I've been extremely impressed (and actually had quite a good time) with GPT-5 and Codex so far. It seems to handle long context well, does a great job researching the code, never leaves things half-done (with long tasks it may leave some steps for later, but it never does 50% of a step and then just randomly mock a function like Gemini used to), and gives me good suggestions if I'm trying to do something I shouldn't. And the Codex CLI also seems to be getting constant, meaningful updates.
Agreed. We're hardcore Claude Code users and my CC usage trended down to zero pretty quickly after I started using Codex. The new model updates today are great. Very well done OpenAI team!! CC was an existential threat. You responded and absolutely killed it. Your move Anthropic.
To be fair, Anthropic kinda did this to themselves. I consider it as a pretty massive throw on their end in terms of the fairly tight grasp they had on developer sentiment.
Everyone else slowly caught up and/or surpassed them while they simultaneously had quality control issues and service degradation plaguing their system - ALL while having the most expensive models comparatively in terms of intelligence.
Agreed. I really wish Google would get their act together because I think they have the potential of being faster, cheaper with bigger context windows. They're so great at hardcore science and engineering, but they absolutely suck at products.
I think this is being downvoted coz it doesn't seem to be really responding to the thread, and maybe it isn't, but for anyone who hasn't tried Gemini CLI:
My experience after a month or so of heavy use is exactly this. The AI is rock solid. I'm pretty consistently impressed with its ability to derive insights from the code, when it works. But the client is flaky, the backend is flaky, and the overall experience for me is always "I wish I could just use Claude".
Say 1 in 10 queries craps out (often the client OOMs even though I have 192Gb of RAM). Sounds like a 10% reliability issue but actually it just pushes me into "fuck this I'll just do it myself" so it knocks out like 50% of the value of the product.
(Still, I wouldn't be surprised if this can be fixed over the next few months, it could easily be very competitive IMO).
Semi-related but I have the same experience with the gemini mobile app on android. ChatGPT and Claude are both great user experiences and the best word to describe how the gemini app feels is flaky.
I really do not want Google to win anything. They're a giant monopoly across multiple industries. We need a greater balance of power.
Antitrust enforcement has been letting us down for over two decades. If we don't have an oxygenation event, we'll go an entire generation where we only reward tax-collecting, non-innovation capital. That's unhealthy and unfair.
Our career sector has been institutionalized and rewards the 0.001% even as they rest on their laurels and conspire to suppress wages and innovation. There's a reason why centicorns petered out and why the F500 is tech-heavy. It's because big tech is a dragnet that consumes everything it touches - film studios, grocery stores, and God only knows what else it'll assimilate in the unending search for unregulated, cancerous growth.
FAANG's $500k TC is at the expense of hundreds of unicorns making their ICs even wealthier. That money mostly winds up going to institutional investors, where the money sits parked instead of flowing into huge stakes risks and cutthroat competition. That's why a16z and YC want to see increased antitrust regulations.
But it's really bad for consumers too. It's why our smartphones are stagnant taxation banana republics with one of two landlords. Nothing new, yet as tightly controlled an authoritarian state. New ideas can't be tried and can't attain healthy margins.
It's wild that you can own a trademark, but the only way for a consumer to access it is to use a Google browser that defaults to Google search (URLs are scary), where the search results will be gamed by competitors. You can't even own your own brand anymore.
Winning shouldn't be easy. It should be hard. A neverending struggle that rewards consumers.
Google has been, for at least a decade, making pretty terrible choices that squander developer and power-user goodwill (see: any thread where they announce a new product and one of the top comments will link to killedbygoogle). When you've burnt bridges with your biggest evangelists, adoption by normies slows, and your products appear to stagnate.
Unfortunately, they've been insulated from the consequences of their bad decisions by the fact the money printer (ads) keeps their company afloat and mollifies shareholders. The moment that dries up, they're in trouble.
- The smartest model I have used. Solves problems better than Opus-4.1.
- It can be lazy. With Claude Code / Opus, once given a problem, it will generally work until completion. Codex will often perform only the first few steps and then ask if I want to continue to do the rest. It does this even if I tell it to not stop until completion.
