TL;DR
Matthew Berman explains model routing — using expensive frontier AI models only for planning and spec-writing, then delegating actual code execution to cheaper models. This simple workflow change can cut AI costs by 60-70% with minimal quality loss.
Key points
1
Model routing means using the best model for planning/research and a cheaper model for execution/coding — the core insight is that a well-written spec enables cheaper models to write code just as effectively.
2
The math is concrete: using a frontier model like Claude Opus for everything costs ~$9.50 per feature vs. ~$3.20 when offloading coding to a cheaper model — a ~68% savings.
3
Output tokens are 5x more expensive than input tokens on frontier models, so code-writing (high output) benefits most from being routed to cheaper models.
4
Third-party tools like Cursor have built-in auto model routing; Cursor auto-routes simpler sub-tasks to Composer 2.5 even when you select a frontier model — unlike first-party tools (Codex, Claude Code) which have no incentive to do this.
5
Coinbase reduced AI spend while increasing usage by routing tasks to cheaper models including open-source ones like GLM 5.2, combined with better caching and context management.
Actionable insights
→
Use a frontier model (e.g. Claude Opus) only for research and writing a detailed spec document, then paste that spec into a cheaper model (e.g. GPT-5.5 via Codex) for actual code writing.
→
After the cheaper model creates a pull request, optionally pass it back to the frontier model for a review pass — only copy feedback back to the cheap model if there are issues.
→
Inside Claude or any platform, actively choose the right model and thinking level for each task: use Haiku 4.5 for simple/fast tasks, reserve Opus for complex planning, and avoid defaulting to max thinking effort on routine jobs.
→
In Cursor, enable auto mode to let the tool handle model routing automatically — it will downgrade simpler sub-agent tasks to Composer 2.5 even when a frontier model is selected.
→
At enterprise scale, consider dedicated model routing services like Not Diamond, and complement routing with caching and context management to flatten costs while usage grows.
Notable quotes
“The frontier labs like OpenAI with Codex, like Anthropic with Claude Code, they do not have model routing. They are not really incentivized to do model routing. Why would they? They want you to use the top tier model.”
“Output is five times more expensive than the input. And now maybe you are starting to see why this strategy pays off.”
Worth watching?
⏭️
Worth watching the full video?
The core strategy and all the math are fully captured here — skip the video unless you want to see the live Cursor/Codex screen demos of the workflow in action.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free