To choose an LLM router in 2026, first decide which layer you are missing. A gateway (LiteLLM, Portkey) owns the pipe: keys, budgets, failover, logs. A model marketplace (OpenRouter, Requesty, Ramp Router) owns access: many models behind one API and one bill. A decision layer (Not Diamond, Weave Router, Nadir) owns the choice: which model and how much reasoning effort each request gets. Most teams need one of each, and most "best router" lists mix the three as if they were rivals. They mostly aren't.
Competitor facts below come from each vendor's own docs, checked on 28 October 2026, and every one is footnoted on that vendor's [comparison page](/compare/). Prices and features change, so check the linked page before you buy. Nadir is one of the products listed, and we say where the others win.
The three layers.
| Layer | What it owns | Examples | You need it when |
|---|---|---|---|
| Gateway | Keys, budgets, failover, spend logs, guardrails | LiteLLM, Portkey, Bifrost | More than one team or app calls models |
| Marketplace | Model access and one bill across providers | OpenRouter, Requesty, Ramp Router | You want many models without many contracts |
| Decision layer | Which model and effort each request gets | Not Diamond, Weave Router, Nadir | One default model is overpaying for easy work |
A gateway or marketplace can include a basic router, and a decision layer can sit behind either one. The full split is in AI gateway vs model router.
The options, side by side.
| Product | Layer | How the model is chosen | Price, as published |
|---|---|---|---|
| LiteLLM | Gateway | Load-balancing strategies; a beta auto router assigns a complexity tier with a hand-weighted heuristic by default, or a classifier | Free self-hosted; enterprise on request |
| Portkey | Gateway | Conditional rules you write over metadata, parameters and path | $49/mo for 100k logs |
| OpenRouter | Marketplace | Auto Router classifies the task type, then ranks models by community spend-share | 5.5% on credits; BYOK free to $25,000 a month, then 5% |
| Requesty | Marketplace | Named routing policies; beta auto routing scores your route labels | 5% markup, 0% on own keys |
| Ramp Router | Marketplace | Caller names the model; cost routing moves eligible calls to a cheaper service tier | Free routing through 2026, list price for tokens |
| Not Diamond | Decision | Trained router predicts model and effort per step for coding agents | $0.05 per million tokens routed |
| Weave Router | Decision | Per-request scorer anchored on the requested model, cache-aware | 5.5% on top-ups, 5% on BYOK, plus seats |
| Nadir | Decision | Trained classifier per request; shadow mode first; optional verifier on complete responses | $0.002 per request, first 50 a month free; no markup on model spend |
Seven questions that separate them.
1. Does it change the model, or just the provider?
Failover, load balancing and cheaper service tiers keep the model you named and change where it runs. That saves on uptime and price per token, not on overpaying for easy prompts. Only a router that picks a different model per request changes which model does the work.
2. What signal picks the model?
Rules you write, keyword heuristics, community popularity and trained classifiers behave very differently on your traffic. Ask for the mechanism in writing, then test it on your prompts, not the vendor's.
3. Is it cache-aware?
On long agent sessions most of the bill is re-read context. A router that switches models mid-session throws away the prompt cache and can cost more than it saves. Ask whether it prices the switch, pins warm sessions, or routes only at task and subagent boundaries. The break-even math is in Your Router Is Fighting Your Prompt Cache.
4. What happens when the cheap model is wrong?
Most routers pick a model and return whatever it says. Some add a judge or guardrail that can block an answer. Fewer can escalate a weak answer to a stronger model. Ask what is checked, on which responses (streaming usually is not), and what happens when the check fails.
5. Can you run it in shadow mode first?
A router that must sit in the response path from day one asks you to trust its savings claim before you have evidence. Prefer one that can log what it would have done on your real traffic, so you can price the decision before you enforce it.
6. Where does your prompt go, and what is the markup on your own keys?
Marketplaces and hosted gateways see every prompt and response. Some decision layers need only the prompt, or only metadata. Check the data path against your privacy rules, and check the fee on bring-your-own-key traffic: it ranges from 0% to 5% across the list above.
7. How do you measure it?
Measure cost per successful task, served accuracy and the rate of bad answers together. A router that cuts cost by sending hard prompts to a weak model looks good on the bill and bad in production. The method is in What is an LLM router?
Which one fits.
- One app, a few models, no routing yet. Start with a gateway (LiteLLM self-hosted, or Portkey hosted) for keys, budgets and logs.
- You want every model on one bill. A marketplace. OpenRouter has the widest catalog; Ramp Router has free routing through 2026.
- Coding agents on one expensive default model. A decision layer built for agent sessions that is cache-aware: Not Diamond, Weave Router or Nadir.
- You need to show finance or security why a model was picked. A decision layer with a per-request receipt and a shadow mode, beside the gateway you already run.
- The cheap model already passes nearly everything, or your caches stay warm. You may not need a router at all. We published our own negative results in When LLM Routing Does Not Help.
Where Nadir fits.
Nadir is a decision layer. It runs beside the gateway you have, including LiteLLM as the native nadir/auto provider, Kong, APISIX and any OpenAI-compatible slot. It starts in shadow mode, returns a receipt per decision (model, effort, cache, policy and economics), and accepts outcomes, so you enforce only after the evidence clears your quality bar. The managed proxy can also execute the plan and verify complete non-streaming answers. For Claude Code and Codex, Nadir Auto picks the model before each turn and checks the cache switch cost first.
Where others win: LiteLLM and Portkey are mature gateways with far larger communities and compliance certifications Nadir does not have yet. OpenRouter has a far larger model catalog. Ramp Router's routing is free through 2026. Nadir publishes no universal savings rate. Run it in shadow mode on your own traffic and decide from the receipt.
Conclusion.
Name the missing layer before you compare products. Keep the gateway you run. Add a router only where one default model is overpaying, test it in shadow mode on your own traffic, and judge it on cost per successful task, not on a vendor's percentage.