Codex and ChatGPT are two doors onto the same models, and on a plan they spend the same allowance. The comparisons that rank for this question explain the interfaces well enough. What none of them says is what the choice does to your bill — which is the part that decides it once you know what each one is.
Two surfaces, two loops
| ChatGPT | Codex | |
|---|---|---|
| What you hand it | A question | A task and a working directory |
| What it touches | Nothing — you copy results out | Your files, and your shell |
| The loop | Ask, read, ask again | Plan, edit, run, check, repeat |
| Cost per unit of work | One exchange | Many billed steps per task |
| Off-plan option | API, per token | Codex CLI on a key, per token |
The fourth row is the one that matters and the one nobody states. Because Codex is agentic, a single thing you ask for becomes many model calls — so on a shared plan allowance, a Codex session drains capacity far faster than the same amount of wall-clock time spent chatting.
What that costs off the plan
Once Codex runs on a key the arithmetic is concrete. A step is one model round trip — read, propose, report — sized here at 25,000 input and 1,200 output tokens, the same sizing used across the other cost pages on this site:
| Model | Kunavo, per 1M in / out | OpenAI list | Per step | 20-step task |
|---|---|---|---|---|
| GPT-5.3 Codex | $0.70 / $5.60 | $1.75 / $14.00 | $0.024 | $0.484 |
| GPT-5.4 | $1.00 / $6.00 | $2.50 / $15.00 | $0.032 | $0.644 |
Divide whatever your plan costs by the per-task figure and you have the break-even in tasks per month. The full version of that arithmetic, including what the free tier does and does not cover, is on Codex pricing. This page deliberately does not reprint OpenAI's plan tiers — they move, and a page quoting a stale one is worse than a page quoting none.
Running Codex with no plan in the path
# The third surface the comparisons leave out: Codex CLI on a key,
# billed per token, with no plan quota in the path at all.
#
# wire_api accepts only "responses" for a custom provider — which is why
# most OpenAI-compatible gateways cannot drive Codex.
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 wire_api line is the whole constraint: Codex accepts a custom provider only over the Responses API, so a gateway that stops at chat completions cannot drive it at all. Field-by-field detail is on the Codex CLI API key page, and if you switch a coding client between providers rather than committing to one, CC Switch covers both directions.
The split most people end up with
Plan in the conversation, execute in the agent. Exploratory work — reading an unfamiliar error, arguing about an approach, deciding what to build — is faster and cheaper in chat, and an agent that starts editing during that phase produces work you then have to undo. Once the plan is settled, handing it to Codex is where the agent loop pays for itself.
If you are choosing between terminal agents rather than between surfaces, Claude Code vs Codex compares the two protocol-bound CLIs, and one key on Kunavo reaches both.
FAQ
What is the difference between Codex and ChatGPT?
They are different surfaces onto the same models, aimed at different loops. ChatGPT is a conversation: you ask, it answers, you read. Codex is an agent: it reads your repository, edits files, runs commands and iterates until a task is done or it gives up. The practical tell is what you hand it — ChatGPT takes a question, Codex takes a task and a working directory. Using ChatGPT for repository work means copying context in and results out by hand, which is the friction Codex removes.
Do Codex and ChatGPT share the same subscription?
On a ChatGPT plan they draw on the same account, which is why the two feel like one product with two doors. That also means they compete for the same usage allowance: a heavy Codex session consumes capacity that would otherwise have been available for chat. The alternative is to take Codex off the plan entirely by running Codex CLI against an API key, where usage is billed per token and there is no shared allowance to exhaust.
Is Codex or ChatGPT cheaper?
Inside a plan neither is cheaper — they spend the same allowance, and the question is only which one uses it faster. Codex uses it much faster, because an agent step resends context and one task is many steps. Off the plan the comparison becomes concrete: at Kunavo rates a step of 25,000 input and 1,200 output tokens on GPT-5.3 Codex is $0.024, so a twenty-step task is about $0.484. Divide your plan price by that per-task figure and you get the number of tasks a month the plan has to save you to be worth keeping.
When should I use ChatGPT instead of Codex?
When the work is thinking rather than editing. Design discussions, reading an unfamiliar error, planning an approach, or anything where you want to argue with the answer before code is touched — those are cheaper and faster in a conversation, because an agent that starts editing during the exploratory phase produces work you then have to undo. A common working split is to plan in chat and hand the settled plan to the agent.
Can I run Codex without a ChatGPT subscription?
Yes — Codex CLI accepts a custom provider through ~/.codex/config.toml, with one constraint that rules out most gateways: wire_api accepts only "responses", so the endpoint has to serve POST /v1/responses rather than just chat completions. With that in place there is no plan in the path, no usage window, and a month where you do not code costs nothing.
Does Codex use a different model from ChatGPT?
Codex runs on models tuned for agentic coding rather than on whatever the chat surface defaults to. On Kunavo that is GPT-5.3 Codex at $0.70 / $5.60 per 1M tokens, against GPT-5.4 at $1.00 / $6.00 for general work — so for mechanical tasks the general model is the cheaper pick, and the tuned one earns its price where the task is genuinely agentic.