like both of them stepping into the other’s turf, to simplify a bit.

I haven’t used Codex but use Claude Code, and the way people (before today) described Codex to me was like how you’re describing Opus 4.6

So it sounds like they’re converging toward “both these approaches are useful at different times” potentially? And neither want people who prefer one way of working to be locked to the other’s model.

> With Opus 4.6, the emphasis is the opposite: a more autonomous, agentic, thoughtful system that plans deeply, runs longer, and asks less of the human.

This feels wrong, I can't comment on Codex, but Claude will prompt you and ask you before changing files, even when I run it in dangerous mode on Zed, I can still review all the diffs and undo them, or you know, tell it what to change. If you're worried about it making too many decisions, you can pre-prompt Claude Code (via .claude/instructions.md) and instruct it to always ask follow up questions regarding architectural decisions.

Sometimes I go out of my way to tell Claude DO NOT ASK ME FOR FOLLOW UPS JUST DO THE THING.

> With Codex (5.3), the framing is an interactive collaborator: you steer it mid-execution, stay in the loop, course-correct as it works.

This is true, but I find that Codex thinks more than Opus. That's why 5.2 Codex was more reliable than Opus 4.5

Just because you can inject steering doesn't mean they stered away from long running...

Theres hundreds of people who upload Codex 5.2 running for hours unattended and coming back with full commits

I am definitely using Opus as an interactive collaborator that I steer mid-execution, stay in the loop and course correct as it works.

I mean Opus asks a lot if he should run things, and each time you can tell it to change. And if that's not enough you can always press esc to interrupt.

I do not trust the AI benchmarks much, they often do not line up with my experience.

That said ... I do think Codex 5.2 was the best coding model for more complex tasks, albeit quite slow.

So very much looking forward to trying out 5.3.

Impressive jump for GPT-5.3-codex and crazy to see two top coding models come out on the same day...

they tested it at xhigh reasoning though, which is probably double the cost of Anthropic's model.

Cost to Run Artificial Analysis Intelligence Index:

GPT-5.2 Codex (xhigh): $3244

Claude Opus 4.5-reasoning: $1485

(and probably similar values for the newer models?)

Did you look at the ARC AGI 2? Codex might be overfit for terminal bench

In my personal experience the GPT models have always been significantly better than the Claude models for agentic coding, I’m baffled why people think Claude has the edge on programming.

Opus was quite useless today. Created lots of globals, statics, forward declarations, hidden implementations in cpp files with no testable interface, erasing types, casting void pointers, I had to fix quite a lot and decouple the entangled mess.

Hopefully performance will pick up after the rollout.

Sounds like the researchers behind https://ai-2027.com/ haven't been too far off so far.

More importantly, this is the early steps of a model self improving itself.

Do we still think we'll have soft take off?

Is "high-capability" a stronger or weaker claim than "team of phd-level experts"?

https://www.nbcnews.com/tech/tech-news/openai-releases-chatg...

That’s just classical OpenAI trying to make us believe they’re closing on AGI… Like all « so called » research from them and Anthropic about safety alignment and that their tech is so incredibly powerful that guardrails should be put on them.

I heard the other day that every time someone claps another vibe coded project embeds the api keys in the webpage.

I wonder if this will continue to be the case.

>Our mitigations include safety training, automated monitoring, trusted access for advanced capabilities, and enforcement pipelines including threat intelligence.

"We added some more ACLs and updated our regex"

The labs have fully embraced the cutthroat competition, the arms race has fully shed the civilized facade of beneficient mutual cooperation.

Dirty tricks and underhanded tactics will happen - I think Demis isn't savvy in this domain, but might end up stomping out the competition on pure performance.

Elon, Sam, and Dario know how to fight ugly and do the nasty political boardroom crap. 26 is gonna be a very dramatic year, lots of cinematic potential for the eventual AI biopics.

They're also coordinating around Chinese New Year to compete with new releases of the major open/local models.

This goes way back. When OpenAI launched GPT-4 in 2023, both Anthropic and Google lined up counter launches (Claude and Magic Wand) right before OpenAI's standard 10am launch time.

The thrill of competition

Wouldn't that be illegal ? i.e. cartel to collude like that ?

A sign of the inevitible implosion !

I wish they’d just stop pretending to care about safety, other than a few researchers at the top they care about safety only as long as they aren’t losing ground to the competition. Game theory guarantees the AI labs will do what it takes to ensure survival. Only regulation can enforce the limits, self policing won’t work when money is involved.

> that hasn't been solved before a bunch of times in publicly available code?

And this matters because? Most devs are not working on novel never before seen problems.

In the 1930s, when electronic calculators were first introduced, there was a widespread belief that accounting as a career was finished. Instead, the opposite became true. Accounting as a profession grew, becoming far more analytical/strategic than it had been previously.