- I have seen severe degradation near max context. For example, I have seen it just repeat the next steps every time I tell it to continue and I have to manually compact.
I'm not sure if the problems are Gpt-5 or Codex. I suspect a better Codex could resolve them.
Claude seems to have gotten worse for me, with both that kind of laziness and a new pattern where it will write the test, write the code, run the test, and then declare that the test is working perfectly but there are problems in the (new) code that need to be fixed.
Context degradation is a real problem with all frontier LLMs. As a rule of thumb I try to never exceed 50% of available context window when working with either Claude Sonnet 4 or GPT-5 since the quality drops really fast from there.
I've never seen that level of extreme degradation (just making a small random change and repeating the same next steps infinitely) on Claude Code. Maybe Claude Code is more aggressive about auto compaction. I don't think Codex even compacts without /compact.
I think some of it is not necessarily auto compaction but the tooling built in. For example claude code itself very frequently builds in to remind the model what its working on and should be doing which helps always keeps its tasks in the most recent context, and overall has some pretty serious thought put into its system prompt and tooling.
But they have suffered quite a lot of degradation and quality issues recently.
To be honest unless Anthropic does something very impactful sometime soon I think they're losing their moat they had with developers as more and more jump to codex and other tools. They kind of massively threw their lead imo.
I definitely agree with all of those points. I just really prefer it completing steps and asking me if we should continue to next step rather than doing half of the step and telling me it's done. And the context degradation seems quite random - sometimes it hits way earlier, sometimes we go through crazy amount of tokens and it all works out.
I also noticed the laziness compared to Sonnet models but now I feel it’s a good feature. Sonnet models, now I realize, are way too eager to hammer out code with way more likelihood of bugs.
Same. We're not hitting limits at all with Codex and it's ridiculously good at managing and preserving its context window while getting a metric fuckton of work done. It's kind of unbelievable actually. I don't know re billing. Not my dept.
It doesn't seem to have any internal tools it can use. For example, web search; It just runs curl in the terminal. Compared to Gemini CLI that's rough but it does handle pasting much better... Maybe I'm just using both wrong...
It does have web search - it's just not enabled by default. You can enable it with --search or in the config, then it can absolutely search, for example finding manuals/algorithms.
This should probably be merged with the other GPT-5-Codex thread at https://news.ycombinator.com/item?id=45252301 since nobody in this thread is talking about the system card addendum.
It would be nice if this model would be good enough to update their typscript sdk (+agents library) to use, or at least support, zod v4 - they still use v3.
Had to spend quite a long time to figure out a dependency error...
Codex with GPT-5-High is extremely good. Like many I was a bit "meh" about the GPT 5 release, however once I started using it with Codex it became clear there was a substantial improvement in a capability I wasn't really paying attention to, which is tool calling. Or more specifically, when to call a tool. Ask GPT-5-High a question about your codebase and watch the things it looks for, and things it searches for (if you use --search). It has very good taste on how to navigate and solve a problem.
this + any coding conventions should ALWAYS be a post process. DO NOT include them in your prompt, you are losing model accuracy over these tiny things.
It helps to actually be able to read the diffs of its proposals/changes in the terminal. The changing from tabs -> spaces on every line it touches generally results in unreadable messes.
I have a pretty complex project, so I need to keep an eye on it to ensure it doesn't go off the rails and delete all the code to get a build to pass (it wouldn't be the first time).
I think the idea is that your IDE or whatever should automatically run the project's autoformatter after every AI edit, so that any formatting mistakes the AI makes are fixed before you have to look at them.
> Reason #3a: Work with the model biases, not against
Another note on model biases is that you should lean into them. The tricky part with this is the only way to figure out a model's defaults is to have actual usage and careful monitoring (or have evals that let you spot it).
Instead of forcing the model to behave in ways it ignores, adapt your prompts and post-processing to embrace its defaults. You'll save tokens and get better results.
If the model keeps hallucinating some JSON fields, maybe you should support (or even encourage) those fields instead of trying to prompt the model against them.
Go uses tabs. Full stop. There is no Go code with spaces. Not if they're using the built-in formatter, anyway. In any case, this is about the diff codex is outputting, not the code I commit. With Claude, I generally don't need to run `go fmt`, but with codex, it is absolutely necessary.
