🔌 Kilo Code Models & Pricing
Kilo doesn't lock you to one model vendor. By default it reaches hundreds of models through its own Kilo Gateway, billed at the provider's rate with no markup. You can also bring your own keys, sign in with a ChatGPT or SuperGrok subscription, or run fully local models. This page covers the routes, the Auto tiers, and what it actually costs.
The previous version said Kilo routes through OpenRouter by default and configures models in ~/.kilo/config.toml. Neither is true today. Everything below is checked against Kilo's own documentation.
Four ways to reach a model
- Kilo Gateway (default when signed in): one account, hundreds of models, provider API rates with no markup, paid from Kilo credits.
- BYOK on the Gateway: add your own Anthropic, OpenAI, Google, Bedrock, Mistral, DeepSeek, Z.ai or other key, or a coding-plan key such as Z.ai Coding Plan or Kimi Code, at app.kilo.ai/byok. You keep Kilo features like cloud agents and code reviews while paying through your own account.
- Direct providers (30+): OpenRouter, Anthropic, OpenAI, Gemini, Vertex, Bedrock and more, configured in the extension or
kilo.jsonc. ChatGPT Plus/Pro and SuperGrok logins work in the IDE and CLI but not for cloud features. - Local: Ollama or LM Studio. To keep everything local, set the main model to the local provider, set
small_modelto a model on the same provider, and log out of the Kilo Gateway.
Auto Model tiers
| Tier | What it does | Cost |
|---|---|---|
kilo-auto/frontier | Latest and most capable paid models, with different picks for reasoning and implementation | Paid |
kilo-auto/efficient | Grades each request by difficulty and sends it to the cheapest model proven accurate enough. You can limit it to a pool of 1–10 models you trust | Paid |
kilo-auto/free | The best currently free models | Free |
Auto Free warning (from Kilo's docs): it may route to providers that log prompts and outputs to improve their services, including NVIDIA's free trial endpoints. Don't send personal or confidential code through it.
Which model does what
- Main model: the one you pick (
modelinkilo.jsonc, or/models). - Small model (
small_model): session titles, commit messages and prompt enhancement. If unset, it uses a cheap sibling on your provider, thenkilo-auto/small. Set it to a free model to avoid spending credits on background tasks. - Subagent model (
subagent_model): the default for subagents; it inherits the main model if unset. - Compaction model (
agent.compaction.model) and a separate autocomplete model. Free autocomplete is possible with a Mistral Codestral BYOK key.
Kilo publishes live usage data on what models people succeed with at kilo.ai/models. Current per-token prices for 40+ models are in our cost calculator.
What it costs
| Option | Price |
|---|---|
| Kilo Code itself | Free (MIT). You pay only for inference and cloud compute |
| Kilo credits | $1 of credit = $1 of usage at provider rates; 5% payment-processing fee on purchases |
| Kilo Pass Starter / Pro / Expert | $19 / $49 / $199 a month, with bonus credits worth up to 50% more ($26.60 / $68.60 / $278.60) |
| Teams | $15 per user per month; Enterprise is custom |
| Cloud Agent compute | $0.60/hour (Docker/Small), $1.20/hour (Standard), billed per second |
| Free use | kilo-auto/free or any model marked "(free)", plus a free small model and BYOK Codestral for autocomplete |
From kilo.ai/pricing and Kilo's Adding Credits page, checked 2026-09-27. Auto top-up is available so a long task doesn't stall when your balance runs out.
Sources: Kilo Code documentation, kilo.ai/pricing and Kilo-Org/kilocode releases (v7.8.1, 2026-09-25). Checked 2026-09-27.
← Back to the Kilo Code hub