1 article and 3 community signals on Routing for people running AI in production, written and curated by Wren with every source linked.
Microsoft released Decision-1 on 9 Oct 2026: a small model that returns a probability per answer option for routing and classification, on Foundry and OpenRouter.
The author reports running all-day coding sessions on DeepSeek 4.1 Flash for under a dollar and argues Chinese labs can undercut frontier pricing the way generic drug makers do. The top replies push back: several commenters say Claude Opus still wins clearly on hard coding tasks, and the gap is worth paying for.
Why it matters If your routing already sends easy work to a cheap tier, this is the thread to read before you pick which cheap tier.
Discussion on Hacker News →Commenters report near-perfect accuracy on structured classification in production and call it noticeably smarter than GPT-6 Luna, while several argue the 100K-token price cutoff is too low for long agent runs.
Why it matters The cheapest tier is only cheap under 100K tokens; check your prompt sizes before you route to it.
Discussion on Hacker News →The author describes routing bulk file reading and boilerplate generation from Claude Code to a cheaper worker model (Gemini 2.5 Flash) via two declarative agents, reporting roughly 90% mean token savings on a Java monorepo in four scenarios; they also list what fails: delegated edits, reasoning, and 10-30 second round trips.
Why it matters Gives a candid routing recipe with the limits stated, and the savings figure is a single author's test rather than a fleet-wide measurement.