The honest guide to choosing an LLM router in 2026
A decision framework, not a product pitch
A decision framework for picking an LLM router in 2026. Seven questions, a flowchart, and honest answers about when each option wins — HiWay included.
If you're searching for "best LLM router 2026" you're going to find a dozen posts ranking them 1-10, and most of them are optimized for whoever runs the blog. This one tries not to be. The goal here is to give you a decision framework that works for your team, honestly, with HiWay showing up as one answer among several — because that's the truth.
Seven questions. At the end you'll have narrowed 20+ options down to 1-3.
Question 1: What's your monthly spend trajectory?
This is the first filter and it sorts everything else.
Near zero / hobby project. A free tier (HiWay Free gives you 2,500 req/mo) or a direct provider call is plenty. Smart routing and fallback are nice-to-have at this scale, not decisive.
A few hundred to a few thousand € / month. A gateway pays off fast. Smart routing alone (auto-downgrading simple requests to cheaper models) typically cuts 40-85% off the inference bill — volume-independent. Integration convenience, observability, and basic cost controls become worth the subscription on top of that. Any of the options below works.
€2,000 to €20,000/month. Pricing model matters a lot. Markup pricing compounds against you. Flat subscriptions or BYOK models become meaningfully cheaper, and smart routing compounds on top. Feature depth (per-endpoint budgets, burn-rate alerts) has concrete ROI.
€20,000+/month. You're in enterprise territory. Custom pricing, dedicated support, compliance commitments matter. Talk to sales teams, not marketing pages.
This isn't a prescription — you can pick any gateway at any scale. It's a prescription about what to weight most when picking. And note the volume-independent levers (smart routing, BYOK, EU hosting) don't need you to cross a spend threshold before they're worth it.
Question 2: Do you need EU hosting?
Binary. Answer yes or no.
Yes, hard requirement (compliance, regulated sector, customer mandate). Your short list drops to: HiWay, self-hosted LiteLLM in an EU region, or Helicone OSS self-hosted in EU. Everything else is off the table until you get a specific regional commitment signed.
Soft preference (nice to have, not a deal-breaker). You can consider US-hosted options with EU regions — some Portkey enterprise tiers, Vercel AI Gateway if you're on Vercel with EU regions. But recognize you're compromising on data residency for feature richness.
No. The full list is available.
Question 3: Do you want BYOK or are you fine with a reseller markup?
BYOK. You want your provider keys in your accounts, providers billing you directly, no markup on inference. Short list: HiWay, LiteLLM (self-hosted), Helicone (self-hosted or cloud with BYOK), Portkey (with BYOK config), Cloudflare AI Gateway.
Reseller is fine. You want one-click signup, one credit card, no provider accounts to manage. Short list: OpenRouter, Requesty, other aggregators.
The BYOK question is often a proxy for "how much do I care about cost alignment over 12+ months?" If your horizon is short and your spend is small, markup is fine. If you're playing long, BYOK pays.
Question 4: Where do you host your app?
Vercel + Next.js + Vercel AI SDK. The Vercel AI Gateway has an integration advantage that's hard to beat if that's your context. Consider it first, evaluate alternatives only if other constraints kick in.
Cloudflare (Workers, Pages). Cloudflare AI Gateway has an edge-proximity advantage. Same logic: first choice if that's your stack.
AWS, GCP, your own servers, Fly, Render, Hetzner, VPS, Kubernetes. No specific platform gateway advantage. Pick based on features, pricing, and hosting region. HiWay, LiteLLM, Portkey, Helicone all work equivalently from this position.
Mixed. The gateway that wins is the one with the cleanest OpenAI-compatible interface, because that's the one easiest to integrate across heterogeneous deployments. Most modern gateways fit.
Question 5: Do you want intelligent routing, or just fallback?
Intelligent routing means: the gateway reads the prompt (or metadata) and picks the best model tier per request. Simple greeting → cheap model. Hard reasoning → flagship model. The gateway has opinions.
Fallback routing means: the gateway calls model A; if A fails, it tries B; if B fails, it tries C. The gateway doesn't have opinions about what's best, it just tries things in order.
If you want intelligent routing as a first-class feature: HiWay, Martian, LiteLLM with custom routing config, or a router-first specialist. Most managed gateways support it at some level but vary in sophistication.
If fallback is enough: basically all gateways support it. Pick based on other criteria.
Intelligent routing's ROI is biggest when your workload is heterogeneous — mixed greetings, complex questions, tool use, structured outputs. If your app does one thing with one model, you don't need intelligent routing. You need a good fallback.
No credit card required
Question 6: Do you have platform engineering capacity?
Yes, dedicated platform engineer, existing Kubernetes/PostgreSQL/Redis infra. Self-hosting is on the table. LiteLLM or Helicone OSS are legitimate answers. You'll spend engineering time, but you'll save cash and keep control.
No, small team focused on product. Managed gateways only. The engineering overhead of self-hosting will eat into your product velocity. Pick a managed option that matches your other answers.
This question kills a lot of "should I use LiteLLM?" debates. If your team can't handle the ops, the OSS savings are an illusion — you pay in engineer hours at 2 AM instead of vendor dollars on a predictable invoice.
Question 7: What matters most over 24 months — features, cost alignment, or compliance?
This is the weighting question.
Features weight highest. You want the richest observability UI, the most polished prompt management, the deepest integration with evals. Helicone, Portkey, LangSmith weight up.
Cost alignment weights highest. You want a gateway whose revenue doesn't grow when yours does. BYOK + flat per-request or OSS self-hosted weight up. HiWay, LiteLLM, Helicone (self-hosted) weight up.
