Blog

A leaderboard can't choose your model for you. The method that actually works: define the job, test two or three candidates on your own traffic, score cost per finished answer, latency and quality, then decide whether to pin or route.

Get schema-valid JSON out of Flux with one field. json_object vs json_schema, runnable Python and curl, why support rides on the served model, how to pin one, how to read which model answered, and the one validation line you still owe yourself.

Asking a model nicely for JSON is a probabilistic hope. A schema-enforced output mode moves the guarantee from the prompt into the decoder. Here is why the difference shows up as quiet, expensive failures your try/except never catches, the honest limits across models, and the decision order to ship it.

What an LLM gateway actually is, how it differs from a router and a raw provider SDK, when you genuinely don't need one, and how to trust any gateway by what it shows you about one live request.

Per-token price is a decoy. The number that runs your AI bill is dollars per completed answer. Here's how to measure it, control it, and govern it, with a real cost receipt on every request.

A router company on when to build your own LLM gateway instead of buying one. The real test: is routing your product or your plumbing? Plus a copy-paste decision checklist.

A spreadsheet only works while someone is watching it. Here's the lightweight operating model that governs AI spend on its own: a standing ceiling, a cost receipt on every request, and one invoice.

A leaderboard scores someone else's task. Build a 20 to 50 example held-out set from your own inputs, pick a grader that matches the work, and score candidates on cost per correct answer. Runnable harness included.

Stop judging routers by the pricing page. Point a trial key at one and read what it shows you about a real request. Seven criteria, each with a five-minute test, and a scorecard you fill in yourself.

Point your OpenAI transcription calls at FluxRouter, use flux-voice, done. Same key, same base URL, one new model id. Python + curl, request fields, response formats, costs and limits. Runs on any paid plan.