Default GPT-5.6 context can cross the 272K higher-usage threshold
What version of Codex are you using?
Codex CLI 0.144.0
What model are you using?
GPT-5.6 Sol and GPT-5.6 Luna
What issue are you seeing?
GPT-5.6's default Codex context configuration appears to allow sessions to enter the higher-usage pricing band without the user explicitly opting into a larger context.
The current model metadata reports a 372,000-token context window with a 95% effective-context multiplier, yielding an effective window of approximately 353,400 tokens. OpenAI's GPT-5.6 model pricing documentation states that prompts with more than 272K input tokens are priced at 2x input and 1.5x output for the full request.
That means the default effective context can extend roughly 81K tokens into the higher-cost band:
- Raw context: 372,000
- Effective context: 353,400
- Higher-cost threshold: 272,000
- Exposure above threshold: approximately 81,400 tokens
This is particularly easy to encounter when Sol uses subagents to audit its own work, because child-agent context and reasoning can add usage without the user intentionally selecting a large-context mode.
Expected behavior
The default auto-compaction threshold should keep normal sessions at or below the 272K pricing boundary, or Codex should clearly warn that continuing past the boundary will increase usage.
A larger context should require an explicit opt-in, or the UI should clearly show the applicable usage multiplier before the request runs.
Possible solutions
- Set the default auto-compaction threshold below 272K for GPT-5.6 Codex sessions.
- Show a visible warning before a request crosses the 272K threshold.
- Make the larger-context behavior an explicit custom-config opt-in.
- Display the active context threshold and pricing multiplier in /status or the usage panel.
- Apply the higher multiplier only to tokens above 272K rather than the full request, if technically feasible.
Additional information
The documented pricing behavior is described here:
https://developers.openai.com/api/docs/models/gpt-5.6-luna
As a local workaround, users can set:
model_auto_compact_token_limit = 270000
However, requiring users to discover and configure this manually makes it easy to enter the higher-usage band unintentionally.
7 Comments
I dont think its that much of a great idea giving users more than 270k context window
its better the models think in short words, and be concise ----
what kind of things you are facing? why do you need higher context window? --- its better to keep your codebase short and lean --- if it get big - maybe your coding agents made too much code
please see
https://learn.chatgpt.com/use-cases/refactor-your-codebase
I also recomend sticking to Terra instead of Luna... I guess you can use SOL for one time, but only to simplify your code base....
part of the 5.6 release is, the Luna and Terra are so great... luna being the same class as very high reasoning in 5.5 .....
generally with gpt 5.5 and 5.6 it gets the right ideas quickly , this cant be said about other non-openai models which non has the same vibe gpt has.... ever since the 5.4 release....
for example 5.3 was a good release, but I was very disapointed at it, and 5.2 and older were very sloppy ,,, and it was weird why codex as so bad few months ago....
i understand if you have some bad memory from the first models.... but see how Sora2 ivdeo gen , is so good at generalizing , also generalizing audio --- and no one able to get near the same kind of concepting things
so its likely comes from a correct idea -- also in terms of training the models --- having deal with 1million tokens seems like a nightmare,, rather than 200k tokens -- which is less of an adhd ...
<img width="892" height="492" alt="Image" src="https://github.com/user-attachments/assets/7d08438a-076c-487b-8575-c78d03c244b3" />
<img width="1215" height="610" alt="Image" src="https://github.com/user-attachments/assets/fa321a27-b559-45cf-9dc3-7bfd6c4bc5fd" />
my argument is that Luna is so great, because the ExploitBench for example shows some a requirement for logic understanding - so it should preform in code stylies its not really familuar with
gpt 5.6luna is very good// even if you can argue this is due to some weird expirement that proven very efficent -- it doesnt come along without the correct ideas -- even historically at the gpt5.2 era which was really really bad -- but nice to use in terms of how quickly things have escalated - so its not just good data (or good data synthesis) those models reason for real, and really have logic using words -- its not real agi and general concepting because the models are so damn big now --- but its good that we have an alternative to the internet itself --- its very interesting how things get solved by the models mutating the datasets it was given and great logic out of it --- think about law and how much of it is citation -- and not truly judgement -- but you can also argue how much english evolved from the times of the follio by shakespear plays and how some random british man could rant in the 1720 and how much english has been ruined --- when in our timeframe we can consider english from the 1800 and 1920 as more truthful ,, but its okay because language is to pass logic -- and not emotion (rather than using vague soulless terms or slangs , [ e g , "dark fiber"])
for any task I ever consider using 5.5medium , and not gpt5.4mini or gpt5.5low ---
gpt5.6luna does anything I expect from a bigger model and more///
I assume your should use SOL for code review or whatever , maybe to tell you a different ideas --- or how other products do it better than you
I recomend using
https://learn.chatgpt.com/use-cases/refactor-your-codebase
and refactour your toy code base, for better token usage
dont forget, its not going to decide by itself to make your code smaller,
codex will only add code, and if remove, it will remove lines regarding the logic you want to change its not going to discover flaws which all software suffer from... bloated and good refurbished-or-not
)))
<img width="762" height="488" alt="Image" src="https://github.com/user-attachments/assets/855e536f-981f-4eb9-8e71-f843a8fef335" />
@JD3Lasers what codebase are you using, is it Typescript is it javascript? whats bloating your context window? I dont expect openai to target your specific niche, but I assume there is a way for some users --- maybe force their coding agents , use better REGEX or whateever
so if you are using nodejs , whats happening , where those tokens go for
you assume on the negative-hand , and counter-argument --- if you are going to force more regex and more excessive auto-cleaning-tokens-without-compacting/compacting ,, you assume degraded preformance for a lot of tasks ... as with people I guess you assume some people need to have degraded preformance in some things -- to be very good at something else ----- so for absulote choice you should use the idea that is more general and best preforming for most using --- for the first stages at the product perhaps
))
@Mahkhmood9 its not just one code base but mostly python. I have over 60 billion tokens used with codex so I'm pretty familiar with my usage and how to use it. My typical day would eat 10-15% of the pro plan weekly. Now I can eat 30% of it in the same day. Same work flows
@JD3Lasers I’m working in mostly python as well, changed nothing other than the model slugs to 5.6 sol/terra/luna and my usage goes in a blink. No idea wtf is going on
ITs good that you expirement with SOL, I recomend you try using the smaller models like Terra and Medium ...
try using /goal or /plan and a smaller model -- try defining your task
by design its meant to be just as powerful as older models like 5.5
Honestly what stops you from avoiding SOL completely, Terra is just as good - I recomend find the limits in what it cant do in the weeker models -- before using the best models