Not a gateway. Not a router. The measurement layer.
Model Maverick is a model recommendation engine: it benchmarks continuously, applies your policy, and returns a ranked answer your existing stack executes. Gateways move traffic; routers bundle the decision with inference; eval tools measure what you author yourself. Maverick sits upstream of all three — and works with them.
What each layer actually does.
| Dimension | Model Maverick | Gateways * | Routing APIs † | Eval tools ‡ |
|---|---|---|---|---|
| Executes inference | No — recommends only | Yes — proxies every call | Yes / bundled with routing | No |
| Continuous private benchmark | Yes — nightly, contamination-blocked | No | Partly — trained routers, opaque data | You author and run your own |
| Difficulty-aware ceilings | Yes — 5 authored tiers per capability | No | Rarely | No |
| Tenant policy as hard boundary | Yes — pools, constraints, validated writes | Partly — allowlists, key scoping | Limited | N/A |
| Decision audit trail | Yes — full trace per recommendation | Request logs, not decisions | Score, not reasoning | Eval runs only |
| Vendor lock-in | None — answers are just JSON | Your traffic flows through it | Routing and inference coupled | Low |
* Gateways: OpenRouter, LiteLLM, Portkey, Bedrock, Vertex AI
† Routing APIs: Not Diamond, Martian, Azure Model Router
‡ Eval tools: Braintrust, Langfuse, LM Evaluation Harness
Category examples as of early 2026. Every product above does more than one row can hold — this table compares layers, not feature checklists.
Use Maverick with your gateway, not instead of it
Gateways are excellent at what they do: unified APIs, key management, retries, caching, traffic control. What they don't have is an opinion — grounded in measurement — about which model deserves the call for this workload at this difficulty under your constraints. That's the part Maverick supplies:
- Your app asks Maverick:
POST /v1/recommend. - Maverick returns a ranked list — already filtered to your compliance pool — plus a check spec.
- Your gateway executes on the pick, validates output against the check spec, and escalates down the list on failure.
- Telemetry from the SDK or plugin reports what happened — automatically, closing the loop with nothing for your team to build.
Against trained routers: show your work
Routing APIs like Not Diamond and Martian train a router to predict the best model per prompt — genuinely useful, and validated by research like RouteLLM, where a trained router hit 95% of GPT-4 quality while sending about a quarter of calls to the strong model. The trade-offs are opacity and coupling: you can't inspect why the router chose what it chose, and the decision usually rides with their inference.
Maverick's bet is different: measurement you can audit. Difficulty ceilings from authored tiers, scores from a versioned corpus you can interrogate, a decision trace for every answer, and no inference coupling at all. If you need to explain a routing decision to a customer, an auditor, or your own postmortem — that difference is the product.
Against eval platforms: continuous and comparative by default
Braintrust, Langfuse, and the LM Evaluation Harness are the right tools for evaluating your prompts on your data. Maverick answers the adjacent question they leave open: across every frontier model tracked, tonight, which one should run this class of workload? You don't author the corpus, run the sweeps, or maintain the graders — you consume the ranking through an API, and bring your own evals for the last mile.
Keep your stack. Upgrade the decision.
Maverick slots in front of whatever executes your inference today.
Request early access