A router that picks the cheapest capable model per request — and the engineers explaining why it costs more
Workweave's Show HN promises 40–70% savings by routing each request to the best model in under 50ms. The thread's most credentialed replies say the math fights itself: every model switch throws away the prompt cache that makes agentic coding affordable.
The tool is a drop-in proxy across Anthropic, OpenAI, Gemini and open models, scoring each request and routing accordingly, keys encrypted, fully self-hostable. The claim is 40–70% lower cost “with just an endpoint change.”
The objection is structural: agentic sessions live on prompt caching with a five-minute TTL, and the more distinct models you touch, the more cache misses you eat — so a one-planner, one-executor setup already captures the win without the churn.
The pitch: one endpoint, 40–70% off
Workweave's router is a drop-in proxy that scores each incoming request with an on-device embedder and routes it to the best model across Anthropic, OpenAI, Gemini and open models via OpenRouter, advertising “Cut costs 40–70% with just an endpoint change,” in “less than 50ms” per decision, with keys encrypted and a self-host path. It plugs into Claude Code, Codex, Cursor and opencode.
The cache problem the credentialed replies kept hitting
The recurring expert objection wasn't about routing quality; it was about what routing destroys. An ex-Meta engineer who worked on HHVM laid out the contradiction:
“Caching and optimum routing of a single request are at odds with each other. Higher the distinct model count in a conversation, more cache misses you accept … The two model primitive, one planner and one executor, is already sufficient.”
A reader who builds on these APIs made the same point from the agentic-coding angle: “These are long-chained sessions of tool use that heavily rely on prompt caching. Changing mid-flight is costly.” The implication is uncomfortable for the headline number: in exactly the workflow the tool targets — Claude Code, Cursor — the savings from a cheaper model can be eaten by the cache miss the switch causes.
The other two questions it didn't answer
Beyond caching, a CISO-turned-founder doubted small models can even finish the job: they're “prone to stop before completion, throw errors and produce loops,” which routing has to account for and the README doesn't. And one reader asked for the one thing that would settle it: “Have you done any A/B tests on this with evidence?” for the “no noticeable difference in quality” claim. The idea is clearly attractive — several commenters admitted they're building the same thing — but the thread's verdict is that smart routing only pays off once it's smart enough to respect the cache it's spending.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
The Economist says the AI backlash is just starting — the thread says it's about size, not silicon
The leader warns politicians the revolt is only beginning and offers four pointers to manage it. Readers noticed the piece is quietly pro-AI, that the anger is bipartisan, and that the real target may not be the technology but the handful of companies that now own everyone's attention.