Codex Pro GPT-5.4 long context regressed: ~672k worked Apr 22, ~290k fails Apr 23

Codex desktop app: 26.417.41555
Bundled app-server: codex-cli 0.122.0-alpha.13
Homebrew CLI after upgrade: codex-cli 0.123.0
Auth mode: ChatGPT
Plan shown in logs: pro
Rate limit shown in logs: limit_reached=false

Config:
model = "gpt-5.4"
model_context_window = 1000000
model_auto_compact_token_limit = 700000

Observed:

Apr 22: GPT-5.4 turns succeeded around 635k-672k total tokens with auto_compact_limit=700000.
Apr 23: GPT-5.4 turns started failing around 270k-309k tokens with:
"Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying."

Expected:
Codex Pro + GPT-5.4 should continue using the long-context route as it did Apr 22, or auto-compact before rejection.

Actual:
Client keeps auto_compact_limit=700000, but backend/app-server rejects around the ~272k boundary.

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

Codex CLI freezes near auto-compaction threshold and drifts with very large context windows #19116
Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying. #18052
Context window always 90%+ full #18136
Repeated automatic context compaction after every prompt/task. #18827

Powered by Codex Action

@markobogoevski-dev

@markobogoevski-dev you have paid plan or have pending due payment?

@markobogoevski-dev you have paid plan or have pending due payment?

I have a paid plan, email is markobogoevski@gmail.com.

CURRENT SUBSCRIPTION
ChatGPT Pro Subscription
$200.00 per month
Hide details
ChatGPT Pro Subscription
$200.00
Total
$200.00
Your next billing date is May 2, 2026.

Date paid is April 2nd.

Same issue here.

Issue in both Codex CLI and Codex Windows App (project inside WSL)
GPT5.4 xhigh was at 221k / 950k, then I started plan mode. And it crashed due to

Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.

Between it being at 221k tokens and the crash, it read 7 files and did 3 rg searches. Another codex instance investigated these actions from the Codex Sessions jsonl and estimated a token usage between 35k and 50k token.
I had this issue multiple times today and every time it happened in plan mode or right after confirming a suggested plan.

config.toml

model = "gpt-5.4"
model_context_window = 1050000
model_auto_compact_token_limit = 800000

me too, I also meet this question, since updated the newest Codex app

@etraut-openai

@etraut-openai

Same ! I also ran into the same issue as you when using codex-cli version 0.123

I may know why now. It's probably because yesterday the request rate for GPT 5.4 was routed to GPT 5.5 again. Since the maximum context window size for GPT 5.5 is only 256k, it truncates requests at this point each time.

Navigation Menu

Repository navigation

Codex Pro GPT-5.4 long context regressed: ~672k worked Apr 22, ~290k fails Apr 23 #19142

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Codex Pro GPT-5.4 long context regressed: ~672k worked Apr 22, ~290k fails Apr 23 #19142

Description

Activity

github-actions commented on Apr 23, 2026

hookenful commented on Apr 23, 2026

markobogoevski-dev commented on Apr 23, 2026

JanDuchscherer104 commented on Apr 23, 2026

MOON-DREAM-STARS commented on Apr 23, 2026

hookenful commented on Apr 23, 2026

daothinh commented on Apr 23, 2026

MOON-DREAM-STARS commented on Apr 24, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions