Blog

Posts on how we build GammaInfra — smart routing design, the bug stories from running it in production, and the methodology behind shipping a solo-founder LLM routing service.

2026-05-14
Every agent step deserves a different model
Agent loops compound every weakness of using one model for everything. A field guide to per-step model variance, tail-latency budgets, fallback chains, cost-runaway observability, and the cross-provider tool_call.id papercut — with code samples for Claude Agent SDK, LangGraph, and OpenAI Agents SDK.
2026-05-09
Why every GammaInfra response carries a cost-USD header
LLM API providers report tokens. Developers care about dollars. How we compute X-GammaInfra-Cost-USD per request, why it's harder than it sounds (per-direction split, long-context surcharges, fallback cascades), and how to read it from common SDKs.
2026-04-23
Designing a continuous cost/quality dial for LLM routing
Most LLM routers force a discrete tier — cheap, balanced, quality. We added a continuous 0.0..1.0 dial via one request header. Why continuous beats discrete, what we tried and threw away, how it maps to actual model picks.

Frequently asked questions

What does the GammaInfra blog cover?
Engineering write-ups on smart LLM routing: per-step model selection in agent loops, the continuous cost/quality dial design, why every response carries a cost-USD header, and the trade-offs behind the routing decisions. It is the deeper-dive layer above the API docs and glossary.
Is GammaInfra's routing approach documented elsewhere?
Yes — the API reference is at docs.gammainfra.com and conceptual definitions are in the glossary. The blog is where design rationale and measured results live; the docs are the reference contract.
How do I try the routing behavior described in these posts?
Point any OpenAI-compatible SDK at https://api.gammainfra.com/v1 with a GammaInfra key and use model gammainfra/auto. Every behavior the posts describe — per-step routing, the cost-quality dial, the cost-USD header, fallback chains — is live on that endpoint with no extra configuration.
How often is the blog updated?
Posts ship alongside notable routing or observability work rather than on a fixed calendar. Each post is dated and the engineering claims are reproducible against the live API at publication time.