# Set up Codex CLI with InferSpeed: pay-per-token API

Add an `inferspeed` provider to `~/.codex/config.toml`, put your key in the `INFERSPEED_API_KEY` environment variable, then run `codex`. When your subscription allowance runs out, this lets you keep going and pay per token.

Last checked: 2026-10-14 (the date the site verified this config) · Codex CLI 0.162.1

## One-command setup

```bash
export INFERSPEED_API_KEY="sk-****"     # create one at https://inferspeed.ai/keys
curl -fsSL https://inferspeed.ai/setup.sh | sh -s -- codex --model glm-5.3 --dry-run   # see what it would change
curl -fsSL https://inferspeed.ai/setup.sh | sh -s -- codex --model glm-5.3
```

The script changes only `~/.codex/config.toml`, backs it up first and shows the diff. If you already have a config, it adds an `inferspeed` profile: start it with `codex --profile inferspeed`, and your existing settings stay as they are.

## Manual setup

```toml
model = "glm-5.3"
model_provider = "inferspeed"

[model_providers.inferspeed]
name = "InferSpeed"
base_url = "https://api.inferspeed.ai/v1"
wire_api = "responses"
env_key = "INFERSPEED_API_KEY"

[tools]
web_search = false  # the gateway does not run hosted tools
```

- `wire_api = "responses"`: Codex CLI uses the OpenAI Responses API (`POST /v1/responses`).
- Keep the key in an environment variable, never in project files. Anyone with the key can spend your balance.
- One key calls every live model. To use `kimi-k3`, set `model` to `kimi-k3`; the same key works.

## Check that it works

```bash
codex exec "print the word ok" < /dev/null
```

Usage and balance: call `GET https://api.inferspeed.ai/v1/usage` with your key. It returns today's and total tokens and the amount billed (`actual_cost`).

## Common errors

| Error | Cause | Fix |
|---|---|---|
| Warning that there is no metadata for `glm-5.3` | Codex does not know this model name | Ignore it; Codex still works |
| `codex exec` never returns | When stdin is not a terminal, it waits for more input | Add `< /dev/null` to the command |
| 400 `model_not_available` | No live model answers to this name | `available_models` in the response lists the ones you can call; change the model |
| 403 `INSUFFICIENT_BALANCE` | The balance is used up | Top up, then retry; retrying alone will not work |

Without `model = ...` Codex asks for its default `gpt-*` model; that name is routed to `glm-5.3`, and the response's `model` field names the model that served it.

## Prices (USD per 1M tokens)

| Model | Input | Output | Cache read | Official price (input / output) |
|---|---|---|---|---|
| glm-5.3 | $0.84 | $2.64 | $0.156 | $1.40 / $4.40 |
| kimi-k3 | $1.80 | $9.00 | $0.18 | $3.00 / $15.00 |

All models and live prices: https://inferspeed.ai/model-plaza

## FAQ

**Will my existing Codex config be changed?**
No. With an existing `config.toml`, the setup script only adds an `inferspeed` provider and profile and leaves everything else alone; starting Codex without `--profile inferspeed` uses your original config.

**What are the default limits?**
30 requests per minute and 5 concurrent requests per account.

**How do I check that billing is right?**
Call `GET https://api.inferspeed.ai/v1/usage` with the same key to see tokens and the amount billed. How we check billing (billed amount = billed tokens × published price) and a script to reproduce it: https://inferspeed.ai/benchmarks.md

We are the operator of InferSpeed.