What Not Diamond is
Not Diamond predicts which model should handle each request; execution stays on infrastructure you already run. Its June 2026 explainer on routers versus gateways places the product between your agent and your gateway. Not Diamond is San Francisco based, roughly 15 people, on a $2.3M seed raised in July 2024 with angels including Jeff Dean and Tom Preston-Werner. Fortune's August 9, 2026 coverage of the router market names SAP among its enterprise clients.
There are two product generations. The 2024-2025 model-selection API returns a recommendation from a select_model call and your client executes the request itself, with optional fuzzy hashing so the recommendation engine never receives raw query text. Since August 4, 2026 the flagship is Not Diamond Code: a local proxy beside your coding harness that sends derived request metadata to Not Diamond's optimization service, gets back a model and reasoning-effort choice, and executes through your existing gateway or provider. The pricing page lists $0.05 per million tokens routed, a stated 100-150ms of added latency per recommendation, and claimed inference savings of at least 20-40%. Not Diamond does not resell execution; inference is billed by your providers.
Where it is the right choice
Learned per-prompt selection is Not Diamond's specialty, productized end to end. You upload a CSV of prompts, candidate-model responses, and evaluation scores (15 to 10,000 samples) and get a custom router back within about an hour. Their RoRF router is open source under MIT. For coding agents, Not Diamond Code routes each step of a session, choosing both the model and the reasoning effort from request complexity and conversation history.
The architecture also keeps payloads out of Not Diamond's hands. With privacy-preserving routing enabled, Not Diamond Code decides from locally computed features and derived metadata; the launch post says routing works "without ever requiring agent payloads, inputs, or outputs, to leave a developer's local machine." The legacy API went further by never executing calls at all. And because Code rides on your existing gateway and provider contracts, the execution side has no lock-in: you keep negotiated rates and can drop the router without re-plumbing inference.
The load-bearing difference: selection versus execution
TrustedRouter is the layer Not Diamond assumes you already have. We execute the request: one OpenAI-compatible API over about 550 model routes across roughly 49 providers as of August 16, 2026 (see /models), with provider failover, prepaid billing, and BYOK. Not Diamond's docs list about 80 candidate models across 13 providers as of May 2026, and you hold the credentials for each provider it might select.
On the selection axis, Not Diamond's routing is the more sophisticated. TrustedRouter routes are policy pools with composable preferences: trustedrouter/auto, free, cheap, fast, zdr, e2e, and eu, plus provider ordering, privacy floors, and jurisdiction pins. Nothing in our stack trains a per-prompt predictor on your evals. What we add instead is enforcement: trustedrouter/zdr refuses providers below a contractual zero-data-retention tier, and trustedrouter/e2e routes only to confidential-compute providers with provider-side end-to-end encryption (currently tinfoil and phala).
The fees sit on different bases and cannot be equated. Not Diamond charges a flat $0.05 per million tokens routed, on top of whatever your gateway and providers bill for inference. TrustedRouter's prepaid fee is 5.5% of the provider's token price with a $0.01 per million token floor, and that is the whole bill: at a $3.00/M provider price you pay $3.165/M, execution included. One is a selection fee, the other an execution markup; per-model numbers are on /pricing.
Privacy: two models, both real
Not Diamond's strongest privacy argument is structural: in privacy-preserving mode your payload never reaches them. The caveats live in their privacy policy: hosted-service input and output content is retained about 90 days by default, API logs up to 18 months, and for hosted model services opt-out from payload processing is not available. We could not establish whether privacy-preserving routing is on by default.
TrustedRouter sees your prompt, because we execute the call, so we prove what happens to it instead. Realtime inference keeps no durable prompt or output logs; what we retain is metadata: request ids, model and provider, token counts, latency, cost, region, and API-key hash (full list on /privacy). The gateway build serving api.trustedrouter.com is attested on GCP Confidential Space, AWS Nitro Enclaves, and Azure Confidential Containers; anyone can run the verifier against trust.trustedrouter.com, and /security documents the model. The attestation boundary stops at our gateway: downstream model providers are covered by cited policy tiers, except the E2E routes, where the provider's own confidential-compute mechanisms apply.
What we do not claim
Our repos are young: public since late April 2026, with benchmark history from June 2026 — a short operating record. The 99.99% figure on /status is a published availability target with live burn rates, not a certified SLA. We publish a DPA, BAA, and subprocessor list, but no SOC 2 or HIPAA certification today; Not Diamond claims SOC 2 compliance and ISO 27001 certification in its August 2026 launch post, self-reported, with no public trust portal we could find. And our gateway and control plane are source-available under BUSL-1.1, converting to Apache-2.0 four years after each release. Every line that touches your prompt is public, but it is not OSI open source; the SDKs are Apache-2.0 or MIT.
Migration, or coexistence
Not Diamond routes through a gateway by design, so the first question is whether you are replacing it or pairing it. Not Diamond describes Code as gateway-agnostic, so TrustedRouter can serve as the gateway underneath it — though neither company publishes a tested integration. If you consolidate on TrustedRouter alone: legacy API users delete the select_model call and the client-side execution branching, then point one OpenAI-compatible client at our base URL; Not Diamond Code users remove the local proxy from the harness configuration. What does not carry over is the learned selection itself. A custom router trained on your scored evals has no TrustedRouter equivalent, and our policy routes will choose differently, so re-run the same task-specific evals before moving production traffic; see /docs/evals.