OpenAI likes to time their announcements alongside major competitor announcements to suck up some of the hype. (See for instance the announcement of GPT-4o a single day before Google's IO conference)
They were probably sitting on this for a while. That makes me think this is a fairly incremental update for Codex.
I still want something no one has, which is the ability to launch agents in different git worktrees simultaneously and check the results out on my main branch for testing when they are finished.
I am curious: why would you you like to have that? (Genuine question, I am personally so scared about the AI going crazy and putting slop everywhere that I often ask it to focus on a single well defined area first)
so this was arctic fox it seems, lot of us ended up downgrading to codex 5.0 because of the token burn was too much, i see codex max is a step up which is welcome but still unsure if they solved that github issue around tool use that impacts tokens
going to wait and see after being burned by 5.1 before i upgrade back to 0.58
gemini 3 has been a let down tbh to see agentic coding wasn't a top priority
im sticking with codex for now and using gemini 3 for frontend
"Starting today, GPT‑5.1-Codex-Max will replace GPT‑5.1-Codex as the default model in Codex surfaces."
Wow, I spent last weekend using a tag-team of Claude and Codex and found Codex to more often get better results (TypeScript physics/graphics application). I probably only wrote a few hundred lines of code out of many thousands; it did a really good job.
Now I guess I'll ask the new Codex to review the work of the old!
These 2 sentences right next to each other stood out to me:
> a new step towards becoming a reliable coding partner
> GPT‑5.1-Codex-Max is built for long-running, detailed work
Does this not sound contradictory? It’s been the shorter form work that has built what little confidence I have in these as a coding partner - a model that goes off and does work without supervision is not a partner to me.
> Compaction enables GPT‑5.1-Codex-Max to complete tasks that would have previously failed due to context-window limits, such as complex refactors and long-running agent loops by pruning its history while preserving the most important context over long horizons. In Codex applications, GPT‑5.1-Codex-Max automatically compacts its session when it approaches its context window limit, giving it a fresh context window. It repeats this process until the task is completed.
Wouldn't the model automatically do that using attention techniques? Why do you need to do it at the token layer and not leave it to the model to automatically decide which tokens are worth paying attention to?
My observation has been that Codex tends to hit logical/data-driven/back-end tasks out of the park while doing weird, random nonsense with even simple UI tasks. This could me needing to improve how I phrase my prompts, but it will be interesting to see if it's improved in that arena at all.
Somewhat related, after seeing the praise for codex in the Sonnet 4.5 release thread I gave it a go, and I must say, that CLI is much worse than Claude Code (even if the model is great, I’m not sure where the issue really lies between the two).
It was extremely slow (like, multiple times slower than Sonnet with Claude Code, though that’s partially on me for using thinking-high I guess) to finish the task, with the back-and-forths being on the order of tens of minutes.
Moreover, the context management seems to be really weird. I’m not sure how exactly it works, but - 1. It uses very little tokens / fills up the context slowly (good I guess) 2. Doesn’t seem to actually internalize the contents of files you mention to it, or it edits.
#2 here being the main one - I usually context-dump reference code for Claude Code, and it does a perfect job of adhering to codebase patterns and its architecture, while codex was completely ignorant of the existing code style.
Moreover, it wrote extremely defensive code, even for code where it wrote both ends itself.
All in all, I was really let down after seeing all the praise.
I would love to see all the big players put 1% of the effort they put into model training into making the basic process of paying and signing in suck less.
Claude: they barely have a signin system at all. Multiple account support doesn’t exist. The minimum seat count for business is nonsense. The data retention policies are weak.
OpenAI: Make ZDR a thing you can use or buy without talking to sales, already. And for those using containers or a remote system or really anything other than local development with the codex CLI, you really really need to fix this bug. I bet Codex could do at least the client part for you!
(Hint: Claude Code gets this right by default, despite the fact that everything else about Claude sign-in is a joke.)
Google: get all your B2B AI product managers in one room and tell them that they need to make one single product menu on one single webpage with all the pricing on that page and that the Google Cloud people are not permitted to make anything that isn’t actually logically Google Cloud depend on Google Cloud Billing. Your product cannot compete with OpenAI or Anthropic if people need to ask an LLM to figure out what your product is and if your own fancy LLMs can’t give a straight answer. My company pays for a non-Google product primarily because it’s too complicated to pay for the Google product! Right now, trying to use Google’s AI is like trying to ride Bay Area public transit before the Clipper Card.
