If you want Cursor to run a model on your own key, the plan decides before the settings do. Cursor's pricing docs, read on 16 November 2026, exclude your own API keys on the Free and Plus plans; own keys start at Premium, $30 a month. We tried anyway, on Cursor 3.24.18 with a Free account. The key saved. The model never showed up.
So we ran one bug fix through three terminal agents that take any OpenAI-compatible endpoint. All three passed on Kimi K3.1 Pro, and the most expensive run billed under 10 cents. Configs, timings and bills are below.
Disclosure: Wallaby Token sells API access to Kimi K3.1 Pro, and the endpoint in these tests is ours. Every run used a fresh config directory and a test key on our live gateway, and the bills come from our gateway log.
What Cursor allows on each plan
From Cursor's pricing and bring-your-own-key docs:
- Free: Agent and Chat with limited usage. No third-party models and no own API keys.
- Plus, $8 a month: Cursor's own models, the Grok 4 series and Composer 2.5. Still no third-party models and no own API keys.
- Premium at $30, Super at $100, Ultra at $200 a month: third-party models at API list price from prepaid credit, plus own keys. On these individual plans, own-key requests do not draw from included usage.
- Teams: own-key requests on eligible models still carry Cursor's token rate of $0.25 per million tokens.
The BYOK page lists five providers: OpenAI (standard, non-reasoning chat models only), Anthropic, Google, Azure OpenAI and AWS Bedrock. Own keys cover chat, and Tab completion stays on Cursor's models. Every request also passes through Cursor's servers with your key attached, so an endpoint on localhost is out of reach.
What we saw on Cursor 3.24.18, Free plan
The settings page still has the older controls, though the BYOK page no longer mentions them: an OpenAI API key field, a Use OpenAI API Key toggle, and Override OpenAI Base URL. We pasted a test key, pointed the base URL at https://api.wallabytoken.com/v1, and the page answered "Secret saved".

Then we added kimi-k3.1-pro as a custom model and searched for it in the chat model picker. The answer was "No models found", with Cursor's built-in Kimi K3 Max still selected underneath.

Our gateway log shows no request from Cursor on that key after the settings were saved. On the Free plan, the setup stops at the picker. We did not test Premium; by Cursor's docs, that is the first plan where own keys apply.
If you hit Cursor's usage limit
Cursor sizes its paid plans by included usage: Plus is the base, Premium four times that, Super 15 times, Ultra 50 times. On Free or Plus, running out does not open the own-key route, because those plans exclude it.
That leaves two ways out. Move up to Premium and add a key from one of the five listed providers, or run the agent work in a tool that takes any OpenAI-compatible endpoint and keep Cursor for editing and Tab.
Three agents that ran our endpoint
The test was a three-line Python function with a bug and five unit tests, plus one prompt: run the tests, fix the bug without touching the test file, run them again. Each agent ran Kimi K3.1 Pro on 16 November with no human in the loop.
| Agent | Version | Tests after | Time | Requests | Input / output tokens | Billed |
|---|---|---|---|---|---|---|
| Codex CLI | 0.154.0 | 5 of 5 pass | 47 s | 3 | 29,032 / 638 | $0.034 |
| opencode | 1.19.5 | 5 of 5 pass | 76 s | 6 | 41,831 / 1,075 | $0.070 |
| crush | 0.98.1 | 5 of 5 pass | 68 s | 7 | 57,415 / 888 | $0.097 |
Input counts include cached tokens, which bill at $0.18 per million instead of $1.80. Codex's own footer said "tokens used 4,199" for its run; the gateway billed 29,670. Read the bill, not the footer.
For a deeper look at what these three bills mean for tool choice, see Codex vs OpenCode vs Crush: 2.9x cost gap.
Codex CLI
Codex loads custom providers from a profile file. Save this as ~/.codex/wallaby.config.toml:
model = "kimi-k3.1-pro"
model_provider = "wallaby"
[model_providers.wallaby]
name = "Wallaby"
base_url = "https://api.wallabytoken.com/v1"
env_key = "WALLABY_API_KEY"
wire_api = "responses"
Export WALLABY_API_KEY, then run codex -p wallaby, or codex exec -p wallaby "your task" for a one-shot run. The full walkthrough is in Kimi K3 in Codex CLI.
opencode
opencode's built-in catalog lists Wallaby. On 16 November it had only Kimi K3, so our run used the custom provider below. Since 18 November the catalog also lists wallaby/moonshotai/kimi-k3.1-pro, and opencode run -m wallaby/moonshotai/kimi-k3.1-pro "your task" works with no config file (on opencode 1.x, run opencode models --refresh first). The custom provider still works too; put it in opencode.json at the project root:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"wallaby-k31": {
"npm": "@ai-sdk/openai-compatible",
"name": "Wallaby",
"options": {
"baseURL": "https://api.wallabytoken.com/v1",
"apiKey": "{env:WALLABY_API_KEY}"
},
"models": { "kimi-k3.1-pro": { "name": "Kimi K3.1 Pro" } }
}
}
}
Then run opencode run -m wallaby-k31/kimi-k3.1-pro "your task", or pick the model with /models in the TUI.
crush
crush reads a crushrc from the project root:
provider add wallaby --type openai-compat \
--base-url "https://api.wallabytoken.com/v1" \
--api-key "$WALLABY_API_KEY"
model add wallaby/kimi-k3.1-pro \
--name "Kimi K3.1 Pro" \
--context-window 1048576 \
--default-max-tokens 8192 \
--price-input 1.80 \
--price-output 9.00 \
--price-cache-hit 0.18
Run crush run -m wallaby/kimi-k3.1-pro "your task". In our non-interactive run it edited the file and ran the tests without a permission prompt, so point it at a branch you can throw away.
Claude Code
If you would rather keep Claude Code's interface, it takes an Anthropic-compatible endpoint. We tested that setup on 15 and 16 November in Claude Code usage limit reached? How to keep coding.
Which one to pick
On our single task, Codex made the fewest requests and billed least; crush made the most. One task is not a benchmark. Point two of them at your own repo for a day and compare the gateway log before you settle.
If you want to stay inside VS Code rather than a terminal, Cline and Kilo Code also take a custom OpenAI-compatible endpoint. We did not re-run them for this post.
What Cursor still does better
Tab completion, and an editor built around it. Cursor's Tab runs on Cursor's models whatever key you add, and none of the three agents above completes code as you type. Cloud agents and Bugbot reviews live on Cursor's side too.
One setup keeps both: Cursor for editing and Tab, and a terminal agent on your own key for the long tasks that would otherwise eat the plan's usage.
Try it
New Wallaby accounts start with $0.50 in free credit, enough for five or more runs like ours at the bills above. Kimi K3.1 Pro is $1.80 in, $0.18 cached and $9.00 out per million tokens, 10% off the official list. Create an account.
Sources: Cursor pricing and bring-your-own-key docs, read 16 November 2026; Cursor 3.24.18 on a Free account, tested 16 November 2026; Codex CLI 0.154.0, opencode 1.19.5 and crush 0.98.1, tested 16 November 2026 against api.wallabytoken.com; Wallaby gateway log for the same runs. Plans and versions change; re-check before you buy.