Does codex have a good way of doing post process hooks? For Claude Code hooks I never found a way to run a formatter over only the file that was edited. It’s super annoying as I want to constantly have linting and formatting cleaned up right after the model finishes editing a file…
Does Codex have token-hiding (cf Anthropic’s “subagents”)?
I was tempted to give Codex a try but a colleague was stung by their pricing. Apparently if you go over your Pro plan allocation, they just quietly and automatically start billing you per-token?
I tried Codex with the $20/month plan recently and it did exactly what Claude Code does, stop and tell you “sorry, you’re out of credit, come back in x days.”
Hey, I work on Codex—absolutely no way that a user on a Pro plan would somehow silently move to token-based billing. You just hit a limit and have to wait for the reset. (Which also sucks, and which we're also improving early warnings of.)
Feels slower than GPT-5 and I understood it that medium should be a lot faster than high but for me it's almost the same , so I don't see a reason preferring medium.
The new GPT-5-Codex model isn't yet available in the API, so if you want to try that model using the Codex CLI tool the only way to do that is with a ChatGPT account (I'm more sure if the free account has it, the $20/month definitely does). You need to then authenticate Codex CLI with ChatGPT.
OpenAI say API access to that model is coming soon, at which point till be able to use it in Codex CLI with an API key and pay for tokens as you go.
You can also use the Codex CLI tool without using the new GPT-5-Codex model.
You can use Codex CLI with an API key instead of a subscription, but then you won't have access to this new GPT-5 Codex model, since it's not out on the API yet. But normal GPT-5 in Codex is perfectly fine.
That's what was in my OP, I used the API approach, but the rate limits were insanely low (seemingly?) the agent died after 3 steps in a single question.
I think it would be cool to see *nix “emulation” integrated into coding AIs. I don’t think it’s necessary to run these agents inside of container as most people are right now. That’s a lot of overhead.
You mean instead of them running the code that they are writing they pretend to run the code and the model shows what it thinks would happen?
I don't like that at all. Actually running the code is the single most effective protection we have against coding mistakes, from both humans and machines.
I think it's absolutely worth the complexity and performance overhead of hooking up a real container environment.
Not to mention you can run a useful code execution container in 100MB of RAM on a single CPU (or slice thereof). Simulating that with an LLM takes at least one GPU and 100GB or more of VRAM.
I understand your point but I basically find myself running all my agents in barebones containers and they’re basically short-run make-or-kill types. And once we ramp up agent counts, possibly into the thousands, that could add up rapidly. Of course, you would run milestone tests on actual container/envs but I think there might be a need for lighter solutions for rapid agent dev runs.
There are now many solutions, and full-blown startups, under the "swarm", "agent orchestration" and other similar keywords, for spinning agents in the cloud. I'm not sure if that's what you mean, but I totally see most of vibe coding being replaced by powerhouse agents, placed locally or in the cloud, picking up tasks and working them out until its really done.
You do realize that there is virtually no overhead in running containers, right? That's the entire point of their existence. They're just processes, with specific permissions (to generalize it). Your computer can run thousands of processes without sweating.
Interesting, the new model uses a different prompt in Codex CLI that's ~half the size (10KB vs. 23KB) of the previous prompt[0][1].
SWE-bench performance is similar to normal gpt-5, so it seems the main delta with `gpt-5-codex` is on code refactors (via internal refactor benchmark 33.9% -> 51.3%).
As someone who recently used Codex CLI (`gpt-5-high`) to do a relatively large refactor (multiple internal libs to dedicated packages), I kept running into bugs introduced when the model would delete a file and then rewrite it (missing crucial or important details). My approach would have been to just the copy the file over and then make package-specific changes, so maybe better tool calling is at play here.
Additionally, they claim the new model is more steerable (both with AGENTS.md and generally).
In my experience, Codex CLI w/gpt-5 is already a lot more steerable than Claude Code, but any improvements are welcome!
What worked was getting it to first write a detailed implementation plan for a “junior contractor” then attempt it in phases (clearing task window each time) and told to use /tmp to copy files and transform them then update the original.
Looking forward to trying the new model out on the next refactor!
