Choosing a model
Use our bundled models, or bring your own keys. How to switch and when to.
Every agent runs on a model. You can use the models we bundle, or bring your own API keys (BYOK). You can switch per agent.
Bundled (default)
The simplest option: use the models we ship with your plan. Inference is included in your plan's monthly allowance — no keys to manage. Kimi K2.6 is the free default; the latest frontier flagships — Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro — are included on paid plans (or pay-as-you-go with usage credits).
Bundled inference is available on the Solo, Fleet, and Pro plans. See Plans and billing.
Bring your own key (BYOK)
Prefer a specific model, or already pay a provider? Add your own key and the agent uses it for inference. We don't mark up your tokens — you pay the platform fee, and inference is billed by your provider at their rate.
Supported providers:
- Anthropic (Claude)
- OpenAI (GPT)
- Google (Gemini)
- OpenRouter (a wide catalog behind one key)
- Custom — any OpenAI-compatible endpoint (see below)
Add a key under Settings → Model keys. Keys are encrypted at rest and never written to logs. You can set a different model per agent.
Custom provider
Running a model behind an OpenAI-compatible endpoint — a self-hosted gateway, a regional proxy, or another vendor — add it as a Custom provider. Under Settings → Model keys, choose Custom, then provide:
- the endpoint's base URL (must be HTTPS and publicly reachable — internal and metadata hosts are rejected),
- your API key, and
- the model ids the endpoint serves.
Your key is validated when you save it and encrypted at rest. The agent reaches your endpoint through FlareCode's secure proxy, so your URL and key never enter the sandbox.
Troubleshooting a custom endpoint
- "Key validation failed" on save. FlareCode makes a test call when you save. Check the base URL points at the OpenAI-compatible root (the part before
/chat/completions), the key is current, and the endpoint actually serves one of the model ids you listed. - "Endpoint not reachable." The base URL must be HTTPS and publicly reachable —
localhost, private ranges, and cloud-metadata addresses are rejected for security. Put a self-hosted model behind a public gateway with a real certificate. - The model id isn't offered to an agent. An agent only lists the model ids you registered for the provider. Add the id under Settings → Model keys, then pick it on the agent.
- Calls work in chat but a goal errors. Make sure your endpoint supports streaming and tool/function calling — the autonomous pipeline relies on both. A chat-only endpoint can drive interactive sessions but may stall a goal.
When to switch
- Default / bundled — everyday tasks, cost-predictable work, when you don't want to think about it.
- A frontier model — hard refactors, gnarly debugging, or anything where you want the strongest reasoning. Use a bundled frontier model, or a BYOK key.
- OpenRouter or Custom via BYOK — when you want to experiment across many models, or run your own endpoint.
Per-agent override
Model choice is set when you create an agent and can be changed later. Different agents (or different repos) can run different models — for example, the free bundled model for routine chores and a frontier model for a tricky service.
Whatever you pick, the cost of each task is reported in the app before and after it runs. See Costs and limits.