Not Diamond Review (2026)

It recommends, it does not proxy

This is the architectural decision that makes Not Diamond worth a page.

Integrations are stack-agnostic through its API, and recommendations are executed in your own model gateway and harness. Not Diamond tells you which model to call. You call it.

Every other router in this category is a proxy. Adopting one means:

• A new component in your request path that can fail
• Additional latency on every call, whether or not routing helps
• An availability dependency independent of your model provider

Not Diamond has none of those. If it is unavailable, you fall back to your default model and carry on.

The practical consequence is that it composes with LiteLLM, Vercel AI Gateway, Cloudflare or your own proxy rather than replacing any of them. Trying it does not displace existing infrastructure, which makes it unusually low-commitment for something in the gateway category.

The trade is one extra round trip for the routing decision before your model call. Measure it - though routing to a smaller, faster model frequently saves more latency than the decision costs.

Custom routers are the substantive part

Three products: a pre-trained Chat router, a Code router aimed at coding-agent cost (early access), and custom routers trained on your own data.

The custom option matters most, because it addresses the weakness generic routing always has: your workload is not the benchmark.

A pre-trained router encodes which model tends to win across some distribution of tasks. Your traffic almost certainly does not match that distribution - different prompts, different domain, different quality bar. Training on your own data turns a heuristic into a measurement, and means the value grows with usage rather than being fixed at adoption.

The underlying approach combines models into a meta-model that learns when to use each LLM.

Why coding agents are the right target

The Code router is well-aimed even though it is still early access, and the reason is worth spelling out.

Coding agents are where per-request model spend is least controlled. They make many calls per task, frequently to the most expensive model available, for sub-tasks whose difficulty varies enormously. Reading a file, renaming a variable and designing an architecture do not need the same model - and the default is to send all three to a frontier model.

A router that recognises the easy calls is attacking the largest and most obviously wasteful slice of agentic spend. That is a better-chosen problem than general chat routing, where the savings are thinner and the quality risk is higher.

The claims, and how to read them

Not Diamond markets cutting inference costs 20 to 40 percent for coding agents, and Rootly reports a 39 percent average accuracy increase across its SRE benchmarks.

The Rootly figure is a customer report, which is stronger evidence than a vendor claim - but it is still one workload on one benchmark set, and neither number has been independently reproduced.

The logic underneath is sound: routing simple requests to cheaper models saves money, routing hard requests to stronger models improves accuracy. But the magnitude depends entirely on your traffic mix. A workload where every request is genuinely hard will save nothing, because there is nothing to downgrade.

Treat these as directional. The company is SOC 2 and ISO 27001 compliant.

Pricing we could not resolve

Our sources contradict each other outright:

• One directory lists $0.05 per million tokens routed, plus enterprise with volume discounts, SSO/SAML and SLAs
• A competitor’s comparison states no per-request or per-token price is published, listing only free Early Access and custom Enterprise

Those cannot both be true. We are recording it as unresolved and would check the vendor’s own pricing page.

This is the fourth pricing figure in this category we have been unable to reconcile from third-party sources, after LiteLLM’s enterprise tier, Confident AI’s model change, and PromptLayer’s tiers. For infrastructure tooling, treat any pricing table you find as a starting point for a question rather than an answer.

Should you use it?

Use Not Diamond if you already run a gateway and want smarter per-request model selection without adding a request-path dependency - especially for coding agents.

Don’t use it if you need the Code router generally available today, you cannot absorb an extra round trip, or your traffic is uniformly hard and there is nothing to route away from.

Bottom line: the only router here that does not ask to sit in your request path, which makes it cheap to try and safe to keep. The custom-router story is the real product; the pricing is the part you will have to ask about.

---

Architecture, product lineup, compliance and reported results verified against vendor documentation and third-party sources on 3 August 2026. Pricing sources conflict directly and the figure is reported as unresolved. Cost and accuracy improvements are vendor- or customer-reported and have not been independently verified. This is a researched directory entry - we have not yet instrumented this platform with our reference application.