Codex

Codex speaks the Responses format. Add a provider to ~/.codex/config.toml:

model_provider = "binference"
model = "openai/gpt-6.1-sol"
web_search = "disabled"

[model_providers.binference]
name = "bInference"
base_url = "https://binference.io/api/v1"
env_key = "BINF_API_KEY"
wire_api = "responses"

Codex's built-in image tool is not served here: bInference leaves it out of each request, and everything else works as usual. Codex also offers web search on every call, with no limit on how many searches one call may run, so bInference would have to set aside the cost of many rounds before the call starts (over $10 even on a small model). web_search = "disabled" keeps each call's reserve small; with plenty of credit you can leave it on.

Then start Codex with your key in the environment:

export BINF_API_KEY="binf_..."
codex

Any listed model works as model. Codex retries 503 answers, so a busy model or a short wait on our side resolves on its own.