Skip to main content
Nexus Auto gives you one model ID, auto, and picks a model for each new user turn. Easy turns can go to a Fireworks open model, and harder turns to a closed model such as Claude Opus or GPT Sol. You pay the selected model’s rate for each turn instead of a closed model’s rate on every turn.
Nexus Auto was formerly called FireRouter, and its model ID was firerouter. FireRouter and the firerouter model ID are no longer available. Use auto everywhere, including in model pickers, SDK calls, and gateway configs.

​
How Nexus Auto works

For each new user turn, Nexus Auto predicts which available model will handle it well at the lowest cost. That model serves the whole turn, including its tool calls, until the next user turn. Open models run on Fireworks serverless. Closed models run on your own Anthropic or OpenAI account. The response model field names the model that served the turn, such as glm-5p3 or claude-opus-5-5. Nexus Auto works with the Chat Completions, Messages, and Responses APIs. Fireworks updates the models behind auto after evaluating them, so your model ID stays the same.

​
Choose a model ID

Use auto for most work. To choose a model yourself, pin one.

​
How auto picks its closed model

auto picks one closed model from the credentials available to the request. It then picks between that model and the open-model mix for each turn. The open-model mix is GLM 5.3 and GLM 5.3 Flash today. Fireworks manages it and updates it as new open models are evaluated. Easy turns usually go to the mix and harder turns to the closed model. Set effort in the harness to ask for a lighter or more thorough attempt. If you connected Amazon Bedrock, Nexus Auto prefers a Bedrock model from the family it picked, which can be GPT Astra or an earlier Opus or Sol version. To keep Nexus Auto on open models even when you have Anthropic or OpenAI credentials, limit it to Fireworks, then run your harness command again, replacing claude with your harness:
The setting is global. Nexus CLI saves it in ~/.nexus/config.json and writes it into each harness the next time you connect that harness. fireworks is the only value --providers accepts today. Run nexus auto on its own to see the current setting, and nexus auto --reset to go back to the default. Your account’s model restrictions always apply first. The setting has no effect in Cursor IDE.

​
Set up

Every setup uses auto and your Fireworks API key. Only the way you provide the closed-model credential changes.

​
Nexus CLI

Nexus CLI writes your coding harness settings and restores them with off. To change the files yourself with nothing to install, use manual setup. Pick a harness, a model ID, and a setup method. The commands and config files update as you choose, including any provider-key headers. A connected Provider Key adds Anthropic and OpenAI models in every harness except Claude Desktop. Claude Code and Codex can also pay with their own sign-in: The Claude Agent SDK matches Claude Code when it reads the user settings that Nexus CLI writes. Claude Desktop’s picker lists auto and Fireworks models only. The --anthropic-api-key and --openai-api-key connect flags only worked with FireRouter routes. Apart from the Claude Code and Codex sign-ins, Nexus CLI doesn’t send a local key on auto requests, including keys saved with nexus configure and exported ANTHROPIC_API_KEY and OPENAI_API_KEY values. Without a Provider Key, use manual setup to send x-anthropic-api-key and x-openai-api-key headers yourself. Cursor IDE and the Copilot App cannot send extra headers, so they need Provider Keys. To clear a key saved by an earlier version, run nexus configure --unset anthropic-api-key or nexus configure --unset openai-api-key. Live inference is tested in Claude Code, OpenCode, Codex, Pi, and DeepSeek Harness. The other harnesses are tested at the configuration layer. For a full walkthrough, see Claude Code. For each harness’s files and restore behavior, see Coding Harnesses.

​
LLM gateway

The gateway sends auto and your Fireworks API key to Fireworks, so you do not need Nexus CLI. Clients behind the gateway call auto with their gateway key only. Choose LiteLLM or Portkey, a model ID, and where the provider keys live. The widget writes the gateway config and a test request. For other gateways and each credential option, see LLM Gateways.