I rarely used Codex compared to Claude because it was extremely slow in GitHub copilot
. Like maybe 2-5X slower than Claude Sonnet. I really wish they just made their models faster than “better”
This is a tangent: Has anyone noticed that GPT-5.0 at some point started producing much faster, crappier answers, then 5.1 made it slower + better again? (Both in Thinking mode)
I've been using a lot of Claude and Codex recently.
One huge difference I notice between Codex and Claude code is that, while Claude basically disregards your instructions (CLAUDE.md) entirely, Codex is extremely, painfully, doggedly persistent in following every last character of them - to the point that i've seen it work for 30 minutes to convolute some solution that was only convoluted because of some sentence I threw in the instructions I had completely forgotten about.
I imagine Codex as the "literal genie" - it'll give you exactly what you asked for. EXACTLY. If you ask Claude to fix a test that accidentally says assert(1 + 1 === 3), it'll say "this is clearly a typo" and just rewrite the test. Codex will rewrite the entire V8 engine to break arithmetic.
Both these tools have their uses, and I don't think one approach is universally better. Because Claude just hacks its way to a solution, it is really fast, so I like using it for iterate web work, where I need to tweak some styles and I need a fast iterative loop. Codex is much worse at that because it takes like 5 minutes to validate everything is correct. Codex is much better for longer, harder tasks that have to be correct -- I can just write some script to verify that what it did work, and let it spin for 30-40 minutes.
> If you ask Claude to fix a test that accidentally says assert(1 + 1 === 3), it'll say "this is clearly a typo" and just rewrite the test.
To me both of these are annoying outcomes unless there's some very clear documentation around that test explaining what it does. Ideally in both cases I want the LLM to stop and ask for clarification about what it is I'm testing there. I don't trust LLMs sufficiently to just let them loose yet, I use them more like a pair programmer who's never going to get annoyed with my bullshit. (So yes, I usually have them set to require approval on any edits, and will nitpick my way through them like the most annoying code reviewer you've ever met)
> If you ask Claude to fix a test that accidentally says assert(1 + 1 === 3), it'll say "this is clearly a typo" and just rewrite the test. Codex will rewrite the entire V8 engine to break arithmetic.
Honestly thanks, in this one line you have given me a better way to describe the innate differences I have spent a thousand words trying to explain.
Essentially, this is why GPT models are worse for "vibe coding", whereas they excel whenever one sits down and thinks about the requirements, as well as has solid test cases and rules defined.
I haven’t used Claude Code much, but I found Codex extremely frustrating. It doesn’t pay attention to anything in AGENTS.md, it’s completely incapable of removing code and is frustratingly defensive.
If you use it, the codebase constantly grows. Even when you explicitly instruct it to remove something, you always end up with more lines of code in the project than before the instruction. Also (I used it for Python and TypeScript) the code was littered with getattr(...), .get(...), isinstance(...), and TypeScript equivalents (typeof, ...). Even though I religiously type‑annotate everything.
(I really need a macro for this comment, I keep repeating it :D )
Claude is a pair programmer, you can interrupt it and keep track what it's doing. It's VERY results-oriented, aiming to be "done" as fast as possible. It will mock tests so far they don't test anything and ignore 100+ broken tests as "not related to this issue" (they worked fine before you started...). Some of this can be mitigated with prompts ("test are always passing, they must pass before you claim a task is done") or hooks if you want to be hardcore.
Codex is an outsourced Indian development team. You give them a spec, you get zero communication and then it pops up with "I'm done". Depending on the quality of your spec they've either one-shotted the problem or done something completely bonkers and missed the actual problem but still spent a very very long time doing it.
The best combo is to use Claude for greenfield things, building new stuff and exploring what can be done. Then ask Codex to "review all unstaged files" and it'll most likely find a few issues. Give that report to Claude and ask "do you agree with this review?" and have it fix the ones all three agree (you, Claude and Codex).
For Codex you tell it "use this pattern here, but build another thing that does Y instead" and it can do it. It's also very good at rewriting small stuf from one language to another (I've tested this multiple times with Bash->Python and Python->Go)
Late, but reading all of the replies, and speaking from my own observation using Claude, Codex, as well as (non-CLI) Gemini, Kimi, Qwen, and Deepseek...
It's fun how we are so quick to assign meaning to the way these models act. This is of course due to training, RLHF, available tool calls, system prompt (all mostly invisible) and the way we prompt them.
I've been wondering about a new kind of benchmark how one would be able to extract these more intangible tendencies from models rather than well-controlled "how good at coding is it" style environments. This is mainly the reason why I pay less and less attention to benchmark scores.
