Honest comparison

When Leanroute beats the alternatives, and when it doesn't.

We're in a crowded category. OpenRouter has the most users, LiteLLM has the deepest open-source roots, Bifrost is a newer fast challenger. Below: how we differ on the dimensions developers actually care about, including the cases where you should pick one of them over us. All claims dated to October 2026; competitors evolve and we re-validate quarterly.

AI gateways at a glance

The questions an evaluation turns on, answered for each gateway. Last checked October 2026 from vendor docs and pricing pages; vendors change often, so confirm anything that decides your choice. Click a name for the detailed comparison.

	Leanroute	OpenRouter	Portkey	Bifrost	Kong AI Gateway	Vercel AI Gateway	Cloudflare AI Gateway
Who runs it, and where	Managed by us, hosted in Singapore	Managed; global, with US- or EU-locked inference on its Business tier	Managed (US; EU on enterprise), or self-host its open-source gateway. Now part of Palo Alto Networks	You run it; enterprise edition in your VPC or on-prem	You run Kong Gateway, or use Kong Konnect (cloud)	Managed by Vercel, US primary	Managed on Cloudflare's global edge; data region selectable on paid plans
Pricing	BYOK: flat $15 or $25 a month. Credits: 2% at top-up plus card processing. Models at list price	Models at list price; 5.5% fee on credit purchases	Free tier; Production from $49 a month plus per-log fees; Enterprise custom	Open source free; Enterprise custom-priced	Gateway core free; most AI features need paid Kong Enterprise or Konnect	$5 a month free credit, then tokens at list price; per-request add-ons for ZDR, allowlists and reporting	Free (with Workers plan usage); you pay providers directly
Your own provider keys (BYOK)	Yes, encrypted at rest, for the flat fee	Yes; no fee up to $25,000 a month of inference, 5% above	Yes	Yes (always your keys)	Yes, through Vault or your secret store	Yes	Yes, stored in Cloudflare Secrets Store
Routing and failover	Cheapest live model in a quality tier, automatic failover on provider errors	Automatic provider fallbacks	Fallback chains and load balancing	Automatic fallbacks and weighted load balancing across keys	Multi-model load balancing and semantic routing in AI Proxy Advanced (Enterprise)	Provider routing and fallbacks	Fallbacks and dynamic routing rules
Hard spend caps	Daily and monthly caps per org and per key, checked before the request is sent	Per-key credit limits	Budgets and rate limits	Budgets on virtual keys (open source)	Token-based AI rate limiting (Enterprise)	Check their docs	Spend limits per gateway, by model, provider or metadata
Guardrails	Seven rules built in: PII, secrets, prompt injection, moderation, exfiltration, profanity, language	Not documented as a gateway feature	Built-in library: PII, moderation, regex and more	Check their docs for your edition	AI Prompt Guard (open source); semantic guard and sanitizer in Enterprise	No native guardrails; moderation models can be called through it	Content moderation and DLP profiles in the gateway firewall
MCP tools	Gateway runs the tool loop on any model, allowlisted servers only	Not documented as a gateway feature	Gateway-side execution in beta, plus an MCP gateway for registered servers	MCP gateway with Agent Mode and per-key tool filtering	AI MCP Proxy (Enterprise)	In your app through the AI SDK's MCP client	Not documented as a gateway feature
Agent runs (cost per session)	X-Agent-Session-Id groups requests into runs: cost, models and tools called per run, in the dashboard and API	session_id groups requests; traces sent to your own observability tool (Broadcast)	Trace IDs group related calls into traces; cost recorded per request	Vendor cites session grouping for coding agents; cost by virtual key, team and customer	Cost per request and per consumer; no per-run grouping documented	Spend by user, tag or API key through its reporting API	Up to five metadata tags per request, searchable in logs
Open source	No, managed only	No	Yes, gateway open-sourced in March 2026	Yes	Gateway core yes (Apache 2.0); most AI plugins are Enterprise	No	No
Security and compliance	PDPA-aligned, DPA on request, keys encrypted (AES-256-GCM), monthly automated security scans. No SOC 2 report yet	SOC 2 Type II reported; see their trust center	SOC 2 Type II	Enterprise: SOC 2 Type II cited by the vendor. Open source: your own controls	Self-hosted: your own controls. Konnect: see Kong's trust center	SOC 2 Type II	SOC 2 Type II and ISO 27001

What every router gets right

OpenAI-wire-compatible (you change a base URL, your code doesn't change)
Multi-provider routing (OpenAI, Anthropic, Google all reachable)
Some flavor of caching, observability, and BYOK
A pitch about "saving you money" on LLM bills

The interesting differences are in which providers the catalog leads with, how the bill is shaped (per-request fee vs. flat subscription vs. deposit fee), where data is hosted, and how the routing handles streaming + tools at the edges.