Yes, regardless of tool, I always create a separate plan doc for larger changes
Will try adding the instructions specific to refactors (i.e. copy/move files, don't rewrite when possible)
I've also found it helpful, especially for certain regressions, to basically create a new branch for any Codex/CC assisted task (even if part of a larger task). Makes it easier to identify regressions due to recent changes (i.e. look at git diff, it worked previously)
Telling the "agent" to manage git leads to more context pollution than I want, so I manage all commits/branches myself, but I'm sure that will change as the tools improve/they do more RL on full-cycle software dev
Everyone else slowly caught up and/or surpassed them while they simultaneously had quality control issues and service degradation plaguing their system - ALL while having the most expensive models comparatively in terms of intelligence.
My experience after a month or so of heavy use is exactly this. The AI is rock solid. I'm pretty consistently impressed with its ability to derive insights from the code, when it works. But the client is flaky, the backend is flaky, and the overall experience for me is always "I wish I could just use Claude".
Say 1 in 10 queries craps out (often the client OOMs even though I have 192Gb of RAM). Sounds like a 10% reliability issue but actually it just pushes me into "fuck this I'll just do it myself" so it knocks out like 50% of the value of the product.
(Still, I wouldn't be surprised if this can be fixed over the next few months, it could easily be very competitive IMO).
Antitrust enforcement has been letting us down for over two decades. If we don't have an oxygenation event, we'll go an entire generation where we only reward tax-collecting, non-innovation capital. That's unhealthy and unfair.
Our career sector has been institutionalized and rewards the 0.001% even as they rest on their laurels and conspire to suppress wages and innovation. There's a reason why centicorns petered out and why the F500 is tech-heavy. It's because big tech is a dragnet that consumes everything it touches - film studios, grocery stores, and God only knows what else it'll assimilate in the unending search for unregulated, cancerous growth.
FAANG's $500k TC is at the expense of hundreds of unicorns making their ICs even wealthier. That money mostly winds up going to institutional investors, where the money sits parked instead of flowing into huge stakes risks and cutthroat competition. That's why a16z and YC want to see increased antitrust regulations.
But it's really bad for consumers too. It's why our smartphones are stagnant taxation banana republics with one of two landlords. Nothing new, yet as tightly controlled an authoritarian state. New ideas can't be tried and can't attain healthy margins.
It's wild that you can own a trademark, but the only way for a consumer to access it is to use a Google browser that defaults to Google search (URLs are scary), where the search results will be gamed by competitors. You can't even own your own brand anymore.
Winning shouldn't be easy. It should be hard. A neverending struggle that rewards consumers.
We need a forest fire to renew the ecosystem.
- all the users
- all the apps (Google, GMail, YouTube, Docs, Maps...)
- all the books (Google Books)
- all the video (YouTube)
- all the web pages
- custom hardware
It's honestly weird they aren't doing better. Agree that the models are great and the UX is bad all around.
Unfortunately, they've been insulated from the consequences of their bad decisions by the fact the money printer (ads) keeps their company afloat and mollifies shareholders. The moment that dries up, they're in trouble.
I don't think they care what we think. They're thriving despite our protests.
But yeah, they shouldn't be shielded from antitrust. They have literally everything.
It's super annoying that it doesn't provide a way to approve edits one by one instead it either vibe codes on its own or gives me diffs to copy paste.
Claude code has a much saner "normal mode".
Gemini cli is too inconsistent, good for documentation tasks. Don’t let it write code for you
- The smartest model I have used. Solves problems better than Opus-4.1.
- It can be lazy. With Claude Code / Opus, once given a problem, it will generally work until completion. Codex will often perform only the first few steps and then ask if I want to continue to do the rest. It does this even if I tell it to not stop until completion.
- I have seen severe degradation near max context. For example, I have seen it just repeat the next steps every time I tell it to continue and I have to manually compact.
I'm not sure if the problems are Gpt-5 or Codex. I suspect a better Codex could resolve them.
Very frustrating, and happening more often.
But they have suffered quite a lot of degradation and quality issues recently.