Compliance weights highest. EU hosting, DPA specificity, sub-processor transparency, EU entity. HiWay or EU self-hosted LiteLLM weight up.
Most teams weight all three, but there's usually one that's disproportionately important. Identify it and let it break ties.
The flowchart, in ASCII
START
│
▼
Near-zero spend / hobby project?
├── Yes → HiWay Free (2.5K req/mo) or direct. Stop.
└── No
│
▼
Need EU hosting?
├── Hard yes → [HiWay | EU LiteLLM | EU Helicone OSS]
├── Soft yes → [Portkey EU | Vercel EU | EU self-host]
└── No → continue
│
▼
BYOK or markup OK?
├── BYOK → [HiWay | LiteLLM | Helicone | Portkey BYOK]
└── Markup OK → [OpenRouter | Requesty | aggregators]
│
▼
On Vercel?
├── Yes → Vercel AI Gateway
└── No
│
▼
On Cloudflare?
├── Yes → Cloudflare AI Gateway
└── No → dedicated router
│
▼
Platform engineering?
├── Yes → [LiteLLM | Helicone OSS]
└── No → [HiWay | Portkey | managed Helicone]
Walk your constraints down the tree. You'll hit a leaf with 1-3 names.
How HiWay shows up in the framework
For transparency, here's where HiWay is the right answer:
- Any production spend where the inference bill is worth trimming (smart routing pays back at any scale)
- EU hosting is a hard requirement or strong preference
- BYOK is a hard requirement or strong preference
- You're not locked into Vercel or Cloudflare's ecosystem (or you're there but EU hosting wins)
- You want intelligent routing as a first-class feature
- You don't have dedicated platform engineering capacity (or you do but want to free it for product work)
- Cost alignment or compliance weights highest over features
If your answers put you there, we'd like to earn your traffic. If your answers put you elsewhere, use the elsewhere option. That's not false modesty; it's the truth about what we're optimized for.
Common anti-patterns
Picking by Twitter popularity
The loudest LLM tools on Twitter are not necessarily the best for your team. Volume of marketing doesn't correlate with fit. Do the framework.
Picking by feature count
More features doesn't mean better. A gateway with 47 features where 4 are great and 43 are mediocre is worse than a gateway with 12 features done well. Ask what each gateway's opinion is about what matters.
Picking by "most mature"
Mature gateways have deep ops tooling and dusty features you'll never use. Newer gateways have narrow feature surfaces focused on current best practices. Mature isn't automatically better.
Picking by the cheapest sticker
Sticker price isn't total cost. A cheap gateway with a 5% markup is more expensive at scale than a flat subscription at a higher-looking sticker. Run the 12-month math with your projected spend.
Never revisiting the decision
The market moved a lot from 2024 to 2026 and it will move again by 2028. Whatever you pick now, put a calendar reminder to re-evaluate in 12 months. Your spend will have grown, your constraints will have shifted, and the landscape will have changed.
A worked example
Let's run the framework for a hypothetical team: a 10-person French B2B SaaS serving insurance companies, €4,000/month LLM spend, growing 20% month-over-month. Compliance team is strict. App hosted on their own Kubernetes on OVH. Team of developers, no dedicated platform engineer.
- Q1 (spend): €4K/mo, growing. Pricing model matters. Flat pricing is attractive.
- Q2 (EU hosting): Hard yes, insurance compliance. US-hosted options off the table.
- Q3 (BYOK): BYOK preferred for transparent billing and control. Reseller markup not attractive.
- Q4 (hosting): Own Kubernetes on OVH. No platform-specific gateway advantage.
- Q5 (routing): Mixed workload, intelligent routing has clear ROI. Want first-class support.
- Q6 (platform capacity): No dedicated platform engineer. Self-hosted LiteLLM is risky.
- Q7 (weighting): Compliance weights highest, followed by cost alignment.
The framework narrows this team to: HiWay, or possibly LiteLLM if they can allocate a fraction of an engineer to run it. Given Q6, HiWay is the safer answer. A managed EU-hosted BYOK gateway with flat pricing and intelligent routing matches every constraint.
This is a real pattern. We have customers who look exactly like this.
A worked counterexample
Now a team where HiWay isn't the answer: 3-person US startup, $300/month LLM spend, prototyping fast, no compliance constraints, hosted on Vercel with Next.js and the Vercel AI SDK.
- Q1 (spend): Low, convenience dominates.
- Q2 (EU hosting): Not needed.
- Q3 (BYOK): Markup is fine at this scale.
- Q4 (hosting): Vercel + Next.js + Vercel AI SDK. Integration advantage is huge.
- Q5 (routing): Fallback is enough for now.
- Q6 (platform capacity): Three people, all product.
- Q7 (weighting): Features and speed, not cost alignment.
The framework narrows this team to: Vercel AI Gateway or OpenRouter. HiWay isn't competitive here, and we wouldn't pretend otherwise. Pick Vercel AI Gateway if you want tight integration. Pick OpenRouter if you want the broadest model catalog and one-click onboarding.
The takeaway
There is no universal best LLM router in 2026. There's the right one for your answers to seven questions. Most "top 10" posts skip straight to features because features are easier to compare than constraints — but constraints are what actually decide.
Walk the framework. Narrow your list. Run the numbers against your projected 12-month usage. Pick the one that fits your shape.
If HiWay fits, we'd be glad to see you. If not, genuinely — use what fits, and revisit in a year.
That's the last post in this series. If you want the whole argument in one thread: pricing models matter, EU hosting is a real requirement for some teams, BYOK aligns incentives, and the right gateway for you depends on constraints you probably haven't written down yet. Now you have.
Was this useful?
Comments
Be the first to comment.