OpenRouter alternatives in 2026: an honest comparison (LiteLLM, Portkey, Helicone, Anvat)

OpenRouter is the best-known LLM gateway because it solved one problem extremely well: "I want every model behind one key, with one bill, and I don't want to manage anything." For a huge slice of developers that's still the right answer in 2026.

But the gateway market has matured. Self-host options are real now, specialised gateways have emerged for observability and compliance, and at least one major player shifted into maintenance mode. If you're picking a gateway today, the decision matrix looks different from 2024.

Here's the honest version of where each option fits, including the trade-offs they don't put on the landing page.

The current landscape (mid-2026)

A few things worth knowing before you compare further:

• Helicone was acquired by Mintlify in early 2026 and is in maintenance mode. Existing users are fine; new projects should pick something else unless you specifically want their dashboard.
• LiteLLM suffered a supply-chain incident in 2026. Self-hosted users are fine if they pin versions and audit upstream, but the operational burden went up.
• Portkey open-sourced the gateway under Apache 2.0 in March 2026. The paid tiers ship guardrails and SOC 2 audit trails on top.

When OpenRouter is still the right answer

Three scenarios where OpenRouter genuinely wins:

1. You need the long tail of models. OpenRouter lists ~400 models including open-source variants, fine-tunes, and community-hosted endpoints. If your use case demands access to a Yi variant, a fine-tuned Llama, or a Kimi K2 model that you'd otherwise have to find a separate provider for — nobody else has the breadth.
2. You want absolute zero ops. Sign up, get a key, pay through one billing surface. No server to host, no scale to plan for, no Docker images to maintain.
3. You're optimising for variety, not cost. OpenRouter's markup is small (~5.5%) but it's still positive. You're paying a small premium for convenience and the marketplace's hands-off failover.

When to look elsewhere

You need the lowest possible cost per token

If your bill is dominated by 2–3 frontier models (which is true of most teams running Claude Code, Cursor, or agent stacks), a discount-focused gateway like Anvat prices below list rather than above it: 30% off Anthropic/OpenAI metered rates plus 2× credit match on prepaid packs. The breakeven against OpenRouter happens immediately because Anvat's effective rate is lower than provider-direct.

You need self-hosting + data sovereignty

LiteLLM is the obvious pick. MIT licensed, runs as a Docker container or a Python proxy, supports the entire long tail of providers. The trade-off is operational: someone owns the deployment, the upgrades, the security audit, and now (post supply-chain incident) the dependency review.

If you're a regulated team (HIPAA, FedRAMP, EU data residency), self-hosting LiteLLM is often the only path. Otherwise the operational cost usually exceeds the gateway markup you're trying to avoid.

You need production guardrails + governance

Portkey has the most feature-complete control plane out of the box: PII redaction, jailbreak detection, prompt injection filters, semantic caching, prompt management with versioning, RBAC, SOC 2-ready audit logs. Pricing scales: free OSS tier, $49/mo hosted with 100K logs, enterprise custom (4-figure monthly for SOC 2 / HIPAA / VPC).

Pick Portkey if your auditors actually want to see the guardrails, not just hear about them.

You're already on Vercel / Cloudflare

Vercel AI Gateway is bundled with a Vercel plan and integrates natively with the AI SDK. Zero config beyond a few lines of code. Cheapest path if you're already paying for Vercel.

Cloudflare AI Gateway is free up to a generous tier and pairs naturally with Cloudflare Workers. $5 per 1M requests above free. If your edge is already Cloudflare, this is the lowest-friction option.

You only care about observability

If you've already picked a gateway for routing and you just want a clean logging surface on top, Helicone's dashboards are still excellent — but since Mintlify acquired them and put them in maintenance mode, new projects should pair LiteLLM with Langfuse or use Portkey's built-in observability instead.

A practical decision matrix

Rather than ranking gateways head-to-head, decide on your single most important constraint first:

If you fall in two categories, pick the more expensive constraint to satisfy first — cost optimisation is easy to bolt on later, but governance and data residency are not.

What we run at Anvat (full disclosure)

We're obviously biased — Anvat is a gateway. But here's how we'd actually recommend pairing tools:

• Day one, building a product: OpenRouter or Anvat. Pick by whether cost or breadth matters more. The decision is reversible in minutes.
• Series A, scaling traffic on frontier models: Anvat for the 80% of traffic that hits Claude / GPT / Gemini, OpenRouter on standby for the long tail. Both keys live in the same env.
• Series B+, hiring a platform team: Self-host LiteLLM for the primary path, keep a managed gateway as the failover. Add Portkey for guardrails if you ship to regulated customers.

The interesting pattern in 2026: gateways stack. The right answer is rarely "one gateway forever" — it's "two or three, layered by what each one is best at."

The supply-chain footnote

A short list of operational hygiene every gateway customer should adopt this year:

1. Pin versions for any self-hosted gateway. Don't run :latest in production. The LiteLLM incident showed why.
2. Use dependency scanning (Snyk, Dependabot, npm audit) on any gateway codebase you self-host.
3. Rotate API keys quarterly. Most gateways issue tokens with no built-in expiry. Add your own.
4. Have a fallback. Single-vendor dependency is a real risk — keep one alternative gateway integration tested and ready to flip to.

Bottom line

OpenRouter is fine. It's still the right pick for "I want every model and I don't want to think about infra." But if your real constraint is cost on the frontier models you actually use most, look at a discount-focused gateway. If your constraint is self-hosting or compliance, LiteLLM or Portkey. If your constraint is observability, pair whatever you pick with Langfuse or stay on Helicone for now.

The "best gateway" question doesn't have one answer in 2026 — it has six, and the right one is the one that solves your single biggest constraint.