Comparison · Not Diamond

rightmodeler vs Not Diamond

Both refuse to be your gateway, from opposite ends. Not Diamond makes a live prediction and hands the call back to your provider; rightmodeler never sees a live request: it replays your accepted traces (the logged record of what your agent already did) offline, validates a cheaper or newer pinned model per step, and hands your team a draft pull request to review.

Different clock · prediction vs measurementVisit Not Diamond  (opens in a new tab)

TL;DR

Different jobs on different clocks. Not Diamond is hired to route: a trained meta-model predicts, live, which model (and for coding agents, which reasoning effort) should take the next request or session step, then hands the call back to your own provider or gateway. rightmodeler is hired to validate: the free CLI replays your accepted agent traces through cheaper or newer candidates offline, measures how well each one agrees with the output you shipped, and opens a draft pull request for each swap that clears every release gate, for your team to review and merge. Both refuse the gateway seat, from opposite ends: Not Diamond predicts and hands back; rightmodeler never sees a live request at all. A router picks a model per request at runtime; rightmodeler validates one pinned model per step, offline, before a person merges.

Same refusal, opposite ends

Neither product wants to be your gateway. Not Diamond hands the call back after a live prediction; rightmodeler never sees the call at all. Four dimensions keep them apart.

not-diamond vs rightmodeler
clock
Not Diamond · request time: score the input as it arrives
rightmodeler · release time: replay traces before anything ships
evidence
Not Diamond · a prediction from a trained meta-model
rightmodeler · a measurement against the output you accepted
seat at runtime
Not Diamond · a local proxy beside your harness; the call still runs through your gateway
rightmodeler · none; it runs offline, never in the request path
who approves
Not Diamond · the router's pick applies live, mid-session
rightmodeler · your team; every swap arrives as a draft pull request a person merges

Pick by the job

Three real situations and the honest call in each.

An engineering org runs Claude Code all day and wants each step of a live session on the right model and reasoning effort, including staying on a pricier model when the warm cache makes that cheaper overall.

the right hire: Not Diamond

This is exactly what Not Diamond Code was built for: a local proxy beside your harness makes cache-aware routing calls mid-session, in real time. It is early access behind a gated waitlist as of August 2026, but rightmodeler works offline and cannot act inside a running session at all.

A regulated team wants the savings but cannot let any third-party service influence production requests or see request metadata, and payment and auth steps must not move.

the right hire: rightmodeler

The CLI runs offline on your own traces with your own keys and adds nothing to your request path; there is no rightmodeler server, no account with us, no telemetry. Where the evidence is thin it abstains rather than proposing a swap. It has no setting that pins a step, but nothing moves until a person merges a draft pull request that changes only model identifiers, so a payment or auth step changes only if your reviewers accept that change. Even Not Diamond's proxy, which per their docs sends only derived metadata, still talks to their service at runtime.

A team is evaluating Not Diamond and wants to know whether a router would actually save money on their pipeline before wiring it in.

the right hire: both, together

Run a validation first, on your own traces: for each step it shows whether a cheaper pinned model clears every release gate, or abstains where the evidence is thin. That tells you what a fixed per-step assignment can save before you weigh a router against it. What per-request routing would add on top is for Not Diamond's own evaluation on your traffic to show; rightmodeler does not measure it.

The honest trade

Admissions first, then the reason we made the opposite bet.

  • Runtime adaptivity is real. Not Diamond routes each individual input live, so on high-volume, heterogeneous traffic it can capture savings a fixed per-step assignment leaves behind. Not Diamond Code goes further: mid-session, it adjusts model and reasoning effort step by step and factors in KV-cache state, sometimes staying on a pricier model because the warm cache makes that cheaper overall. rightmodeler never touches a live request and cannot do any of that.
  • There is also no measurement lift on their side. The router works out of the box, with no traces to collect and no replays to run, and it has operated at real scale: Not Diamond's meta-model has powered OpenRouter's Auto router since January 2025, and OpenRouter's CEO still credits it on Not Diamond's homepage as of August 2026. Their launch post claims 20%+ cost reduction without quality degradation for Not Diamond Code; that is Not Diamond's own number, so weigh it as you would any figure, ours included: against your own traffic.
  • Here is why we still built offline validation instead of a router. Researchers from Michigan, Stanford, MIT, Google DeepMind, and All Hands AI, the team behind the OpenHands agent, ran eight frontier models across 500 real GitHub issues, four separate times each, and recorded every token (Bai et al., arXiv:2604.22750). Across those repeats, token usage on the identical task, agent, and model varied by as much as thirty times. A single live prediction, like a single replay, is one draw from that wide distribution. So the report shows reference agreement (how often a candidate agreed with the output you shipped), sample size, and abstentions, and it abstains (declines to recommend) where the evidence is thin, rather than stretching one draw into a verdict. The verdicts stay hedged accordingly: agreement with what you shipped, never ground-truth correctness, because the production result is the reference.

Stay on Not Diamond when

  • Your traffic is high-volume and genuinely heterogeneous, and the best model varies input to input; a fixed assignment leaves that variance on the table.
  • You want live, cache-aware picks of model and reasoning effort inside a running coding session.
  • You want savings with zero measurement lift: no traces to collect, no replays to run.

Bring in rightmodeler when

  • You want evidence against a quality floor (the minimum agreement bar you set) that a reviewer can approve before any model change ships.
  • Steps whose evidence is thin must produce an abstention, never a swap.
  • You want one pinned model per step, validated on your own traces and changed only through a draft pull request a person merges.

Frequently asked questions

Does rightmodeler replace Not Diamond?

No, and the reverse is also false. Not Diamond predicts, live, which model should take the next request or session step. rightmodeler validates, offline, whether a cheaper or newer pinned model can take each step of your stack, and a person merges every change. They answer different questions on different clocks.

Does rightmodeler integrate with Not Diamond?

No integration ships today: there is no Not Diamond trace adapter and no Not Diamond replay provider. The honest relationship is sequencing, not plumbing. The CLI reads the traces your harness or gateway already logs, in a format it supports, so a stack that routes through Not Diamond still produces traces rightmodeler can read.

Is Not Diamond Code generally available?

No. As of August 2026 it is in early access behind a gated waitlist, per their docs. Its launch post claims 20%+ cost reduction without quality degradation, plus benchmark wins on Poly-SWE-Bench and LongCodeQA; those are Not Diamond's own numbers.

Can rightmodeler tell me whether a router would pay off on my pipeline?

Not directly. The report shows, per step, whether a cheaper model's agreement with the output you shipped reaches your quality floor at a statistical lower bound, with sample size attached, and it abstains where the evidence is thin. That tells you what a fixed per-step assignment can save on your traces. It does not measure what per-request routing would add; that comparison needs Not Diamond's own evaluation on your traffic.

Does Not Diamond see my prompts and code?

Per their docs, no: the Not Diamond Code proxy computes derived metadata about the session, sends only that to their optimization service, and the model call itself runs through your configured provider or gateway. rightmodeler's answer to the same worry is structural: there is no rightmodeler server, no account with us, no telemetry; the CLI runs on your machine against your own traces, and its replays go only to the provider you configure, under your own key.

Start with the traces you already have.

The CLI runs from npx, nothing to install, and validates cheaper or newer models on your own traces.

View on GitHub