Unlocking 1.05M Tokens in Codex with Three Config Lines—Why It’s a Costly Trap

OpenAI engineer Tibo announced that the GPT‑5.6 Sol model supports a 1.05 million‑token context window, but Codex ships with a smaller default; enabling the full window only requires three lines in ~/.codex/config.toml:

model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000

After inserting these lines and restarting Codex, the author ran a large project and observed that token consumption jumped 1.7 times compared with the default setting.

To quantify the effect, four test projects were created:

Small (5 k tokens): consumption 5 k, ratio 1.0×.

Medium (15 k tokens): consumption 15 k, ratio 1.0×.

Large (30 k tokens): consumption 34.2 k, ratio 1.1×.

Extra‑large (50 k tokens): consumption 74.2 k, ratio 1.5×.

Extreme (90 k tokens): consumption 154.2 k, ratio 1.7×.

The jump begins once the default context window of 258 k tokens is exceeded; OpenAI designed the system so that token‑consumption speed doubles beyond that point, as confirmed by Tibo’s warning.

OpenAI’s MRCR v2 benchmark shows accuracy of 91.5 % for the 256 k–512 k range, dropping to 73.8 % for the 512 k–1 M range—a loss of 17.7 percentage points, roughly 3.6 % per additional 100 k tokens.

GPT‑5.6 Sol’s MRCR v2 evaluation: 256K‑512K accuracy 91.5 %; 512K‑1M accuracy 73.8 %.

When asked what this decay means, an AI assistant replied that at 1 M context the model may forget details after the 800 k token mark, risking critical omissions.

Tibo also noted that Codex’s default 258 k window is carefully tuned for performance and cost, and that the automatic compact feature can summarise older tokens to keep consumption normal.

Two usage strategies were compared:

Enable 1 M context: all old code retained, but token usage roughly doubles and accuracy falls to 73.8 %.

Keep default 258 k with automatic compact: older code is summarised, token usage stays normal, though some detail may be lost.

Guidelines for enabling the 1 M window:

Use it for one‑off large‑code‑base analyses that require full‑text search across many files.

Avoid it for routine coding, bug fixing, or testing where each interaction only consumes a few thousand tokens.

For temporary needs, run Codex with the three‑line parameters in a single session, e.g.:

codex-mgpt-5.6-sol \
  -c model_context_window=1000000 \
  -c model_auto_compact_token_limit=900000

This avoids polluting the global configuration, though the same token‑doubling penalty still applies within that session.

In summary, while the three‑line tweak unlocks a larger context window, it incurs higher cost and a noticeable drop in recall; the default 258 k setting remains sufficient for most tasks, and the 1 M option should be reserved for specific, large‑scale analyses.