​
API and SDK

Point the SDK you already use at Fireworks. Its API key is your Fireworks API key. For an Anthropic or OpenAI model, connect a Provider Key or send the provider’s key as a header (see Provide Anthropic and OpenAI credentials). Anthropic-style clients add /v1/messages to the base URL themselves, so their base URL stops at /inference.
Pick an API and a language. The code adds the provider header that API needs, or leaves it out if you have Provider Keys connected. The Anthropic SDK’s base URL is https://api.fireworks.ai/inference, without /v1. Some models return a thinking block before the text, so the snippet reads the first text block rather than content[0]. For errors and limits, see APIs and SDKs.

​
Provide Anthropic and OpenAI credentials

Nexus Auto calls closed models through your own Anthropic, OpenAI, or Amazon Bedrock account. For Anthropic and OpenAI, store the key once as a Provider Key or send it with each request. Bedrock uses Provider Keys only, with each model mapped to a Bedrock model ID and region. A request header takes precedence over the stored Provider Key for that provider.
Provider Keys is not yet available on every account. If Provider Keys does not appear in Settings, contact the Fireworks team to enable it. Self-serve setup is coming soon.
Without a credential for a closed provider, Nexus Auto leaves that provider’s models out and serves the turn with the others. A request returns 400 no_credential when you call a closed model ID such as claude-opus-5-5 directly and that credential is missing.

​
Check which model served a turn

  1. Send a request with auto:
  2. Check the response model field. It names the model that served the turn, such as glm-5p3 or claude-opus-5-5.
  3. An easy prompt usually names an open model. A harder prompt, such as debugging a concurrency bug or writing a proof, is more likely to use the closed model. No single request is guaranteed to use either. In a coding harness, set effort to ask for a lighter or more thorough attempt.
  4. To confirm a closed-model credential, call that model ID directly, such as claude-opus-5-5. A 400 no_credential error means the credential is missing.

​
Watch orchestration live

In Claude Code, you can watch each turn’s model choice as you work. Connect Claude Code with a Nexus Auto model, install tmux, then run:
Claude Code opens on the left, and a live meter on the right adds one row per user turn. See Usage and Cost for the columns and pricing. To follow a specific session, run nexus claude live --session <id>.

​
Balance quality and savings

Set effort in the harness. Higher effort asks for a more thorough attempt. Lower effort asks for a lighter one. Nexus Auto uses that when it picks a model.

​
Pay for the model that serves each turn

Open models bill to your Fireworks account at their serverless rates. Closed models bill to the Anthropic or OpenAI account whose credential served the turn. In the Codex CLI, a ChatGPT sign-in covers most OpenAI model turns; see Who pays for OpenAI model turns. Cached input bills at the serving model’s cached-input rate. Use Usage and Cost to review Fireworks usage and Claude Code session estimates.

​
Supported models

auto uses Claude Opus or GPT Sol plus the open-model mix. To send turns to any other model below, pin it. Anthropic serves Claude Fable only to organizations with data retention enabled. Otherwise, Fable requests return a 400 error from Anthropic.

​
Pin a specific model

To choose a model yourself, use its ID instead of auto, in your harness with --model or as the model value in a request:
A pinned model serves every turn, with no orchestration. A full model ID such as glm-5p3 stays on that version. Aliases such as glm-latest track the current version in a family and are called directly, without Nexus Auto; see Open Models.
Custom routes such as firerouter/opus are no longer available, and Nexus Auto does not accept auto/... route IDs. Pin a model ID as shown above, or use auto.

​
Limitations

  • Nexus Auto is not available through Microsoft Foundry. See Nexus for Microsoft Foundry.
  • Accounts with data residency enabled cannot use Nexus Auto. Use a residency-compatible pinned serverless model instead.

​
Troubleshooting

No closed-model credential is reaching Nexus Auto. Connect the matching Provider Key, or send x-anthropic-api-key and x-openai-api-key headers with manual setup. Keys saved with nexus configure don’t reach Nexus Auto. In Codex, OpenAI model turns need a Codex sign-in or an OpenAI Provider Key.