blog01

Flux Blog

Tips, benchmarks, and Flux updates.

Posts cover run-level budgets, step-aware routing and how the Flux savings benchmark was measured. New to Flux? Start with the Flux documentation or the comparison with OpenRouter and LiteLLM.

engineeringSep 1, 2026

Show HN: Flux, a self-hosted LLM router with per-agent-run budgets

Why request-level routing isn't enough for agents, and how run-level budgets work

A router asks which model should handle this request. Flux also asks what an agent's next call should be allowed to cost, given what it's already spent. Includes the live 21-step multi-agent test, the routing table's known errors, and one open question for feedback.

Read Article →
topics02

What we write about

Flux is a self-hosted LLM gateway for AI agents and AI-powered apps. The blog covers the problems it is built around and is open about what has and hasn't been proven yet.

Run-level budgets

Why a budget should belong to a whole agent run rather than a single request, and how Flux degrades to cheaper models from 70% of the run budget, warns at 90% and hard-stops with a 429 at 100%. The per-run budget example shows the setup.

Step-aware routing

How planning, tool selection, tool-result interpretation, reflection and final answers each get their own routing treatment, so agents keep premium models for the steps that need them. See how this differs from per-request gateways.

Benchmarks and limits

How the offline synthetic benchmark was run, what up to 80% estimated savings does and doesn't mean, and where the routing table is still wrong. The benchmark methodology has every number.

newsletter03

Stay in the loop

New posts on LLM cost optimization, benchmarks, and Flux updates. No spam.