Leanroute vs Cloudflare AI Gateway

Cloudflare AI Gateway

Free edge-network LLM proxy

Cloudflare's AI Gateway: edge caching, rate limiting, DDoS protection and request logs across their global network.

Choose Leanroute when

MCP tool calling on any model with a per-org allowlist, guardrails that also cover prompt injection, secrets and exfiltration, APAC + India provider depth, and Singapore hosting, under one flat monthly price.

Choose Cloudflare AI Gateway when

Mostly free, on a global edge network, with caching, guardrails, stored provider keys and spend limits built in. If your app already runs on Cloudflare, they're hard to beat on price.

Detailed comparison →

Leanroute vs AWS Bedrock

AWS Bedrock

Managed foundation-model service on AWS

AWS's managed foundation-model service: Anthropic, Meta, Amazon Nova and Mistral hosted on AWS infrastructure with IAM + VPC integration. Bedrock-specific API, not a gateway.

Choose Leanroute when

We front Bedrock via BYOK (shipped 2026-08-30): your AWS credentials, your region, AWS invoices you directly. On top: OpenAI-compatible wire format, MCP tool calling on any model, cross-provider failover (a Bedrock 5xx doesn't take your app down), semantic cache, per-org guardrails and spend caps. Use both together.

Choose AWS Bedrock when

Deep AWS ecosystem integration (IAM, VPC PrivateLink, EDP contracting, Provisioned Throughput SLAs, Amazon Nova exclusive). Bedrock stays underneath. Leanroute doesn't replace it; we add the routing and observability layer on top.

Detailed comparison →

Leanroute vs Groq

Groq

LPU-accelerated open-weight inference

Custom LPU hardware for Llama/Mixtral/DeepSeek inference. Fastest tokens/sec on the market for supported OSS models. An inference host, not a gateway.

Choose Leanroute when

Multi-provider routing across 100+ models on 14 curated providers with one OpenAI-compatible endpoint, MCP tool calling on any model, per-org guardrails, cross-provider failover. AWS Bedrock BYOK path shipped 2026-08-30: route Anthropic Claude through the customer's own AWS account (ap-southeast-1 launch region) alongside every other provider.

Choose Groq when

Nothing matches Groq on raw tokens/sec for Llama and Mixtral. For real-time voice agents or streaming that has to feel instant on OSS models, direct-to-Groq is the answer.

Detailed comparison →

Leanroute vs Together AI

Together AI

Open-weight inference host with fine-tuning

100+ OSS models (Llama, Mixtral, Qwen, DeepSeek, Yi, Nous, community fine-tunes) with per-token pricing, LoRA fine-tuning, and an OpenAI-compatible API.

Choose Leanroute when

Multi-provider gateway that reaches every closed model (OpenAI, Anthropic, Google, xAI) plus APAC/India providers, with MCP tool calling on any model and per-org guardrails. Together adapter on v2 roadmap for OSS routing through the same endpoint.

Choose Together AI when

Deepest OSS catalog including community fine-tunes we don't carry. LoRA fine-tuning on their infra. SOC 2 + HIPAA today. If your workload is exclusively OSS models or you need fine-tuning, direct-to-Together wins.

Detailed comparison →

Leanroute vs Vercel AI Gateway

Vercel AI Gateway

Hosted gateway bundled with Vercel + AI SDK

Zero-markup routing across Western frontier providers, deeply integrated with the Vercel AI SDK. $5/mo free credits, then pay-as-you-go tokens with per-1k-request add-ons for ZDR, allowlists, and reporting.

Choose Leanroute when

MCP tool calling on any model via `mcp_servers`, ZDR + allowlists + reporting included in the flat $15/$25 subscription (Vercel bills per-1k for each), APAC + India catalog first-class (Qwen, GLM, Doubao, Kimi, Sarvam, Krutrim), and Singapore hosting.

Choose Vercel AI Gateway when

Deep bundling with Vercel and the AI SDK, one env var away if you already deploy on Vercel. Best fit when your workload is small enough that $5/mo free credit + pay-as-you-go tokens beats any flat subscription, and you don't need the feature add-ons.

Detailed comparison →

Leanroute vs OpenRouter

OpenRouter

Hosted LLM router

The category-defining hosted router. 400+ models, 5.5% fee on credit purchases, fee-free BYOK up to $25k/month.

