Fireworks Nexus routes easy code tasks to cheap models, claims 3-5x cost cuts

The numbers coming out of engineering orgs lately are wild. Uber burned through its entire 2026 AI budget in four months. Claude Code hit 5,000 engineers almost overnight. The problem isn’t that companies are spending money — it’s that they’re paying frontier prices for routine work. Fireworks AI just released Nexus, and the pitch is refreshingly direct: stop sending every coding request to the most expensive model on earth.

Nexus is three pieces bolted together. Enterprise controls let teams set budgets and track ROI across models. FireConnect, released under Apache 2.0, is a one-line install that maps existing coding tools to Fireworks’ managed layer so Claude Code, Codex, and OpenCode keep working exactly as they did. The interesting bit is a custom-trained router that scores each request by difficulty. Easy stuff goes to a cheaper open-weight model. Hard stuff gets passed through to your existing frontier provider using your own API key, which Fireworks says it never stores. The research team claims a typical 3-5x reduction. One catch: the router is still a research preview and currently shuttles between Claude Opus 5 and GLM-5.2, so you’ll need an Anthropic key. An all-open configuration routes between Kimi K3 and GLM-5.2 instead.

Two independent evaluations give the approach some weight beyond vendor claims. Faros AI ran 211 real engineering tasks across 12 repositories and found Claude Code on GLM-5.2 slightly outperformed Claude Code on Opus 4.8 — a score of 0.568 versus 0.521 — while costing roughly half as much. Arize tested 10 models over 2,400 runs on Terminal-Bench tasks and found something Nexus is built on: easy tasks don’t need frontier models. Kimi K2.6 passed 73% of easy tasks where GPT-5.5 managed 69%. On hard tasks, GPT-5.5 hit 51% and Kimi K3 landed at 32%. The gap is real. Arize simulated routing and found a deliberate escalation ladder reached $0.525 per successful task while reliably solving 32.3 of 40 tasks. GPT-5.5 alone cost $0.636 and solved 25. Naive escalation through all ten models cost $1.319 — worse than any single model. Ladder design matters.

There are three ways teams can turn this on, and they range from zero-friction to full-router. The FireConnect path installs in one command, registers a plugin marketplace, and exposes slash commands for toggling routing on and off. The second path is even simpler: change a base URL and model ID. Arize noted adding a model to its harness was one config line. The third path puts the router in front of an existing frontier contract. Preliminary results from Notion and Doximity show a one-third reduction in cost per merged pull request. Those are vendor figures from a preview program, so calibrate expectations accordingly. The Faros and Arize data are the more useful signal — and they suggest the core idea holds up, at least for the cohort tested.