GPT-5.1-Codex-Max: A Lower List Price That Costs 3x More (We Measured All Four)

GPT-5.1-Codex-Max lists at $1.25 per million input tokens and $10 per million output — cheaper on both sides than GPT-5.2-Codex at $1.75 and $14. Put those two rows in a pricing table and Codex-Max looks like the bargain. We ran both, plus GPT-5.1-Codex and GPT-5.3-Codex, on the same nine executed Python tasks. All four scored 9 out of 9. Codex-Max cost $7.11 per 1,000 tasks. GPT-5.2-Codex cost $2.30. That is 3.09x more money for an identical score, from the model with the lower sticker price, and it took roughly twice as long to get there. The mechanism is not a hidden fee. It is token volume, and it is the reason list-price comparisons quietly mislead.

Nobody publishes the number that decides your Codex bill. Providers publish price per million tokens. Your invoice is price multiplied by how many tokens the model chooses to emit, and that second factor varies more between Codex tiers than the price does.

All four, measured

Same nine tasks, same prompts, temperature 0, one scored attempt each. Cost is derived from the measured token counts at the list price captured 2026-08-22.

Read the first and fourth rows together. The model with the higher list price on both sides produced the smallest bill, by a factor of three.

The inversion: cheaper price, bigger bill

Why Codex-Max spends more

It writes more. On the identical nine prompts, Codex-Max emitted 6,327 output tokens against GPT-5.2-Codex's 1,404 — 4.5x the volume. Of that, 569 tokens were reasoning against 50, an 11.4x gap.

Reasoning tokens bill at the output rate. You do not see them, you cannot cache them, and on these tasks they bought nothing: both models solved all nine. The extra thinking also shows up as wall-clock — 9.3 seconds median against 4.7.

This is the same variable we found dominating agent costs when we looked at model choice for a multi-turn agent harness: reasoning spend, multiplied by every turn in a session, is the whole invoice. In a single call it is an annoyance. In a Codex loop that fires dozens of calls per task, it compounds.

The cleanest comparison: same list price, same score

Strip out the price variable entirely. GPT-5.1-Codex and GPT-5.1-Codex-Max carry the identical list price: $1.25 in, $10 out. Both scored 9 out of 9.

Same price, same result, 2.8x the bill and roughly twice the wait. Whatever the Max tier is for, our nine tasks did not surface it — which is a limit of the tasks as much as a verdict on the model. See the limits section.

Which Codex to actually use

• Default: GPT-5.2-Codex. Cheapest of the four on measured cost, within 0.2 seconds of the fastest, 9/9.
• If you want the newest: GPT-5.3-Codex. Fastest of the four at 4.5 seconds, and 35% more expensive than GPT-5.2-Codex for the same score on these tasks.
• GPT-5.1-Codex is the value floor if you are already pinned to the 5.1 generation — $2.54, 4.9 seconds.
• GPT-5.1-Codex-Max: we cannot justify it from this data. If your workload is genuinely harder than nine self-contained Python functions, measure it on your tasks before paying 3x.

What this costs against the wider field

Codex is not competing on price with the open field. Across the 51 models in our benchmark set, DeepSeek V3.2 scored 9/9 at $0.08 per 1,000 tasks and GPT-5.4 mini scored 9/9 at $0.53 in 2.3 seconds. GPT-5.2-Codex, the cheapest Codex here at $2.30, ranks 21st cheapest among our 9/9 models.

That is not an argument against Codex — it is a tuned coding model with tooling around it that a raw chat endpoint does not have, and our nine tasks measure the model, not the harness. It is an argument for knowing the multiple you are paying. Our cheap coding model roundup covers the low end, and GPT-5 Codex vs GPT-5 covers when the Codex tuning earns its keep at all.

How these numbers were produced

Nine Python tasks, each a function signature plus a spec and no example test cases. The model's generated code is executed against hidden asserts in an isolated python3 -I subprocess with a 12-second timeout. Temperature 0, max_tokens 4000, one scored attempt per task. Cost is derived: measured input and output token counts multiplied by the list price on the date shown — not a billing statement. Prices move, so every figure here carries the 2026-08-22 capture date, and our note on price volatility explains why that matters. Full write-up on the methodology page. Runs go through OpenRouter.

What we did not measure

• Long-horizon agentic work. Nine self-contained functions is not a repository refactor. If Codex-Max is built for extended autonomous runs, this suite would not show it.
• The Codex CLI or IDE integration. We called the models through an API. Tooling is not in scope here.
• Repeated runs. One scored attempt per task per model. These are single-run figures, not averages, and we do not average across runs.
• Anything beyond Python. The suite is Python only.

FAQ

Is GPT-5.1-Codex-Max worth the money? Not on our nine tasks. It scored the same 9 out of 9 as three cheaper Codex models and cost $7.11 per 1,000 tasks against $2.30 for GPT-5.2-Codex.

Why does Codex-Max cost more when its list price is lower? It emits 4.5x the output tokens on identical prompts. Price per token is only half of a bill; volume is the other half.

Which Codex model is cheapest to run? Of the four we measured, GPT-5.2-Codex at $2.30 per 1,000 tasks.

Which Codex model is fastest? Of the four, GPT-5.3-Codex at 4.5 seconds median, with GPT-5.2-Codex 0.2 seconds behind.

Did any Codex model fail a task? No. All four scored 9 out of 9.