Best LLM Routers in 2026: A Field Guide From Inside the Category
Disclosure first, because every "best LLM router" roundup you'll read this year is written by a vendor and most of them hide it: this one is published by TierUp, and TierUp is on the list. We'll state our trade-offs as bluntly as everyone else's, cite sources you can check, and tell you plainly when we're the wrong pick.
Second thing worth knowing: "LLM router" now means at least three different products, and roundups that rank them on one axis are comparing a truck to a bus lane. Sort the field by one question instead — who owns the model choice? — and it falls into place.
The question that splits the field
Every router sits between your code and the model providers. The differences:
- You pick the model, the router carries it. Unified APIs and gateways: one endpoint, many models, but
"model": "..."is still a string you chose and maintain. - The router picks per prompt. A classifier reads each request and picks a model for it. Cost-efficient; also means output can vary request to request.
- You pick a quality class, the vendor maintains what's behind it. The model choice — and the churn of re-evaluating it every time pricing or checkpoints change — becomes the vendor's job.
Most of the market is type 1. Type 2 exists at the edges. Type 3 is where we are, nearly alone — more on why that is at the end.
The field, honestly
OpenRouter — the incumbent, and the default answer. One hosted API over 400+ models, pass-through pricing plus a 5.5% fee ($0.80 minimum) on credit purchases, mature, battle-tested at scale. If you want breadth without ops, start here; we say the same on our own comparison page. The honest caveats: you're still picking model IDs (type 1, with routing assists), and quality across its provider pool is uneven — HN threads this spring complained about being routed through degraded quantized models and not being able to see which provider actually served a request. Not a reliability knock — their recent uptime record is clean — but provider variance is a real thing you'll manage.
LiteLLM — the self-host answer. Open-source proxy/SDK over essentially every provider, with routing, fallbacks, budgets, and spend tracking you configure yourself. Maximum control, no per-token middleman, keys never leave your infra. The cost is that you run it: the project's own production docs recommend recycling workers to bound memory growth, long-standing memory-leak issues mean real on-call load, and self-hosting means owning supply-chain risk on every upgrade — the repo's most-reacted issue ever is the March 2026 PyPI compromise. Right pick for teams with infra appetite; see our LiteLLM comparison for the longer version.
Portkey — routing as part of an ops/governance platform: observability, guardrails, prompt management, flat SaaS pricing (Production at $49/month plus request overages, per their pricing page). If your problem is "we need visibility and policy over LLM traffic," Portkey is competing for that job, not the model-choice job. Our comparison covers where the two overlap.
Helicone — observability-first, routing as a feature. If you mostly want to see, debug, and analyze your LLM traffic, it's a strong tool with usage-based pricing (Pro at $79/month). Like Portkey, it treats routing as plumbing rather than the product.
Requesty — a unified gateway over 400+ models at 5% of spend. Closest in shape to OpenRouter; the interface is still model names (type 1).
NotDiamond — the most prominent type 2: per-prompt intelligent routing, now positioned mainly at coding agents ("the world's most powerful intelligent model router for coding agents," per their site). If per-prompt cost/quality optimization inside an agent loop is your problem, it's the specialist.
Inworld Router — one endpoint, model selection via rules you author (their docs show tier-style routing as a config example — e.g. free-plan users get a cheaper model). Tiering exists here, but as DIY config you write and maintain, which makes it a very flexible type 1.
TierUp — us, the type 3. You send tier-1 through tier-4 (Speed → Reasoning) instead of a model name, on an OpenAI-compatible endpoint; we map each tier to what we assess as the best value in that class and re-shop it as the market moves, so a better model shipping means your tier silently upgrades. Flat per-tier pricing at ~50% under the retail of the underlying models (tier details); tier 1 is currently free. The trade-offs, stated as plainly as we stated everyone else's: you give up model pinning entirely — if you need a specific checkpoint, use LiteLLM or OpenRouter, full stop. We're very early, with roughly zero production users. And the discount is transparently subsidized — we route through OpenRouter today and price below cost on purpose while we test whether tier-routing is something developers actually want.
Two names you'll still see in older roundups
Martian, the best-known routing pioneer, has pivoted to interpretability research and no longer sells routing. Unify.ai left the category entirely and now sells AI agent staffing. Worth knowing for two reasons: don't evaluate products that no longer exist, and note what their exits imply — routing mechanics alone didn't sustain a company. That's exactly why we sell the maintained tier contract rather than the plumbing, and honestly, it's also the risk you're pricing in when you bet on anyone in this category, us included.
So which is "best"?
Depends entirely on which question you're asking:
| Your actual question | Start with |
|---|---|
| "I want one API over every model, hosted" | OpenRouter (Requesty as the comparison quote) |
| "I want that, but in my infra, under my control" | LiteLLM |
| "I need observability and governance over LLM traffic" | Portkey or Helicone |
| "Pick the best model per prompt inside my coding agent" | NotDiamond |
| "I want to stop owning model choice altogether" | TierUp — that's the product |
If you're cost-driven, do the arithmetic before picking anything: our cost calculator has dated, sourced prices for 13 models and tiers, and the cost-reduction guide covers the five levers that matter more than which router you choose. For the raw per-token numbers behind all of this, see the July 2026 pricing survey.
And if "pick a tier, not a model" sounds like the version you'd actually maintain, the playground needs no signup — pasting your real prompt into tier 2 answers the quality question faster than any roundup, ours included.
Competitor claims verified 2026-07-07 against the linked primary sources (vendor docs, pricing pages, GitHub issues, HN threads). Pricing and positioning in this market drift fast — re-check before you commit.
