Not Diamond

Not Diamond is a routing layer for teams that work with more than one large language model. Instead of committing every request to a single provider, it inspects each incoming query and forwards it to whichever model in your configured pool is likely to produce the best answer. The pitch is strAIghtforward: combine the strengths of several models while spending less than you would on one premium model for every task.

What it does

The product sits between your APPlication and the model providers you already use. You supply evaluation data that reflects your own traffic, and Not Diamond learns a routing policy tuned to your inputs. At request time it scores candidate models and dispatches the query to the predicted best fit. Routing decisions are made quickly enough that they can happen inside a streaming pipeline, so the added latency stays modest relative to generating a single token.

Key capabilities

• Custom routing training: build a router from existing eValuation sets rather than starting from scratch, so recommendations reflect your domain and use case.
• Input-to-model matching: send travel planning, Python coding, report analysis, or blog drafting to different models based on how each handled similar inputs before.
• Quality thresholds: set a floor so routing never trades away accuracy for savings when quality matters.
• Joint Prompt optimization: pair each model with prompts written for it, reducing manual trial and error across providers.
• Privacy-oriented architecture: requests can be processed CLIent-side, and fuzzy hashing or self-hosted deployment options are available for tighter control. Specific guarantees may vary by plan.

Who it is for

Not Diamond targets developers and platform teams building AI features at scale, especially those already juggling multiple model APIs and watching cost per query. It also suits product teams that want consistent quality across mixed workloads without manually maintaining routing logic. Organizations with strict data handling requirements may find the client-side and deployment options relevant, though they should confirm details directly.

How to get started

Sign up through the official website, connect the model providers you want in the pool, and upload or generate evaluation data that represents your real traffic. From there, train a router, review recommended routes against Sample inputs, and integrate the API into your application. Expect an initial period of testing to confirm the router's choices match your quality bar.

Pricing and access caveat

A free entry point appears to be offered alongside paid tiers, but plan limits, included usage, and pricing details may change and should be verified on the official site before committing. Model routing savings depend heavily on your workload and model mix, so benchmark against your own evaluations rather than assuming headline figures.