Choose Leanroute when

APAC + India catalog first-class (Qwen, GLM, Doubao, Kimi, Sarvam, Krutrim), Singapore hosting, MCP tools on any model, guardrails and per-key hard caps. Flat BYOK subscription that beats a 5% fee once BYOK volume passes $25k/month.

Choose OpenRouter when

Model breadth (400+ including a long tail of fine-tunes), free BYOK for most teams, longer track record, larger community.

Detailed comparison →

Leanroute vs LiteLLM

LiteLLM

Self-hosted proxy + Python SDK

Open-source LLM proxy. Self-host the gateway yourself or use their managed cloud tier. Strong Python SDK.

Choose Leanroute when

Managed-first with billing, dashboard, request log, BYOK encryption, and per-org cache scoping wired up out of the box. No infrastructure to run.

Choose LiteLLM when

Self-host for full control. Open source. Largest model coverage (200+) via their unified Python SDK. Best fit when you already operate Kubernetes and want one less SaaS bill.

Detailed comparison →

Leanroute vs Portkey

Portkey

Enterprise gateway, now part of Palo Alto Networks

Closest functional peer in feature scope: routing, cache, observability and guardrails. Being folded into Palo Alto Networks' Prisma AIRS; the gateway core is open source.

Choose Leanroute when

APAC + India provider depth, Singapore hosting, $15/$25 BYOK subscription or credits with a 2% top-up fee (no tier ladder, no seat costs). MCP tools on the standard chat-completions API, plus a 7-rule guardrails library on every plan.

Choose Portkey when

Guardrails library breadth (PII / moderation / regex / profanity) first-class. SOC 2 Type II today. US enterprise sales motion. Prompt management included.

Detailed comparison →

Leanroute vs Helicone

Helicone

Observability-first proxy

Open-source LLM observability layer. Rich analytics, prompt versioning, traces, evals. More complementary to us than competitive.

Choose Leanroute when

Routing, BYOK encryption, multi-account failover, APAC catalog: the parts Helicone passes through to whatever upstream you point at.

Choose Helicone when

Deeper analytics, prompt versioning, evals against production traffic, MIT open source. Many teams use both: point Helicone at our base URL.

Detailed comparison →

Leanroute vs Bifrost

Bifrost

Go-based LLM gateway

Newer Go-based open-source gateway. Performance-focused. Self-hosted.

Choose Leanroute when

Fully managed, nothing to deploy. Prepaid billing or a flat BYOK subscription, APAC + India provider catalog, Singapore hosting.

Choose Bifrost when

Low-latency Go runtime, self-host control, no per-request markup if you run it yourself.

Detailed comparison →

Leanroute vs Kong AI Gateway

Kong AI Gateway

AI plugins on general-purpose API gateway

LLM-specific plugins riding on Kong's existing API gateway. Best fit when your org already runs Kong for general REST/gRPC traffic.

Choose Leanroute when

Purpose-built AI gateway: curated provider catalog with APAC + India depth, transparent flat-subscription or top-up-fee billing, BYOK encryption, multi-account failover, all in 5 minutes of setup, not a Kubernetes deploy.

Choose Kong AI Gateway when

One control plane for all your APIs (REST + gRPC + WebSocket + GraphQL + AI). Open source. Plugin composability with rate-limiting / auth / transformation plugins you already use.

Detailed comparison →

The honest take

If you're US-resident, want the broadest model selection you can get, and don't care about Asia-pacific providers or Singapore residency, OpenRouter is the safe pick. If you're a platform team that already runs Kubernetes and wants no SaaS bill at all, LiteLLM self-host is the right answer. If you're building something where every millisecond matters and you can operate Go services in production, Bifrost is worth a look.

Leanroute is for the developer who wants hosted ease + APAC/India provider depth + Singapore residency + a flat BYOK subscription or 0%-per-request credits and isn't willing to give up any of those to get the others. That's a specific bet on a specific market shape; if it's yours, we're the cleanest option that meets the brief.

Alternatives guides

Longer write-ups that compare several gateways at once, with decision trees and the cases where you should not switch.

Ruled out by SaaS entirely?

The comparison above is SaaS-first (Portkey, OpenRouter, Helicone) vs. open-source self-host (LiteLLM, Bifrost, Kong). We're currently SaaS-only, but actively gauging demand for a self-hostable or private single-tenant version for teams that need on-prem, air-gap, or sovereignty guarantees SaaS can't offer. If that's a blocker for you, email sales@leanroute.dev with your use case; we're keeping an early-access list.