Fireworks, which hosts open AI models for companies, announced a desktop version of its Nexus service at its Forge conference in November. Nexus is a layer between staff and the models. It sends each request to an open model or to a closed one such as Claude, and it shows finance what each model cost. It already worked inside coding tools; the desktop version brings it to Claude Desktop, the Codex app and OpenCode, so it now reaches everyday office work as well.
A bill that grows with every request
Uber's numbers show the shape of the problem. After it rolled out Claude Code, the average engineer cost $150 to $250 a month, and heavy users cost $500 to $2,000. None of those is a wild sum for one person. Multiplied across thousands of people, and across agents that make many model calls to finish a single task, they used up a year's budget in four months. Uber's chief operating officer later questioned whether the spending was producing anything customers could see.
A router goes after the cause. It judges how hard each request is, keeps the easy ones on a cheap model, and sends only the hard ones to the expensive model. It works like an office where junior staff answer the routine questions and the senior partner sees only the odd clause.
Picture a finance team of 40 that uses an assistant to match invoices, draft supplier emails and read contracts. Most of those requests are routine. At $200 a person each month, the team pays $8,000. If routing moves most of the routine requests to a cheaper model, more than $4,000 of that disappears, and the contract with the strange clause still goes to the best model.
Some models are also built to spend less on every request. Fireworks trained one, Ember-1, to think through problems with about 40% fewer tokens while keeping the quality of the model it came from. A token is the small piece of text a model reads or writes, and it is what the bill counts.
Speed and patience have separate prices
An agent that takes 30 steps waits for the model 30 times. Fireworks has a fast option, now in public preview, that runs at up to 1,500 tokens a second. On a benchmark task of 30 steps, it brought the finish time down from 1 minute to 8 seconds. Nobody would notice that for one question; they would notice it for an agent that does a hundred tasks before lunch.
The other direction is cheaper. Work that can wait, such as overnight research, now runs at half price in Fireworks's Background Mode. That is the logic of off-peak electricity, and the discount for waiting is not new: Fireworks was already selling batch jobs at half price in 2025. What is new is that the discount, the fast lane and the routing now work inside the tools staff already open.
So every AI task now has three settings: how smart the model must be, how fast the answer must arrive, and how soon it is needed. Companies that set those per task will pay a fraction of what companies pay who leave everything on the default. Within two years, finance teams will budget AI the way they budget phone plans, per person, with rules about what goes where.
Before that reaches your team, list the ten things it asks AI most often. Mark the ones where a wrong answer would cost real money. Those belong on the best model, and a router should never be allowed to move them. The rest are routine, and routine work is where the savings are.