For what it's worth: I still best converse with Claude when doing code. Its reasoning sounds like me, and it finds a good middle ground between conservative and crazy, being explorative and daring (even although it too often exclaims "I see the issue now!"). If Anthropic would lift the usage rates I would use it as my primary. The CLI tool is also better. E.g. Codex with 5.1 gets stuck in powershell scripts whilst Claude realizes it can use python to do heavy lifting, but I think that might be largely due to being mainly on Windows (still, Claude does work best, realizing quickly what environment it lives in rather than trying Unix commands or powershell invocations that don't work because my powershell is outdated).
Qwen is great in an IDE for quick auto-complete tasks, especially given that you can run it locally, but even the VSCode copilot is good enough for that. Kimi is promising for long running agentic tasks but that is something I've barely explored and just started playing with. Gemini is fantastic as a research assistant. Especially Gemini 3 Pro points out clear and to the point jargon without fear of the user being stupid, which the other commercial models are too often hesitant to do.
Again, it would be fun to have some unbiased method to uncover some of those underlying persona's.
If you want to try out other models try opencode. Right now grok is free to use. I am using it now. I think its a little better than codex or Claude. But it's so so much faster. Gemini 3 can also be used, but is often overloaded.
66 comments
[ 6.4 ms ] story [ 69.9 ms ] threadThey were probably sitting on this for a while. That makes me think this is a fairly incremental update for Codex.
going to wait and see after being burned by 5.1 before i upgrade back to 0.58
gemini 3 has been a let down tbh to see agentic coding wasn't a top priority im sticking with codex for now and using gemini 3 for frontend
Wow, I spent last weekend using a tag-team of Claude and Codex and found Codex to more often get better results (TypeScript physics/graphics application). I probably only wrote a few hundred lines of code out of many thousands; it did a really good job.
Now I guess I'll ask the new Codex to review the work of the old!
> a new step towards becoming a reliable coding partner
> GPT‑5.1-Codex-Max is built for long-running, detailed work
Does this not sound contradictory? It’s been the shorter form work that has built what little confidence I have in these as a coding partner - a model that goes off and does work without supervision is not a partner to me.
Wouldn't the model automatically do that using attention techniques? Why do you need to do it at the token layer and not leave it to the model to automatically decide which tokens are worth paying attention to?
- New benchmark SOTAs with 77.9% on SWE-Bench-Verified, 79.9% on SWE-Lancer, and 58.1% on TerminalBench 2.0
- Natively trained to work across many hours across multiple context windows via compaction
- 30% more token-efficient at the same reasoning level across many tasks
Let us know what you think!
It was extremely slow (like, multiple times slower than Sonnet with Claude Code, though that’s partially on me for using thinking-high I guess) to finish the task, with the back-and-forths being on the order of tens of minutes.
Moreover, the context management seems to be really weird. I’m not sure how exactly it works, but - 1. It uses very little tokens / fills up the context slowly (good I guess) 2. Doesn’t seem to actually internalize the contents of files you mention to it, or it edits.
#2 here being the main one - I usually context-dump reference code for Claude Code, and it does a perfect job of adhering to codebase patterns and its architecture, while codex was completely ignorant of the existing code style.
Moreover, it wrote extremely defensive code, even for code where it wrote both ends itself.
All in all, I was really let down after seeing all the praise.
https://github.com/openai/codex/releases/tag/rust-v0.59.0
Claude: they barely have a signin system at all. Multiple account support doesn’t exist. The minimum seat count for business is nonsense. The data retention policies are weak.
OpenAI: Make ZDR a thing you can use or buy without talking to sales, already. And for those using containers or a remote system or really anything other than local development with the codex CLI, you really really need to fix this bug. I bet Codex could do at least the client part for you!
https://github.com/openai/codex/issues/2798
(Hint: Claude Code gets this right by default, despite the fact that everything else about Claude sign-in is a joke.)
Google: get all your B2B AI product managers in one room and tell them that they need to make one single product menu on one single webpage with all the pricing on that page and that the Google Cloud people are not permitted to make anything that isn’t actually logically Google Cloud depend on Google Cloud Billing. Your product cannot compete with OpenAI or Anthropic if people need to ask an LLM to figure out what your product is and if your own fancy LLMs can’t give a straight answer. My company pays for a non-Google product primarily because it’s too complicated to pay for the Google product! Right now, trying to use Google’s AI is like trying to ride Bay Area public transit before the Clipper Card.
