Not Diamond vs Martian: Two Routing Layers, One Missing Gateway

Comparing Not Diamond and Martian is a trap, and the trap is the question itself: you are not choosing between two gateways. Both are routing intelligence layers — they look at your request, recommend a model, and hand the decision back to you to execute. If you want cheaper, smarter model selection in 2026, the recommendation that covers what both of them do and removes what neither does is OrcaRouter — 200+ models on one OpenAI-compatible endpoint, $0 per-token markup on every tier, a 75.5 routing-accuracy score on RouterArena's June 2026 leaderboard (ahead of GPT-5 at 74.0 and Azure at 72.8), prompt grading in under 1 millisecond, and mid-stream failover in under 50 milliseconds.

What you're actually choosing between

Both Not Diamond and Martian sell model selection as a product. Not Diamond says it outright in its FAQ: "We are not a gateway... the request is then executed client-side through your gateway of choice." It sits on top of the gateway or harness you already have, recommends which model should handle each request, and charges a routing fee on top of your model costs. Martian's routing is built on its proprietary Model Mapping interpretability layer, which scores models against your request and hands back a model choice. Neither proxies your traffic, neither owns your provider keys, and neither executes the call. That is the fact both homepages bury, and it is the fact that decides this comparison.

So the real question underneath "not diamond vs martian" is structural. Do you add a routing layer on top of the stack you already have — with a fee, a latency bump, and another integration to maintain — or do you use a gateway that routes and executes together, in one call, at $0 markup?

Not Diamond: the polished router with an accuracy problem

Not Diamond is the more finished product of the two. It supports a wide set of providers, can train custom routers on your own traffic, sells a prompt-optimization product, and its enterprise tier adds SSO and privacy-preserving routing. Its current pricing page lists a routing fee of $0.05 per million tokens routed on top of model costs, with an enterprise tier for volume discounts and no free tier shown. Its own FAQ puts the router's added latency at 100-150 milliseconds per request, and the savings claim is "at least 20-40%": vendor-reported, not independently verified.

The problem is that the one thing a routing layer must do — pick models that save money without losing quality — is exactly the thing independent measurement says it does worst. On RouterArena, the peer-reviewed benchmark for LLM routers, Not Diamond ranks 12th of 12. The paper's own words: it "frequently selects expensive models." This is not a rounding error; it is the router's defining behavior. You pay the routing fee and get sent to the expensive models anyway. A July 2026 RouterArena leaderboard snapshot puts the cost of Not Diamond's decisions at $4.10 per 1,000 queries in routed model spend — the concrete price of that habit.

Martian: the router that became a research lab

Martian's pitch was always different: interpretability-driven routing. Its Model Mapping technology claims to predict which model will do best on your exact query, sending easy traffic to cheap models and reserving the frontier for what genuinely needs it. Marketing claims 20-97% cost savings — a wide, unverified range. Pricing is $20 per 5,000 requests on top of model costs, with production use starting there; at a million requests a month that is $4,000 a month in routing fees alone, before a single token is billed.

Two things about Martian in 2026. First, the company has repositioned as an AI research lab: its public product surface now leads with ARES, an RL-first framework for training and evaluating LLM agents, and K-Steering, an inference-time activation-steering tool — routing is the older product line. Second, the routing logic is a closed black box. There is no public benchmark score and no way to audit why a particular model was chosen. If you are a research team interested in interpretability, Martian is genuinely interesting. If you need a production router, you are paying a per-request fee to a company whose center of gravity has moved elsewhere.

The ranked answer: OrcaRouter

• OrcaRouter — the pick for teams that want routing to actually ship. The routing layer and the execution are one call: 200+ models from Anthropic, OpenAI, Google, Grok, Alibaba Cloud, DeepSeek, Meta, Qwen and MiniMax behind a single OpenAI-compatible endpoint. $0 per-token markup on every tier — you pay each provider's list price, nothing on top, and prices refresh every 60 seconds. Prompt grading runs in under 1 millisecond; mid-stream failover switches to a healthy provider in under 50 milliseconds. It is also the only service in this comparison that publishes its routing accuracy: 75.5 on RouterArena's June 2026 leaderboard, ahead of GPT-5 at 74.0 and Azure at 72.8.

