Codex pricing — what a task costs per token, and the break-even against a plan

Codex is sold on two different axes, which is why “how much does Codex cost” has no single answer. A ChatGPT plan bundles Codex into a flat monthly fee with usage limits. An API key bills per token, has no monthly fee and no limits, and costs nothing in a month you do not code. Which is cheaper is arithmetic, and the arithmetic is short.

This guide does not reprint OpenAI's plan tiers — those move, and a page that lists them is wrong the week after they change. It gives you the per-token side, one worked task, and a break-even you fill in with whatever your plan costs on the day you read this.

What a Codex task costs per token

Codex is agentic: one thing you ask for is many billed round trips. The unit that matters is therefore the step — one model call that reads context, proposes something and reports back — not the request and not the hour.

A step here is 25,000 input and 1,200 output tokens — the same sizing this site uses on the Claude Code page, so the two are comparable. Rates come from the catalog at render time, not from a hand-typed table, and each is printed next to OpenAI's own list price so you can check rather than trust.

The break-even

This is the whole decision, and it takes one division. A plan is worth its price only if it saves you more than its price in per-token spend:

# Break-even, in tasks per month.
#
#   plan_price_per_month / cost_per_task = tasks the plan must save you
#
# With a 20-step task on GPT-5.3 Codex at Kunavo rates:
#
#   cost_per_task = $0.484
#
#   $20 plan  ->  $20  / $0.484  = 41 tasks/month
#   $100 plan ->  $100 / $0.484  = 206 tasks/month
#
# Fewer than that and per-token is cheaper. More and the plan is.
# Substitute whatever your plan actually costs today.

Two things skew it in practice, both in the same direction. Plans have usage windows, so the heavy days you bought the plan for are the days you are most likely to hit a limit — and a limit you hit is capacity you paid for and did not get. Per-token has no window; the bill just tracks the work. Against that, a plan's cost is predictable and a per-token bill is not, which is worth something on a team budget.

Running Codex CLI on a per-token key

Codex CLI reads a custom provider from its config file. One constraint rules out most gateways: wire_api accepts only "responses", so the endpoint has to serve POST /v1/responses — not just chat completions.

# ~/.codex/config.toml — Codex CLI on a per-token key instead of a plan.
model = "gpt-5-3-codex"
model_provider = "kunavo"

[model_providers.kunavo]
name = "Kunavo"
base_url = "https://api.kunavo.com/v1"
env_key = "KUNAVO_API_KEY"
wire_api = "responses"

The key never goes in the file: env_key names an environment variable Codex reads at launch, so the config stays safe to commit. Full walkthrough, including what each field does and how to tell a protocol failure from an auth one, is on the Codex CLI API key page. If you drive Codex through a provider switcher instead, the CC Switch setup covers both directions.

Three things that move the real number

Reasoning tokens. Codex-class models produce internal tokens billed at the output rate that never appear in the diff you see. An estimate built from visible output length understates a reasoning-heavy task, sometimes by a lot. Read billed output from your usage data, not from the answer.

Prompt caching. Agentic tools re-send the same context every step, so most of that 25,000-token input is unchanged between turns. Cached input on the OpenAI protocol is billed at a fifth of the input rate, which on a long session is the single largest lever available — larger than switching models. The OpenAI pricing calculator has a field for it.

Retries. A request that consumed tokens and then failed downstream is still billed. On a flaky network or a rate-limited upstream this is a real percentage, and it is invisible in any per-request estimate.

Codex or Claude Code

They are protocol-bound in opposite directions — Codex speaks the Responses API, Claude Code speaks Anthropic Messages — but both are agentic CLIs billed by the step, so the arithmetic above transfers unchanged. The per-step comparison across both families is on Claude Code vs Codex, and the Claude side of the break-even is on Claude Pro and Max limits. One key on Kunavo reaches both, so the choice stays a --model argument rather than a second account. If the question is which OpenAI surface to open rather than which CLI to install, Codex vs ChatGPT covers what a shared plan allowance does when one of the two doors is agentic.

FAQ