Fireworks AI has launched Fireworks Nexus, a management and routing platform designed to shift routine coding tasks from expensive frontier models to cheaper open-weight alternatives.
The motivation is financial pressure. Forbes reported that Uber exhausted its 2026 AI budget within four months after rolling out Claude Code to roughly 5,000 engineers in December. Agentic adoption among its staff climbed from one-third to more than four-fifths in just two months.
Fireworks argues the issue is a pricing mismatch rather than overspending. Most organisations run standard work at frontier prices. Operational complexity has previously discouraged platform teams from switching to open-weight models.
Fireworks Nexus
The system consists of three parts.
First, enterprise controls and cost observability allow teams to set budgets at the company or team level. They can track return on investment across different models and tools while enforcing policy from a single location. Requests run on the Fireworks production inference platform with US-hosted endpoints, zero data retention, and coverage across 20 global data centers.
Second, workflow continuity relies on FireConnect. This is a one-line installation that maps harness model slots to Fireworks models. Released under the Apache 2.0 licence, it installs from the Fireworks Dashboard in a single command. Tools like Claude Code, Codex and OpenCode continue to function unchanged. FireConnect runs on Fireworks Serverless APIs compatible with Anthropic and OpenAI standards, allowing most tools to connect using a base URL and model ID.
Third, intelligent traffic management uses a custom trained model to score request difficulty. Routine requests go to a cost-effective open-weight model served by Fireworks. Difficult requests pass through to your existing provider on your own key, which Fireworks states is never stored server-side. The research team reports this typically delivers a 3–5× cost reduction.
A note for those planning a rollout: the router is in research preview. It currently routes between Claude Opus 5 and GLM-5.2, so the pass-through path requires an Anthropic key. An all-open configuration routes between Kimi K3 and GLM-5.2 instead.
Performance
The Fireworks research team tested Nexus with developer teams including Notion and Doximity. Preliminary results show a one-third reduction in cost per merged pull request. The research team also reports a blended token rate roughly a quarter of the closed model labs. These are vendor figures from a preview program and should be read as such.
The more useful evidence comes from two independent evaluations Fireworks cites.
Faros AI ran 211 real engineering tasks drawn from 12 repositories across seven model-and-harness routes. Claude Code on GLM-5.2 scored 0.568 on a model-based rubric judge. Claude Code on Opus 4.8 scored 0.521. The cost gap was larger than the quality gap: $0.92 per task against $1.76. Cache share was 89.7% and 99.7% respectively, so caching does not explain the result. Faros stresses that the cohort is company-specific and not a universal leaderboard.
Arize, in a joint study with Fireworks, evaluated 10 models on 40 Terminal-Bench tasks at six trials each. That is 2,400 runs and $626 of API spend, graded by each task’s own test suite. The headline metric is cost per successful task, which counts every failed and retried attempt.
Two findings from that study bear directly on the Nexus design. On easy tasks the frontier premium buys nothing: Kimi K2.6 passed 73% where GPT-5.5 passed 69%. On hard tasks only the top tier competes, with GPT-5.5 at 51% and Kimi K3 at 32%. Arize’s evaluation harness is open source, so teams can rerun it on their own workloads.
Routing simulated over those runs beat every single-model strategy. 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 reliably solved 25 of 40. Naive escalation through all ten models cost $1.319, worse than any single model tested. Ladder design is not optional.
How teams actually turn it on
There are three practical paths, and they have different blast radii.
The FireConnect path is the lowest friction. The installer requires the Claude Code CLI already on PATH, prompts once for a Fireworks API key, registers a plugin marketplace, and writes Claude Code settings with a timestamped backup. It exposes /fireconnect:on, /fireconnect:off, /fireconnect:setup, /fireconnect:models and /fireconnect:set-models.
The second path skips the plugin entirely. Set ANTHROPIC_BASE_URL and a Fireworks key, or use the OpenAI-compatible client with a changed base URL and model ID. Arize noted that adding a model to its harness was one config line. That interchangeability is what makes an escalation ladder buildable.
The third path is the router itself, which sits in front of an existing frontier contract.
What it means
Engineers do not need to change their workflow to see the cost drop. The system handles the routing automatically. A developer writes code as usual, but the platform sends simple tasks to cheaper models and reserves expensive ones for complex problems. This keeps the quality high while the bill stays low.