You are correct that these models primarily address problems that have already been solved. However, that has always been the case for the majority of technical challenges. Before LLMs, we would often spend days searching Stack Overflow to find and adapt the right solution.

Another way to look at this is through the lens of problem decomposition as well. If a complex problem is a collection of sub-problems, receiving immediate solutions for those components accelerates the path to the final result.

For example, I was recently struggling with a UI feature where I wanted cards to follow a fan-like arc. I couldn't quite get the implementation right until I gave it to Gemini. It didn't solve the entire problem for me, but it suggested an approach involving polar coordinates and sine/cosine values. I was able to take that foundational logic turn it into a feature I wanted.

Was it a 100x productivity gain? No. But it was easily a 2x gain, because it replaced hours of searching and waiting for a mental breakthrough with immediate direction.

There was also a relevant thread on Hacker News recently regarding "vibe coding":

https://news.ycombinator.com/item?id=45205232

The developer created a unique game using scroll behavior as the primary input. While the technical aspects of scroll events are certainly "solved" problems, the creative application was novel.

Yeah, I would LOVE to see attempts at significant video games that are then open-sourced for communities to work on. E.g. OpenGTA or OpenFIFA/OpenNHL.

Well, it took opus 4.5 five messages to solve a trivial git problem for me. It hallucinated nonexistent flags three times. Hallucinating nonexistent flags is certainly a novel solution to my git ineptness.

Not to be outdone, chatgpt 5.2 thinking high only needed about 8 iterations to get a mostly-working ffmpeg conversion script for bash. It took another 5 messages to translate it to run in windows, on powershell (models escaping newlines on windows properly will be pretty nuch AGI, as far as I’m concerned).

Why even come to this site if you're so anti-innovation?

Today with LLMs you can literally spend 5 minutes defining what you want to get, press send, go grab a coffee and come back to a working POC of something, in literally any programming language.

This is literally stuff of wonders and magic that redefines how we interface with computers and code. And the only thing you can think of is to ask if it can do something completely novel (that it's so hard to even quantity for humans that we don't have software patents mainly for that reason).

And the same model can also answer you if you ask it about maths, making you an itinerary or a recipe for lasagnas. C'mon now.

Can you point me to a human written program an LLM cannot write? And no, just answering with a massively large codebase does not count because this issue is temporary.

Some people just hate progress.

Great question, here is the link from the future:

I just wanted to say that's a pretty cool demo! I hadn't realised people were using it for things like this.

yea but i feel like we are over the hill on benchmaxxing, many times a model has beaten anthropic on a specific bench, but the 'feel' is that it is still not as good at coding

Benchmarks are useless compared to real world performance.

Real world performance for these models is a disappoint.

I only switched to using the terminal based agents in the last week. Prior to this I was pretty much only using it through Cursor and GH Copilot. The Anthropic models when used through GH Copilot were far superior to the codex ones and I didn't really get the hype of Codex. Using them through the CLI though, Codex is much better, IMO.

My guess is that it's potentially that and just momentum from developers who started using CC when it was far superior to Codex has allowed it to become so much more popular. Potentially, it's might be that, as it's more autonomous, it's better for true vibe-coding and it's more popular with the Twitter/LinkedIn wantrepreneur crew which meant it gets a lot of publicity which increases adoption quicker.

I'm with you. Codex's plans seems to be a much more generous offering than Claude

I just.. can't tell a different in quality between them.. so I go for the cheapest

Codex is great and I hit the usage once doing multiagent full 5 hour absolute degen session for the nornal workflow alongside never hit it and now x2 useage even and now with the planmode switch back and forth absolute great.

AI is mostly garbage at creating useful abstractions. I'd feel threatened if I was a competitive programmer or IMO kid

No. AI does not work well enough, you still need a person to look on it and CODE. It probably never will, until AGI which probably also in my opinion will never come.

I would feel threatened if I didn't invest in learning how to best use AI

Jevons paradox hints that the situation is not as bleak as it sounds.

I wonder if their "5.3" was continuously being updated, with regenerated benchmarks with each improvement, and they just stayed ready to release it when claude released

It’s also functionally not likely without some sort of insider knowledge or coordination

Yeah they are both fighting for survival. No surprise really.

Need to keep the hype going if they are both IPO'ing later this year.

What happened to you?

What's funny is that most of this "progress" is new datasets + post-training shaping the model's behavior (instruction + preference tuning). There is no moat besides that.

As coding agents get "good enough" the next differentiator will be which one can complete a task in fewer tokens.

Literally just find and replace

All of them can do it but Codex has the least frustrating usage limits.

Doesn't matter which one. All of them can do things like this now, given a good enough feedback loop. Which your problem has.

The notes explicitly call out you may want to dial the effort setting back to medium to reduce latency/tokens (high being default, apparently there is a max setting too).

From Claude 4.6 Thinking:

OSWorld is the full 369-task benchmark. OSWorld Verified is a ~200-task subset where humans have confirmed the eval scripts reliably score success/failure — the full set has some noisy grading where correct actions can still get marked wrong.