To be honest unless Anthropic does something very impactful sometime soon I think they're losing their moat they had with developers as more and more jump to codex and other tools. They kind of massively threw their lead imo.
have to say not sure what this even means and what the exact definition of a message is in this context.
with claude code max20 I was constantly hitting limits, with codex not once yet
Had to spend quite a long time to figure out a dependency error...
Updates (v0.36) https://github.com/openai/codex/releases
I have a pretty complex project, so I need to keep an eye on it to ensure it doesn't go off the rails and delete all the code to get a build to pass (it wouldn't be the first time).
In fact you should convert your code to spaces at least before LLM sees it. It’ll improve your results by looking more like its training data.
> Reason #3a: Work with the model biases, not against
Another note on model biases is that you should lean into them. The tricky part with this is the only way to figure out a model's defaults is to have actual usage and careful monitoring (or have evals that let you spot it).
Instead of forcing the model to behave in ways it ignores, adapt your prompts and post-processing to embrace its defaults. You'll save tokens and get better results.
If the model keeps hallucinating some JSON fields, maybe you should support (or even encourage) those fields instead of trying to prompt the model against them.
https://cdn.openai.com/pdf/97cc5669-7a25-4e63-b15f-5fd5bdc4d...
I was tempted to give Codex a try but a colleague was stung by their pricing. Apparently if you go over your Pro plan allocation, they just quietly and automatically start billing you per-token?
Even shorter version:
- New coding-specialist model called GPT-5-Codex, coming soon to the API but for now available in their Codex CLI, VS Code and Codex Cloud products
- New code review product (part of Codex Cloud) that can review PRs for you
- New model promises better code review, less pointless comments and can vary its reasoning effort for simple vs complex tasks
So a bit in line with what Theo mentioned in his video that he was not happy with the ui capabilities
This model is available inside all OpenAI codex products. Yet to be available on Api
The model is supposed to be better at code reviews and Comments than the other GPT-5 variant. It can also think/work upto 7 hours.
But when I installed Codex and tried to make a simple code bugfix, I got rate limited nearly immediately. As in, after 3 "steps" the agent took.
Are you meant to only use Codex with their $200 "unlimited" plans? Thanks!
OpenAI say API access to that model is coming soon, at which point till be able to use it in Codex CLI with an API key and pay for tokens as you go.
You can also use the Codex CLI tool without using the new GPT-5-Codex model.
I don't like that at all. Actually running the code is the single most effective protection we have against coding mistakes, from both humans and machines.
I think it's absolutely worth the complexity and performance overhead of hooking up a real container environment.
Not to mention you can run a useful code execution container in 100MB of RAM on a single CPU (or slice thereof). Simulating that with an LLM takes at least one GPU and 100GB or more of VRAM.
SWE-bench performance is similar to normal gpt-5, so it seems the main delta with `gpt-5-codex` is on code refactors (via internal refactor benchmark 33.9% -> 51.3%).
As someone who recently used Codex CLI (`gpt-5-high`) to do a relatively large refactor (multiple internal libs to dedicated packages), I kept running into bugs introduced when the model would delete a file and then rewrite it (missing crucial or important details). My approach would have been to just the copy the file over and then make package-specific changes, so maybe better tool calling is at play here.
Additionally, they claim the new model is more steerable (both with AGENTS.md and generally).
In my experience, Codex CLI w/gpt-5 is already a lot more steerable than Claude Code, but any improvements are welcome!
[0]https://github.com/openai/codex/blob/main/codex-rs/core/gpt_...
[1]https://github.com/openai/codex/blob/main/codex-rs/core/prom...
(comment reposted from other thread)
What worked was getting it to first write a detailed implementation plan for a “junior contractor” then attempt it in phases (clearing task window each time) and told to use /tmp to copy files and transform them then update the original.
Looking forward to trying the new model out on the next refactor!
Will try adding the instructions specific to refactors (i.e. copy/move files, don't rewrite when possible)
I've also found it helpful, especially for certain regressions, to basically create a new branch for any Codex/CC assisted task (even if part of a larger task). Makes it easier to identify regressions due to recent changes (i.e. look at git diff, it worked previously)
Telling the "agent" to manage git leads to more context pollution than I want, so I manage all commits/branches myself, but I'm sure that will change as the tools improve/they do more RL on full-cycle software dev