Recently, there was a change in the token context limit for the codex(chatgpt) app and the codex cli + any connectors to codex app-server. The change was from roughly 1M tokens down to 258K tokens. which is a substantial decrease in the token context limit.
So instead of the model keeping track of everything it is and has done, it auto compacts quite often and loses track of what i originally set it to do. this is absolutely unacceptable for the kinds of long horizon work that I like to do. My code quality has went down a good bit compared to using a competitors latest frontier model.
if it isnt resolved soon, im going to switch to claude. changing this out of the blue is stupid of openai!
they do these things on purpose, dont expect them to care
this is unfortunate. in my opinion, codex has been slowly been overtaken by enshittification.
VeitB
August 2, 2026, 8:41pm
5
Yes, the context size limit was reduced from 372K to 272K in mid-July. I suggest joining the corresponding issue reports in the public repo to get the original context window reinstated.
opened 09:39PM - 21 Jul 26 UTC
enhancement
CLI
context
app
config
# Restore GPT-5.6 Sol’s 372k Codex context window, or provide an opt-in setting
…
### What variant of Codex are you using?
Codex desktop app and/or Codex CLI with a ChatGPT subscription.
* Subscription: `Pro 20x`
* Codex version: `26.715.70719 `
* Platform: `Mac app Apple Silicon`
* Model: `gpt-5.6-sol`
### What feature would you like to see?
Please restore GPT-5.6 Sol’s original **372,000-token Codex context profile**, which provided approximately **353,400 effective tokens** under the current 95% effective-window policy.
If restoring 372k as the default is not currently possible, please provide a supported per-model or per-thread option that lets users choose the 372k profile, with a clear warning about any additional usage impact.
## Summary
GPT-5.6 Sol launched in Codex with:
```text
Raw context window: 372,000
Effective context window: 353,400
```
The current server-delivered profile provides:
```text
Raw context window: 272,000
Effective context window: 258,400
```
That is a reduction of **95,000 effective tokens**, or approximately **26.9%**.
This is a material product regression for long-running, detail-sensitive workflows. The earlier 372k profile noticeably improved continuity and reduced the loss of important information caused by compaction.
## OpenAI’s public statements
On July 12, Tibo stated that GPT-5.6 Sol’s context threshold had been:
> “tuned … to be perfect at the default limit”
Source: https://x.com/thsottiaux/status/2076201049086648705
At that time, the launch default was 372k.
On July 13, Tibo explained that the product had been temporarily returned to 272k because the move to 372k caused more subscription usage to be charged than intended. The same post said OpenAI:
> “will work to roll back out to 372k in the days to come”
Source: https://x.com/thsottiaux/status/2076495156757577895
This indicated that 272k was a temporary usage-accounting measure—not the intended permanent GPT-5.6 Sol configuration.
Please either follow through on restoring 372k or provide a clear update explaining:
* Whether restoration is still planned
* Which plans and Codex surfaces will receive it
* Whether there is a technical or usage-accounting blocker
* When users should expect the original profile to return
## Real-world impact
For many workflows, 272k is not working nearly as well as 372k.
My use cases include researching and reasoning over substantial amounts of internal business information, multiple documents, detailed requirements, prior decisions, exceptions, and relationships between facts.
At 272k:
* Automatic compaction occurs significantly earlier.
* Important nuances and earlier constraints are more likely to disappear after compaction.
* The model misses relationships between details that were previously available in its active working context.
* Users must repeatedly re-anchor the model by restating requirements and reloading information.
* Long non-coding research and analysis tasks become less reliable.
* Compaction interrupts otherwise coherent, long-running work.
The 372k range provided a reasonable and noticeable improvement. It allowed GPT-5.6 Sol to maintain a larger coherent working set before lossy summarisation became necessary.
This is not merely a request for the largest possible context window. It is a request to restore the launch configuration that OpenAI publicly described as tuned for the model and said would return.
## Requested resolution
Please implement one of the following:
### Preferred
Restore the original GPT-5.6 Sol profile for paid Codex users:
```text
context_window: 372000
max_context_window: 372000
effective_context_window_percent: 95
```
### Acceptable alternative
Expose a supported setting in the app and CLI:
```toml
gpt_5_6_sol_context_profile = "372k"
```
The setting could include a clear notice that larger contexts may consume subscription usage more quickly.
### Additionally requested
* Publish context-window changes in the Codex changelog rather than only through X posts.
* Clearly distinguish raw context, effective context, and auto-compaction thresholds in the UI.
* Do not silently reduce the context profile of existing threads.
* Allow users to select between a usage-efficient profile and a long-context profile.
* Provide an official status update on the promised restoration of 372k.
## Scope
This issue is specifically requesting restoration of the **372k Codex launch profile**.
It is not demanding that subscription users immediately receive the API model’s full 1.05M window. That is a separate discussion. Restoring the already deployed and publicly promised 372k profile would be a meaningful immediate improvement.
## Related issues
* #32803 — Effective context window shrank from approximately 353k to 258k
* #32806 — Regression report covering the 372k-to-272k reduction
* #31860 — GPT-5.6 Sol Codex context profile versus the published API capacity
Please restore 372k—or at minimum give users the option to select it.