Fireworks Nexus Desktop: Pricing, Model Routing and App Support

Fireworks launched Nexus Desktop on November 3, a macOS app for connecting existing AI tools to its Nexus Auto model router. Pro starts at $8 per month and Team at $10 per seat, but those prices do not replace the bill for every model used through the service.

That distinction is the first thing to understand. Nexus sits between a familiar application and the models serving its work. It can make model choice less manual, while leaving usage split across accounts and providers.

What Nexus Desktop changes

The Forge announcement introduces Desktop alongside the existing Nexus CLI. Instead of configuring each tool from a terminal, users can enable supported applications through switches and restart them to make Nexus Auto available.

According to the desktop documentation, the macOS download supports Apple silicon and Intel Macs. Windows and Linux desktop versions are coming later. The CLI remains the alternative, with limited Windows support through Git Bash.

Supported tools include Claude Code, Codex, ChatGPT, Cursor, OpenCode and others. Desktop and CLI share settings. Turning a tool off and restarting it restores its original setup.

This is useful for someone who already has a preferred coding or knowledge-work interface. A routing layer is easier to evaluate when it does not also require learning a new editor. The quality test, however, remains whether the selected models perform well inside that particular tool.

Nexus pricing and what the plans include

The plan documentation lists three options:

Plan	Listed price	Important limit
Pro	$8 per month	One person, included Fireworks usage
Team	$10 per seat per month	Up to 200 seats, more included usage per seat
Enterprise	Usage-based	Enterprise controls and direct access to serverless open models

Pro and Team include Nexus Auto, but not general direct selection of Fireworks open models. Auto can still route work to those models. Enterprise adds the ability to select them directly.

Anthropic and OpenAI turns use the customer's own provider account and are billed there. Included Nexus usage counts turns that run on Fireworks; on Pro and Team, those turns pause when the allowance is exhausted until the relevant window resets. Nexus accounts and billing are also separate from the Fireworks platform used for inference and training.

The subscription price is one part of the total cost. Comparing it with an existing AI subscription requires looking at which provider serves the work, how frequently the router uses that provider and whether the included Fireworks allowance is enough.

How Nexus Auto chooses models

Fireworks' Auto guide describes routing easier turns to open models and more demanding turns to Anthropic or OpenAI models. Closed-model access requires the relevant account credentials. Without them, the router uses the models it can access.

The response's model field identifies what served a turn. That visibility matters because a router introduces another variable into evaluation: a good result does not necessarily tell you which model will handle a similar request next time.

For a team, a sensible comparison is the full workflow. Measure accepted outputs, retries, latency and total spend against the current setup. A lower-priced first attempt can be less economical if it creates more review work or needs several corrections.

Claude Desktop has a more restrictive path

The Claude Desktop integration guide describes limited macOS support through the CLI and a separate local workspace. While using Nexus, the app cannot remain signed in to claude.ai, and existing claude.ai chats, projects, memory and scheduled tasks are hidden rather than deleted. New work is saved locally.

That detailed guide is narrower than the Desktop overview, which lists Claude Desktop among connected tools. Users should follow the integration-specific instructions instead of assuming every listed app has the same one-switch experience. Existing connectors can also need fresh sign-in.

This makes Nexus most interesting as an experiment in moving model choice closer to the task while keeping the working interface familiar. Whether it becomes a useful everyday layer depends on how much routing can save without making billing, context and tool behavior harder to understand.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain product launches and technical tradeoffs in stories that developers and founders want to read and share.

Book a 30-minute call.