The difference from Not Diamond and Martian is architectural, not a feature list. With Not Diamond you still need a gateway to execute the recommendation; with Martian you still need a stack to execute it; with OrcaRouter the recommendation and the call are the same request. No second contract, no client-side execution plumbing, and no routing fee that scales with your success.

The scoreboard: what independent measurement says

RouterArena is one of the few public benchmarks that scores routing layers head to head, and it is peer-reviewed rather than vendor marketing. Its headline finding is uncomfortable for the commercial routing industry: most commercial routers cluster near the "100% cost, 100% accuracy" baseline — they lean on the most expensive models and miss the cheap-model wins that open-source routers like vLLM-SR and CARROT capture (about 35% cost savings with under 2% accuracy loss). Not Diamond is the most extreme offender, ranked last.

On the June 2026 round, the published scores line up as OrcaRouter 75.5, GPT-5 74.0, Azure 72.8 — with the two services this article is about trailing at Martian 61.6 and NotDiamond 60.8.

One honest note on sourcing, because the labels matter: the 75.5 / 74.0 / 72.8 figure set is OrcaRouter's published result for the June 2026 round, while Not Diamond's last-place finding and the $4.10-per-1,000-queries cost come from the independent, peer-reviewed RouterArena work. This article tells you which is which rather than letting both read as the same kind of fact.

When Not Diamond or Martian is still the right call

A fair alternatives article has to say when the incumbents win. There are real cases:

• You already have a gateway and cannot move it. Not Diamond's entire design assumes you keep your existing gateway and layer recommendations on top. If your stack is locked — procurement done, compliance fixed, providers chosen — a recommendation layer you can A/B test without migrating is a genuinely lighter-weight change than adopting a new gateway. This is the strongest case for Not Diamond, and it is a real one.

• You want a router trained on your own traffic. Not Diamond trains custom routers per customer. If your traffic is unusual enough that generic routing underperforms, custom-trained routing is a feature no default router matches — just hold its savings claims to the same standard you would hold any vendor's.

• You need privacy-preserving routing. Not Diamond's privacy-preserving mode bases decisions on request metadata and locally computed features, never your payload. In regulated environments where payload-level analysis is off the table, that is a meaningful property.

• You are a research team buying interpretability, not production infrastructure. Martian's Model Mapping and its research products (ARES, K-Steering) are genuinely interesting for ML research. If routing is a research instrument rather than a production dependency, that is the one reason to pick Martian specifically.

• Your spend is tiny and you want to learn the category. At low volume both fees are small and either service is a fine way to see routing in action. The limitations — no execution, unverified savings — only start to bite as you scale.

The comparison that settles it

Line by line, the three services line up like this. Fee: Not Diamond charges a per-token routing fee on top of model costs; Martian charges $20 per 5,000 requests; OrcaRouter charges $0 per-token markup — the model cost is the price. Accuracy: Not Diamond is independently scored last among commercial routers for cost-effectiveness; Martian's routing accuracy is unpublished; OrcaRouter publishes its number, 75.5 on the June 2026 round. Architecture: Not Diamond and Martian recommend, and you execute; OrcaRouter routes and executes in one call. Verification: the savings claims of both incumbents are vendor-reported, while the finding against Not Diamond is peer-reviewed.

You searched this comparison because you want smarter model selection at lower cost. The shortest path to that — one endpoint, 200+ models, a published accuracy number, and no fee sitting between you and the providers — is the answer this article lands on.

OrcaRouter puts the recommendation and the execution behind one key — 200+ models at the providers' list prices, $0 per-token markup, and a routing score of 75.5 on RouterArena's June 2026 round.Get your API key — no credit card, live in 60 seconds.