One huge difference I notice between Codex and Claude code is that, while Claude basically disregards your instructions (CLAUDE.md) entirely, Codex is extremely, painfully, doggedly persistent in following every last character of them - to the point that i've seen it work for 30 minutes to convolute some solution that was only convoluted because of some sentence I threw in the instructions I had completely forgotten about.
I imagine Codex as the "literal genie" - it'll give you exactly what you asked for. EXACTLY. If you ask Claude to fix a test that accidentally says assert(1 + 1 === 3), it'll say "this is clearly a typo" and just rewrite the test. Codex will rewrite the entire V8 engine to break arithmetic.
Both these tools have their uses, and I don't think one approach is universally better. Because Claude just hacks its way to a solution, it is really fast, so I like using it for iterate web work, where I need to tweak some styles and I need a fast iterative loop. Codex is much worse at that because it takes like 5 minutes to validate everything is correct. Codex is much better for longer, harder tasks that have to be correct -- I can just write some script to verify that what it did work, and let it spin for 30-40 minutes.
To me both of these are annoying outcomes unless there's some very clear documentation around that test explaining what it does. Ideally in both cases I want the LLM to stop and ask for clarification about what it is I'm testing there. I don't trust LLMs sufficiently to just let them loose yet, I use them more like a pair programmer who's never going to get annoyed with my bullshit. (So yes, I usually have them set to require approval on any edits, and will nitpick my way through them like the most annoying code reviewer you've ever met)
Honestly thanks, in this one line you have given me a better way to describe the innate differences I have spent a thousand words trying to explain.
Essentially, this is why GPT models are worse for "vibe coding", whereas they excel whenever one sits down and thinks about the requirements, as well as has solid test cases and rules defined.
If you use it, the codebase constantly grows. Even when you explicitly instruct it to remove something, you always end up with more lines of code in the project than before the instruction. Also (I used it for Python and TypeScript) the code was littered with getattr(...), .get(...), isinstance(...), and TypeScript equivalents (typeof, ...). Even though I religiously type‑annotate everything.
Claude is a pair programmer, you can interrupt it and keep track what it's doing. It's VERY results-oriented, aiming to be "done" as fast as possible. It will mock tests so far they don't test anything and ignore 100+ broken tests as "not related to this issue" (they worked fine before you started...). Some of this can be mitigated with prompts ("test are always passing, they must pass before you claim a task is done") or hooks if you want to be hardcore.
Codex is an outsourced Indian development team. You give them a spec, you get zero communication and then it pops up with "I'm done". Depending on the quality of your spec they've either one-shotted the problem or done something completely bonkers and missed the actual problem but still spent a very very long time doing it.
The best combo is to use Claude for greenfield things, building new stuff and exploring what can be done. Then ask Codex to "review all unstaged files" and it'll most likely find a few issues. Give that report to Claude and ask "do you agree with this review?" and have it fix the ones all three agree (you, Claude and Codex).
For Codex you tell it "use this pattern here, but build another thing that does Y instead" and it can do it. It's also very good at rewriting small stuf from one language to another (I've tested this multiple times with Bash->Python and Python->Go)
It's fun how we are so quick to assign meaning to the way these models act. This is of course due to training, RLHF, available tool calls, system prompt (all mostly invisible) and the way we prompt them.
I've been wondering about a new kind of benchmark how one would be able to extract these more intangible tendencies from models rather than well-controlled "how good at coding is it" style environments. This is mainly the reason why I pay less and less attention to benchmark scores.
For what it's worth: I still best converse with Claude when doing code. Its reasoning sounds like me, and it finds a good middle ground between conservative and crazy, being explorative and daring (even although it too often exclaims "I see the issue now!"). If Anthropic would lift the usage rates I would use it as my primary. The CLI tool is also better. E.g. Codex with 5.1 gets stuck in powershell scripts whilst Claude realizes it can use python to do heavy lifting, but I think that might be largely due to being mainly on Windows (still, Claude does work best, realizing quickly what environment it lives in rather than trying Unix commands or powershell invocations that don't work because my powershell is outdated).
Qwen is great in an IDE for quick auto-complete tasks, especially given that you can run it locally, but even the VSCode copilot is good enough for that. Kimi is promising for long running agentic tasks but that is something I've barely explored and just started playing with. Gemini is fantastic as a research assistant. Especially Gemini 3 Pro points out clear and to the point jargon without fear of the user being stupid, which the other commercial models are too often hesitant to do.
Again, it would be fun to have some unbiased method to uncover some of those underlying persona's.
(Maybe it would be a waste of time.)