Fireworks Nexus: маршрутизация кода в open-weight модели и снижение AI-расходов в 3–5 раз

Fireworks AI released Fireworks Nexus on July 28, 2026 — a routing and spend-control platform that sends routine coding tasks to open-weight models and, according to the developer, cuts costs by 3–5x without changing familiar tools like Claude Code.

Why This Was Needed

AI agent spending is growing faster than companies can plan for it. According to Forbes, Uber burned through its entire 2026 AI budget in just four months. Fireworks cites the same report: adoption of agentic tools among engineering teams grew from roughly a third of staff to more than four-fifths in two months, and Claude Code itself reached around 5000 engineers after its launch in December 2025.

Fireworks frames the root of the problem as a mismatch, not overspending: most organizations do routine work at "frontier" prices because switching to open-weight models is operationally inconvenient for platform teams.

How Fireworks Nexus Works

Fireworks Nexus consists of three parts and plugs in on top of the development tools teams already use. Key features:

• Control and observability — team- or company-level budgets, ROI tracking by model, US-hosted endpoints, zero data retention, and coverage across 20 global data centers.
• FireConnect — a one-line install under the Apache 2.0 license; it maps model slots in the harness to Fireworks models. Claude Code, Codex, and OpenCode keep working unchanged via Anthropic- and OpenAI-compatible APIs.
• Smart routing — a trained model scores the complexity of each request: routine work goes to a cheap open-weight model, while complex tasks go to your previous provider using your own key (which, according to Fireworks, is not stored on the server).

The router is currently in research preview status. Right now it routes between Claude Opus 5 and GLM-5.2, so the pass-through path requires an Anthropic key; the fully open configuration routes between Kimi K3 and GLM-5.2.

How Much Cheaper Is It?

The savings in independent tests turned out to be significant and are not explained by caching alone. Faros AI ran 211 real engineering tasks from 12 repositories across seven "model+harness" routes. Claude Code on GLM-5.2 scored 0.568 on the model-judge rubric versus 0.521 for Claude Code on Opus 4.8 — at a cost of $0.92 versus $1.76 per task. The cache-hit share was 89.7% and 99.7%, respectively.

"This cohort is specific to a particular company and is not a universal leaderboard,"

Faros AI cautions based on its measurements.

A second study by Arize, together with Fireworks, covered 10 models on 40 Terminal-Bench tasks across six runs — 2400 runs and $626 in API spend. On easy tasks, frontier models offer no advantage: Kimi K2.6 passed 73%, GPT-5.5 — 69%. On hard tasks, top-tier models lead: GPT-5.5 — 51%, Kimi K3 — 32%. Simulated routing beat every single-model strategy: an escalation ladder cost $0.525 per successful task and consistently solved 32.3 out of 40, while a lone GPT-5.5 cost $0.636 and solved 25 out of 40.

What This Means

Fireworks Nexus turns model choice into a manageable parameter: routine work goes to cheap open-weight engines, while expensive frontier models are reserved for genuinely hard tasks. The numbers come from the vendor and its preview program, but independent runs from Faros AI and Arize back up the core claim: for a share of tasks, overpaying for a top-tier model doesn't buy you better quality.

FAQ

Does Fireworks Nexus work with Claude Code?

Yes. The FireConnect component is installed with a single command from the Fireworks dashboard and maps model slots, so Claude Code, Codex, and OpenCode keep working unchanged via Anthropic- and OpenAI-compatible APIs.

How much does Fireworks Nexus cut costs?

According to Fireworks, the typical reduction is 3–5x. In tests with the Notion and Doximity teams, the cost per merged pull request dropped by roughly a third, and the average token price came out to around a quarter of that of closed-model labs.