GPT-5.6 Sol is 272K inside Codex
same account, drop originator, 872K
From 2026-08-22, on one ChatGPT Pro account, Codex coding clients clamp gpt-5.6-sol max_context_window to 272000 (about 258400 effective). Omit the coding originator and the same token returns 872000. The model spec page still says 1,050,000. This is a catalog split, not a takedown.
Four numbers you can quote alone
Coding-client catalog
With originator codex_cli_rs, codex-tui, Codex Desktop, or codex_vscode, the same account receives max_context_window = 272000.
Same account, non-coding originator
In the issue's control, the same token without a coding originator returned 872000. Terra and Luna stay 872000 on both catalogs.
Effective UI window
272000 × 95% = 258400. openai/codex#40258: 353400 on 2026-08-21, 258400 on 2026-08-22.
Still on the model spec page
Public GPT-5.6 Sol docs still advertise a 1.05M model context. That is the model spec, not the Codex coding-surface catalog cap.
This is not context_length_exceeded — the catalog cuts the window first
openai/codex#40258 (opened 2026-08-23, fetched 2026-08-30) ran a controlled test on one ChatGPT Pro account, one URL, client_version=0.149.0: changing only the HTTP originator header flips gpt-5.6-sol max_context_window between 272000 and 872000. Coding surfaces (CLI / TUI / Desktop / VS Code) get the 272K catalog (one ETag); non-coding originators get 872K. Codex clamps a local model_context_window override to the remote max_context_window, then applies a 95% effective window, so the UI shows 258400. The raw body has no cohort / experiment / downgrade-reason field. Related #32806 recorded the earlier 353400 → 258400 drop. The official model page still says 1,050,000. This page is about the Codex product catalog. It does not claim the inference endpoint rejects prompts above 272K, and it does not claim any reseller window equals 1.05M or 872K.
Why the window dropped on 2026-08-22
The report’s user-visible timeline: same CLI 0.149.0, same model_context_window = 372000, effective window 353400 at 2026-08-21T08:58:30Z and 258400 at 2026-08-22T13:16:48Z with no local version change. Upstream PR #39102 had raised the bundled Sol/Terra/Luna override to 872000, but coding originators are still clamped to 272K by the server catalog. Sol also spent July–August burning Codex limits faster than expected (5-hour window, weekly cap). A smaller catalog window means earlier compaction and a “limits again” feeling. Read the catalog number before you chase usage-limit docs.
Timeline
Community issues recorded Codex delivering Sol at ~353400 effective (catalog 372000 × 95%) against a 1.05M spec. Terra / Luna catalogs stayed 872000.
Same account, no CLI bump: effective window fell from 353400 to 258400 (catalog 272000 × 95%).
openai/codex#40258 published the originator split. The issue was still open when fetched on 2026-08-30.
Confirmed vs not proven
Confirmed (community controlled test, not an OpenAI post)
On the public GitHub issue: one Pro account, originator only, Sol max_context_window flips 272000 ↔ 872000. The four coding clients share the 272K ETag. Terra / Luna do not flip. 258400 = 272000 × 0.95.
Not proven / do not extrapolate
OpenAI has not documented “coding originator → 272K” as a product rule. Do not call 272K the model’s hard ceiling, and do not promise 872K or 1.05M on QCode or any proxy. The experiment shows Codex returning different catalog metadata per originator.
Coding client vs model spec — not two models
Codex coding surfaces: the catalog wins
CLI / TUI / Desktop / VS Code send a coding originator. Remote 272000 overrides a larger local model_context_window. Threads compact against 258400 from the first turn. A local config knob does not beat the remote clamp.
Model spec / non-coding originator: larger numbers
The developer spec still says 1.05M. The same account without a coding originator gets 872000 for Sol in the Codex catalog. Terra / Luna are 872000 on both catalogs. That is a surface split, not “Sol was removed”.
Confirm it is the catalog before you chase context_length_exceeded
Step 1: read model_context_window in the UI or task_started. ~258400 matches 272000 × 95% — you do not need a “shadow ban” theory. Step 2: do not start with /usage and the 5-hour window; that is a different limiter, already documented on this site. Step 3: if you need the spec-page long window, leave the coding-originator catalog and use an API client that does not consume that Codex coding catalog. This page does not describe spoofing originator, bypassing ChatGPT auth, or sharing accounts.
An API catalog is not the Codex coding catalog
gpt-5.6-sol logged 106,837 calls on QCode in 30 days (CRS 2026-08-30; positive control claude-sonnet-5 was 188,213). That means the model id is routed, not that QCode’s context window is 1.05M, 872K, or 272K. This page does not promise a window size. To avoid the Codex coding-client clamp you need an OpenAI-compatible API, not a different originator inside Codex CLI.
FAQ
How is 258400 computed?
Remote max_context_window = 272000, times effective_context_window_percent = 95, equals 258400. The previous 372000 × 95% = 353400. It is not UI rounding.
Did Pro accounts get singled out?
In #40258, plan_type stayed pro and both catalogs listed pro in available_in_plans. The split moved with originator, not with “being Pro”. That is a controlled result in the report, not a promise for every plan.
Were Terra and Luna also cut to 272K?
In the same dump, Terra and Luna stay at max_context_window 872000 on both catalogs. The public originator split is documented only for gpt-5.6-sol.
Will setting model_context_window to 372000 or higher unlock it?
No. The report says the remote 272000 clamps an explicit override. A larger window needs a different surface, not a bigger local integer.
Is this the same as context_length_exceeded?
No. context_length_exceeded fires after a request already exceeded the live window. Here the client cuts the window metadata to 272K before the request. Read the catalog first, then the 400.
Does the official 1.05M spec still count?
It counts as “the number on the model spec page”, not as “Codex coding clients guarantee 1.05M”. Both numbers exist. Trust the catalog or API you actually hit.
Sources
The originator split, 272000 / 872000, 258400, and the 2026-08-22 regression timeline are from GitHub openai/codex#40258 (opened 2026-08-23, fetched 2026-08-30) plus related #32806 / #31860. 1,050,000 is the public GPT-5.6 Sol spec cited in the issue (developers.openai.com/api/docs/models/gpt-5.6-sol). QCode call counts are CRS 30-day usage, fetched 2026-08-30. OpenAI’s changelog does not document this split as a product term.
If you wanted the API spec window, stop staring at the Codex catalog
272K is the coding-client catalog, not “Sol disappeared”. For an OpenAI-compatible endpoint, prepaid credit at official rate × a service multiplier is one path.
Related
Sol vs Terra vs Luna
How the three GPT-5.6 tiers differ; do not mix window numbers with tier quality.
Codex 5-hour limit
A smaller window is easy to misread as a quota problem. Separate catalog clamp from the 5-hour window.
context_length_exceeded
The 400 / silent truncation after you actually overflow — different from “the catalog shrank first”.
Based on public GitHub issues and the OpenAI model spec page. OpenAI has not confirmed the originator split as a lasting product rule. QCode does not promise Codex’s catalog window or the 1.05M spec, and does not document spoofed headers or ChatGPT-auth bypasses.