1 Million Context to enable professional workloads
title matches the content - anything below 1m is not usable pretty much . it is just a hassle and claude wins even if less capable. So why not 1m or more?
Kind regards
The part that irritates me the most about this is that when these models come out, their context limit is marketed like it’s real.
What kind of workloads do you guys have?
whether the context is 100k or 1M or 10M is illusory as @htperkins points out - at the end of the day it’s just the usable ‘internal’ ‘search’ space available to the model.
At the end of the day it’s not that different from embedding search - or RAG as the kids call it - wrapped into the model.
So from a product standpoint, it seems like it’s a balance between context length and capability. Pound for pound you can’t really have both.
The operational method to deal with massive corpuses is to provide the agent with search tools, and the knowledge that the knowledge is available - instead of just loading everything in.
So, depending on your workload, it might make sense to task codex to create a search and indexing tool, or perhaps simply to generate a summary index of what data you want it to know about it.
C# coding in a small to mid sized repository just done by myself - not even multi user (like you would have in a business environment).
256K is sometimes filled on loading up a new task and go back an forth 2-3 times with the agent.
A detailed prompt, pointing to specific files and exact results needed, is all it takes.
The agent is also smart enough with memory files to know what files might be needed - so not really a win to be more precise than needed really.
The argument to compact often and work a lower context is not valid - some tasks are simply multi-file and thousands of lines of code naturally.
Claude enabled such tasks easily and 500K is reached often - while 1M is reached here and there.
I am talking about private repos of C# coding → “HOBBY STUFF” : 256K is de facto “defective”
The argument stands: Why can Claude provide 1M and OpenAI not? With Claude there is no care in the world and it simply works while GPT SOL is broken simply put.
I would also add it’s important to look at a problem piece by piece. It’s not “enter the prompt and walk away” yet.
You are right and still missing the point - Claude can provide that and simply put → it is like 80% less time spent dealing with the context window. You are clearly not working with proper code bases if you argue for the “limited” context window and not the “full” one. I don’t get it - in what dimension do you think, where you simply accept “the bare minimum” is enough. I don’t understand. And that is for the paid version. Paying money for the bare minimum while Claude provides you with 1 million context for the same price pretty much? Don’t you understand the situation I am referring to?
CODEx is same price as CLAUDE for 25% the context window? It falls in the category of "what the f***? "
Please stop joining the conversation to defend the limited context window - the post is about the limitation while the competition is providing 4x the “performance” pretty much. This is factual true and cannot be argued against.
my comment was merely aimed at “workaround”. Your request stands and is reasonable - who doesn’t want longer context? I still use 4.1 for a lot of work for that reason.
I’d be very surprised if OpenAI wasn’t pouring research and investment into longer context.
However, there’s a lot you can do with search and limited context.
Most of my codex work relates to an enterprise level codebase and codex seems to traverse it pretty deftly.
“workaround” is not true. It is factually broken in some use cases. Not usable without any chance of fixing the task. GPT fails since the 256K context window is crippling the agent into a disaster. It is like Windows with 8GB of SYSTEM RAM these days → Open a Browser and Discord + Launch a game → Sorry, not enough system resources. CODEX is at this stage.
ANALOG to your solution: Close all executable’s running and only run the game at the lowest possible settings.
De facto not acceptable.
I do not find it a dealbreaker and I can get on with my work, but sure, longer context would be great!
great ^^ - we are half way there: “would be great!” is the real thing and matching your competitor at the same price point is what is needed. The current state is (when you compare your choices) → you buy the Pro Plan of Claude since OPUS does a similar JOB with 1M context and GPT SOL (might be better in some tasks sure) is simply crippled.
This has just been shared tonight by Tibo on Twitter
Here is how to enable a 1M-token context window in Codex for GPT-5.6 Sol.
Codex is tuned to use a context limit that balances performance and cost, but a larger context window is a common request, so here is how to configure it.
A larger context window lets Codex retain more code, tool output, and conversation history before summarizing older material. You need a model that supports it. GPT-5.6 Sol, for example, has a documented 1,050,000-token context window.
Open ~/.codex/config.toml and add or update these settings at the top level, before any [section] headers:
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
The first setting selects the model. The second tells Codex to use a one-million-token context budget. The third starts automatic history compaction around 900,000 tokens, leaving some headroom.
Restart the Codex client and start a new session after saving.
To try the configuration for a single CLI session without changing your defaults:
codex -m gpt-5.6-sol
-c model_context_window=1000000
-c model_auto_compact_token_limit=900000
Have fun, but keep in mind that the default context limit has been tuned carefully for performance and cost.
Thank for the information - but this is a bit of the “hacked” method which could be “fixed” in the future or similar - or might get you “banned” since you kinda adjust the true parameters of the VScode extension ( i am just guessing here ) . Furthermore → a simple default for everyone is not hurting for less advanced users who put good money into their subscription <3.
I’m glad this helps.
Just know that no one is getting banned for making changes to config.toml.
It is a more advanced configuration change and can lead to degraded performance. I have also personally messed up the config in the past and could only fix it with a reinstall.
From that perspective, I wish there were a better way to configure Codex, too.
^^ Do I come around as toxic or so 
?
I just read my stuff from before and I am a bit afraid of the situation we are in.
To me it is like buying a very similar car → SOL and OPUS are quiet competitive (on coding tasks)
And the real world example would be to buy a car; same model; same everything; same price; but one simply is throttled to 100 km/h - while the other can (if the highway is free) drive 230 km/h at a higher cost in gasoline at that specific moment (which would translate to our token cost at higher context window)
Maybe Tibo read this Topic 
I can deny that what you wrote is simply not working in VSCODE using the official CODEX extension. I checked all configuration files and can also confirm that what you write is simply not true. So fairness wise I am asking for a video proof showing me what I did wrong. The video production is easy in takes less than 60 seconds if it works like you described.
I am curious what will happen now from your end. But I already know - guessing how the Internet works.