Per-token price now shapes architecture more than capability: the case for tiered model routing
JeremyCMorgan · x · 2026-09-02
A dbreunig essay argues the "free lunch" era is over: frontier models like Fable are strong but expensive, pushing agentic coders toward tiered routing instead of defaulting to the best model for everything.
- GLM 5.2 shipped the same week as Fable at roughly 1/9th the cost (and 1/5th of Opus 5); for most rote coding with good context, it's more than sufficient.
- The author's workflow: interrogate and shape designs with Fable, then hand off a brief to GLM.
- The analogy is Moore's Law slowing down: cheap model improvements used to paper over crude harnesses and context strategies; now you must think about what work goes where.
- He's skeptical falling prices will bring back send-everything-to-the-big-model routing, since cheaper inference benefits mid-tier models too, and better harnesses make weaker models viable.
Takeaway: per-token price has become a bigger architectural constraint than raw capability — a useful frame for anyone designing multi-model systems.
More from coding & agent
- Unitpost launches on Product Hunt: one email platform AI agents can send through via MCP — thisiskp_ · 2026-09-02
- Google ships agentic video understanding with Gemini, explainer video released — patloeber · 2026-09-02
- Post-training, custom spec decoding and vLLM tuning: a hands-on inference cost-saving playbook — dhruv2038 · 2026-09-02
- Simon Willison vibe-codes a GeoJSON map tool, with ChatGPT Work pulling government boundary data — MaxLenormand · 2026-09-02
- Dev on the Codex hype: less unreliable than claimed, with marketing on both sides — D3VAUX · 2026-09-02
- Developer runs a 'reality-check' Claude Code skill on every new frontier model to audit his half-finished projects — doodlestein · 2026-09-02