Codex without a ChatGPT subscription: pointing it at another API

Codex is OpenAI's console agent, and by default it calls their API. But the provider comes from a config file, which means any compatible address can take OpenAI's place. For anyone who cannot pay for the subscription, this is the only working route.

The config

Codex reads ~/.codex/config.toml:

model = "gpt-5.6-sol"
model_provider = "zukko"
model_reasoning_effort = "high"

[model_providers.zukko]
name = "zukko"
base_url = "https://api.zukko.pro/v1"
wire_api = "responses"
requires_openai_auth = true

The key lives separately, in ~/.codex/auth.json:

{ "OPENAI_API_KEY": "fetch_your_key" }

The non-obvious fields:

wire_api = "responses" — Codex speaks the Responses API rather than classic chat/completions. Your provider must support it; any other value sends requests to the wrong place.

base_url with the /v1 suffix — mandatory for OpenAI-compatible mode. Omitting it produces a 404 on every call.

model_reasoning_effort controls how much the model thinks before answering. Higher means better and more expensive. On simple tasks high earns little.

Automatic setup

If you would rather not edit config files by hand, there are ready scripts. For macOS, codex-setup-macos.command↗: download it and double-click, or from a terminal:

curl -fsSLO https://zukko.pro/downloads/codex-setup-macos.command
chmod +x codex-setup-macos.command
open codex-setup-macos.command

For Windows, codex-setup-windows.vbs↗: download the file and double-click it.

The script creates both files and asks for your key. If you dislike running scripts from the internet, that is sound caution — set it up by hand from the config above, it is only two files.

Choosing a model

The model field takes any model from the provider's catalogue. Codex is an agent: it makes many calls in sequence and keeps loading context, so input token price matters most.

Rough guide in the Zukko catalogue: gpt-5.6-sol is the flagship ($0.30/$0.30 per million), gpt-5.6-luna the cheaper option ($0.10/$0.10); test which suits you on your own tasks. Nothing stops you putting an Anthropic model in Codex either: the API is compatible, and claude-opus-5-5 works the same way. The current model list is GET https://api.zukko.pro/v1/models.

Errors

404 on every request. A missing /v1 in base_url, or the wrong wire_api.

401 unauthorized. The key was not picked up: check it sits in ~/.codex/auth.json and the field is named OPENAI_API_KEY — Codex looks for that name regardless of provider.

402. An empty balance.

Model not found. Compare the name against the catalogue: curl https://api.zukko.pro/v1/models -H "Authorization: Bearer fetch_your_key".

What it costs

You pay for the tokens you used, with no monthly fee. Zukko↗ accepts card, bank transfer and crypto, and prices sit well below list: Claude Opus 5.5 costs $0.30/$0.30 per million tokens against Anthropic's official $4 input and $20 output. A caveat for agents with a large context: vendors charge ~10% of the input price for prompt-cache reads, while we bill cached tokens at the regular price, so the real gap is smaller than a straight comparison suggests.

To learn what your own Codex will consume, start small: top-ups start at $1. There is no starting balance, but subscribing to the @zukkopro↗ channel credits $0.50 once (link Telegram in the dashboard).