> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fireworks.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Overview
> Route LLM requests between closed-source and open models with FireRouter
FireRouter is a managed routing service for **any LLM workload**. Request a FireRouter model through the [Fireworks inference API](/tools-sdks/openai-compatibility), and FireRouter scores each turn to select a model from the configured set. By default, it routes between a Fireworks open model and a closed-source model.
The result is lower cost on simpler requests without giving up closed-source quality on harder ones.
FireRouter is in **research preview**. APIs, routing behavior, and the models in the routing pair may change. This documentation is updated to reflect the current configuration.
## When to use FireRouter
Use FireRouter when you want **automatic cost optimization** without picking a different model per request:
* You want closed-source quality (for example Claude Opus) on hard prompts but do not need it on every call.
* Many of your requests are straightforward (summaries, formatting, simple Q\&A) and can be served by a Fireworks open model.
* You want to select the models available to the router without choosing a target for every request.
* For the default configuration, you are able to send both a Fireworks API key and a provider API key on each request.
## How it works
FireRouter evaluates each new user request. The default `firerouter` configuration picks one of two paths:
| Path | When | What runs | Billing |
| ---------------- | ----------------------------------------- | ------------------------------------------------- | --------------------------------- |
| **Redirect** | Simple or low-complexity work | A Fireworks open model (for example GLM 5.2 Fast) | Your Fireworks API key |
| **Pass-through** | Hard reasoning, judgment, or long context | A closed-source model (by default Claude Opus 5) | Your provider API key (Anthropic) |
FireRouter uses a bring-your-own-key (BYOK) model:
* Your **Fireworks API key** authenticates to FireRouter and pays for calls to Fireworks models.
* A third-party **provider API key** pays for calls to that provider's models.
FireRouter never stores your provider keys server-side. You send them on each request.
## FireConnect
The easiest way to use FireRouter in coding harnesses is [FireConnect](/ecosystem/fireconnect/overview). As of FireConnect **v0.9.0**, select FireRouter like any other model:
```bash theme={null}
fireconnect login
fireconnect claude on --model firerouter
```
See [FireConnect Models](/ecosystem/fireconnect/models) for how `firerouter` fits next to other short IDs, and [Harness support](/ecosystem/fireconnect/overview#harness-support) for which harnesses support FireRouter. Per-harness details (Claude slot flags, Codex Anthropic env, routing preference) live on each harness page.
Upgrade FireConnect before enabling FireRouter on an older install. See [Upgrade FireConnect](/ecosystem/fireconnect/overview#upgrade-fireconnect).
Pass `--anthropic-api-key sk-ant-...` on `on`, or store a key once with `fireconnect configure --anthropic-api-key sk-ant-...`. The default `firerouter` requires an Anthropic key because Claude Opus 5 is its primary model; FireRouter fails closed instead of silently restricting the route to GLM. See [Authentication](/ecosystem/firerouter/authentication).
Cursor and Deep Agents need workspace BYOK for FireRouter (no local Anthropic key).
## Endpoint
FireRouter is available through the Fireworks inference API:
```text theme={null}
https://api.fireworks.ai/inference/v1
```
Common API paths:
| Wire format | Path |
| ------------------ | -------------------------------------------------------- |
| Chat Completions | `https://api.fireworks.ai/inference/v1/chat/completions` |
| Anthropic Messages | `https://api.fireworks.ai/inference/v1/messages` |
## Model ID
FireRouter is a first-party model on the Fireworks provider. Use `firerouter` in the `model` field:
```text theme={null}
firerouter
```
These longer forms are also accepted:
```text theme={null}
fireworks/firerouter
accounts/fireworks/routers/firerouter
```
For [LiteLLM](/ecosystem/firerouter/litellm), use the full router path in `litellm_params.model`.
FireRouter decides the target model for each request based on request complexity. The model ID selects the set of models FireRouter can route to.
### Current routing pair
The short `firerouter` model ID currently routes between:
| Role | Model |
| ---------------------------- | --------------------------------- |
| Pass-through (closed-source) | **Claude Opus 5** |
| Redirect (open) | **GLM 5.2 Fast** (`glm-5p2-fast`) |
These models are subject to change as FireRouter is updated. This page reflects the current configuration.
Because the default pass-through target is Claude Opus 5, supply an Anthropic API key to use that path. See [Authentication](/ecosystem/firerouter/authentication).
### Choose different models
Use a slash-delimited FireRouter slug to change the models available to the router. The first model is the primary; the remaining models are alternatives that FireRouter ranks for each request. Each member must be an exact deployed model ID or a unique model alias, and a slug can contain up to eight distinct models.
Examples:
| Model ID | Models |
| -------------------------------------------------------- | --------------------------------------------------- |
| `firerouter/kimi-k3/glm-5p2-fast` | Kimi K3 and GLM 5.2 Fast |
| `firerouter/kimi-k3/deepseek-v4-flash-0731` | Kimi K3 and DeepSeek V4 Flash (0731) |
| `firerouter/kimi-k3/glm-5p2-fast/deepseek-v4-flash-0731` | Kimi K3, GLM 5.2 Fast, and DeepSeek V4 Flash (0731) |
| `firerouter/claude-opus-5/kimi-k3/glm-5p2-fast` | Claude Opus 5, Kimi K3, and GLM 5.2 Fast |
| `firerouter/gpt-5.6-sol/glm-5p2-fast` | GPT 5.6 Sol and GLM 5.2 Fast |
Set the selected slug in the `model` field just as you would use `firerouter`. FireRouter still chooses the target for each request.
The primary model's credential is required. A later model is eligible only when its credential is present. Use `x-anthropic-api-key` for Claude models and `x-openai-api-key` for OpenAI models. See [Authentication](/ecosystem/firerouter/authentication).
## Client integrations
| Integration | When to use |
| --------------------------------------------------------------- | --------------------------------------------------------------------------------------- |
| [FireConnect](/ecosystem/fireconnect/overview) | One-command setup for coding harnesses (recommended). Select with `--model firerouter`. |
| [Quickstart](/ecosystem/firerouter/quickstart) | Direct HTTP calls (curl, OpenAI SDK, any OpenAI-compatible client) |
| [Claude Code (manual setup)](/ecosystem/firerouter/claude-code) | Manual `settings.json` setup without FireConnect |
| [LiteLLM](/ecosystem/firerouter/litellm) | Add FireRouter to a LiteLLM Proxy deployment |
## What FireRouter is not
* **Not a deployment router.** FireRouter is different from [deployment routers](/deployments/routers), which load-balance traffic across your own Fireworks deployments.
## Next steps
Enable FireRouter with one command
Make your first API call
Edit settings.json for FireRouter in Claude Code
BYOK headers and API key requirements
Tune the quality vs. savings dial
Add FireRouter to LiteLLM Proxy