GPT-5.6 Sol vs GPT-5.3 Codex

Context window	1,050,000 tokens	400,000 tokens
Knowledge cutoff	Feb 16, 2026	Aug 31, 2025
Standard input price	$5.00/MTok	$1.75/MTok
Standard output price	$30.00/MTok	$14.00/MTok
Reasoning effort levels	Not published	low, medium, high, xhigh
Long-context repricing rule	Published	Not published
Tier 1 rate limit	500 RPM · 500,000 TPM	500 RPM · 500,000 TPM

No OpenAI page states this comparison directly, which is exactly why it's worth stating here: the Codex CLI's actual default model and the model literally named for Codex are two different things, and the choice between them is a real, three-axis trade-off rather than an obvious pick.

The three axes that actually differ

Context window is the first — one of these models offers meaningfully more headroom for a large codebase or a long agentic session than the other. Knowledge cutoff is the second — one model's training data is noticeably more recent. Price is the third, and it moves in the opposite direction from context window: the model with the larger window and more recent knowledge also costs more per token, on both input and output.

What doesn't differ

Both models sit in the same published rate-limit group, tier for tier — so choosing the Codex-branded model over the CLI default costs nothing in throughput. That rules out one axis you might otherwise assume matters: this is not a "the cheaper model is also the slower one" trade-off. It's specifically cost against context headroom and recency, with throughput held equal.

One more asymmetry worth knowing

The long-context repricing rule that applies to crossing the larger model's input threshold is published for that model and explicitly not published for the Codex-branded one — recorded honestly as "not published" here rather than assumed to carry over. If your workload runs consistently large inputs, that's a real, separate reason the comparison isn't purely about the three axes above.

How to actually decide

If your work leans on a large amount of accumulated context — a big repository, a long multi-turn agentic session — the larger window is often worth its price premium on its own. If your sessions are typically smaller and reasoning-effort control matters more to you than raw window size, the Codex-branded model's published effort levels and lower price are the stronger pull. Neither is a universally correct default; that's precisely why the CLI's own choice of default doesn't settle the question for every workload.