Scores on Verified tend to run higher, so they're not directly comparable.

I think codex got much faster for smaller tasks in the last few months. Especially if you turn thinking down to medium.

I think the slow feeling is a UI thing in codex

I just asked Opus 4.6 to debug a bug in my current changes and it went for 20 minutes before I interrupted it. Take that as you will.

How so?

Isn't the best eval the one you build yourself, for your own use cases and value production?

I encourage people to try. You can even timebox it and come up with some simple things that might look initially insufficient but that discomfort is actually a sign that there's something there. Very similar to moving from not having unit/integration tests for design or regression and starting to have them.

I usually wait to see what ArtificialAnalysis says for a direct comparison.

It's better on a benchmark I've never heard of!? That is groundbreaking, I'm switching immediately!

I’m hoping they add better IDE integration to track active file and selection. That’s the biggest annoyance I have in working with Codex.

Honest question: have you tried evolving your code architecture when adding features instead of just "promting more and more things"?

> What if one day AI will fall down and there will be no real programmers to write the software.

What if you want to write something very complex now that most people don't understand? You keep offering more money until someone takes the time to learn it and accomplish it, or you give up.

I mean, there are still people that hammer out horseshoes over a hot fire. You can get anything you're willing to pay money for.

Sorry but companies will not hire you but instead a person who learned how to code with AI. Get with the times or lose.

OpenAI has a whole history of trying to scoop other providers. This was a whole thing for Google launches, where OpenAI regularly launched something just before Google to grab the media attention.

I assume some sort of corporate espionage. This is high stakes after all

Tell me that you are hurt without telling me that you are hurt this applies to Sam right now

"But we anonymize prompts before training!"

Meanwhile the prompt: Crop this photo of my passport

Oh yeah that’s in the “These Are The Illegal Things We Did” section 7.4 in the Model Card.

Today is the best day to rewrite everything in React. You may not enjoy React, but AI agents do. And they are the ones writing the code.

5.2 was already very good with svelte 5, at least when you have the svelte MCP server set up.

You can also use via Opencode Zen, Github Copilot, or probably any number of other model providers that Opencode integrates with.

Not sure why everyone stays focused on getting it from Anthropic or OpenAI directly when there are so many places to get access to these models and many others for the same or less money.

OpenAI models in general, yes - `opencode auth login`, select OpenAI, then ChatGPT Pro/Plus. I just checked and 5.3-codex isn't available in opencode yet, but I assume it will be soon.

I've tried opus 4.5 in opencode via the GitHub Copilot API, mostly to see if it works all. I don't think that broke any terms of service? But also I haven't checked how much more expensive I made it for myself over just calling them directly.

You can use Anthropic models in Opencode, make an api key and you're good to do(you can even use the in house Opencode router, Zen).

What you can't do is pretend opencode is claude code to make use of that specific claude code subscription.

Yes, OpenAI said they'd allow usage of their subscriptions in opencode.

You should look into Kilo Pass by Kilo Code (https://kilo.ai/features/kilo-pass). It's basically a fixed subscription and your credits roll over each month, and you get free extra credits too which are used up first before paid credits. It's similar to paying for Cursor except the credits roll over which is why I'm contemplating moving to it, because I don't want to be locked into any one LLM provider the way Claude Code or Codex make you become.

I'm coding about 6-9h per day with Codex CLI on the $20 Plus sub, occasionally switching to extra-high reasoning and feeding it massive contexts, all tools enabled, sometimes 2-3 terminal sessions running in parallel and I've never hit limits... I operate on small-ish codebases but even so I try to work in the most local scope possible with AGENTS.md at the sub-directory levels.

Are you really hitting limits, or are you turned off by the fact you think you will?

I use Codex in OpenCode through the API and find the experience quite enjoyable.

I guess the jump is on purpose. You can buy Codex credits and also use codex via the API (manual switching required).

It's not in the API yet - "We are working to safely enable API access soon.", but I assume the rate-limits won't be worse than for 5.2 Codex.

gemini-3-flash-preview will be GA soon i hope. /s

My codex CLI didn’t notice version bump available, but I manually did pnpm add -g @openai/codex and 5.3 was there after.

Because Claude Code is stealing the thunder so OpenAI is focusing on coding now.

Why is it absurd?

I agree, I was confused about where 5.3 non Codex was. 5.2-Codex disappointed me enough that I won't be giving 5.3 Codex a try, but I'm looking forward to trying 5.3 non Codex with Pi.

Why is it loser? He very well could be a sore winner here.

They both are doing this to each other.

BTW, loser is spelled with a single o.

You could also claim that Anthropic is trying to scoop OpenAI by launching minutes earlier, as OpenAI has done with Google in the past.

For downvoters, you must be naive to think these companies are not surveilling each other through various means.