Not Diamond Learns About You. Alone.
Every Developer Gets Their Own Router. Nobody Else Gets Smarter.
Most coding-agent spend goes to a lazy default: every step goes to the most expensive model. Intelligent model routing replaces that default with a choice made at each step, and it's one of the best cost-and-quality levers teams running agents have. Once you accept that routing is worth doing, the question that matters is where the router's intelligence comes from.
This piece compares two routers that answer that differently: Not Diamond Code, a commercial product, and the Router running on the Hokusai protocol. It covers routing only. The goal is a clear breakdown you can decide from.
Why route at all
A coding task is made of dozens of steps, and only a few of them need a frontier model thinking hard. Everything else runs fine on something cheaper and usually faster. Picking the right model for each step by hand doesn't scale, and static rules go stale every time a lab ships a new model. A learned router makes the call for you. Not Diamond has published large cost reductions at Opus-level quality. We haven't verified those numbers and aren't disputing them. The direction is what everyone in the space sees: most steps get more model than they need.
How Not Diamond Code works
Not Diamond Code runs as a local, privacy-preserving proxy next to your harness (Claude Code included) and picks the model and reasoning effort for each step. According to the company, only anonymized metadata leaves your machine. It's sold to startup and Fortune 500 engineering teams and carries SOC 2 and ISO 27001.
It learns per developer. It infers whether a routing choice worked from what you did next (kept the change, re-prompted, reverted) and adapts to "each developer's workloads and preferences." You don't have to do anything beyond using it normally, and it tunes itself to your habits.
How the Hokusai Router works
The Hokusai Router sits beside your harness as an advisor, outside the request path. Your system asks which model to call and then makes the call itself. The router learns from outcomes that people and systems explicitly report, and those outcomes are pooled across every organization using it into one shared model.
Hokusai is a protocol, not a vendor. The SDK is open, attribution is recorded on-chain, and contributors whose data measurably improves the model (measured in DeltaOnes) earn a stake in it. What the model learns lives in the shared model, not in any one team's install, so it moves with you when you switch models or harnesses. Swap the model, keep the expertise.
Better data, better decisions
A router can only be as good as the evidence behind its choices. So look at what each one learns from.
- Volume. The most expensive routing mistakes are rare for any one team. How many of them does the router ever get to see?
- Coverage. A new model ships. Is the router learning about it from your trial and error alone, or from everyone trying it at once?
- Starting point. On day one, does a new install know what earlier users already found out?
- Clarity. Keeping a change might mean it worked, or that you ran out of time. Is the signal inferred from behavior, or reported and auditable?
Tradeoffs are covered below.
Side by side
| Not Diamond Code | Hokusai Router | |
|---|---|---|
| What it is | Commercial product | Router on an open protocol |
| Where it runs | Local proxy next to your harness | Advisor beside your harness, outside the request path |
| Learns from | Implicit signals from each developer | Explicitly reported outcomes, pooled across organizations |
| Improves for | You | Everyone using it, including you |
| New install starts from | Vendor baseline, then personalizes | The shared model as it stands |
| Portability of what it learned | Stays with your install | Carries across models and harnesses |
| Feedback setup | None | Requires reporting outcomes |
| Ownership | Vendor-owned | Contributors earn a stake via DeltaOnes |
| Published cost benchmarks | Yes (vendor-reported) | Not making cost claims at this time |
Which to choose
Not Diamond Code fits if you want routing tuned to individual developers, zero feedback setup, a vendor with enterprise compliance certifications, or published benchmarks today.
The Hokusai Router fits if you want routing informed by more outcomes than your team will ever generate, learning that survives a model or harness switch, feedback you can audit, or a share of the intelligence your usage creates.
The difference is structural, not about who executes better. A product built to sell private, per-tenant optimization can't also pool everyone's signal into shared intelligence owned by its contributors without undercutting what it sells. A protocol can. We think the decision about which model to call belongs in shared infrastructure, not inside someone's margin.
Both are reasonable bets. Pick the one whose data you'd rather trust.