Vercel AI Gateway vs. Cloudflare AI Gateway vs. OpenRouter (2026): Which LLM Gateway Should You Actually Use?

TL;DR: The Short Answer

Quick Comparison

Figures sourced from each vendor's own documentation as of August 2026 — see References & Sources at the end of this article.

What an AI Gateway Actually Buys You in 2026

An AI gateway sits between your application and the LLM providers you call — OpenAI, Anthropic, Google, and increasingly a long tail of open-weight models served by Groq, Together, Fireworks, and others. Instead of your codebase holding a different SDK, API key, and error-handling path for each provider, you call one endpoint and the gateway handles routing, retries, and — to varying degrees — caching and spend tracking.

This has become harder to ignore in 2026 than it was two years ago, for three reasons that come up in nearly every infrastructure conversation about model usage: provider outages are still common enough that single-provider apps see real downtime; frontier models get deprecated or repriced with only a few weeks' notice, so hardcoding one model into your application logic turns into a recurring migration project; and production LLM usage has shifted from single chat-completion calls toward multi-step agents that call tools, chain models for different steps of a task, and sometimes talk to an MCP server mid-run — see our roundup of the best MCP servers for B2B workflows if you're wiring up that layer on top of whichever gateway you choose. A gateway doesn't fix any of those problems by itself, but it gives you one place to configure fallback behavior, one bill to reconcile, and one dashboard to check when something breaks at 2 a.m.

Vercel AI Gateway: Pricing, Features, and Limits

Vercel AI Gateway is built into the same dashboard developers already use to manage deployments, and its core pricing pitch is unusually simple: no markup and no platform fee on tokens, on the free tier or the paid tier, and even when you bring your own provider key (BYOK). You pay the provider's own list price; Vercel's revenue comes from the surrounding platform, not a spread on inference.

The free tier gives every team access to a subset of models with lower, per-model rate limits — enough to prototype, not enough to run production traffic at scale. Purchasing AI Gateway Credits moves your team to the paid tier, which unlocks the full model catalog and higher rate limits. BYOK is paid-tier only: if a request using your own key fails, AI Gateway automatically retries it with Vercel's system credentials for reliability, and that fallback usage draws from your credit balance.

Beyond raw pricing, the feature set is clearly aimed at teams already inside the Vercel ecosystem: one API key across hundreds of models, OpenAI Chat Completions and Anthropic Messages API compatibility (so existing SDK code often needs minimal changes), automatic retries to a different provider on failure, embeddings support, and per-generation cost, latency, and token data available from the dashboard or a GET /v1/generation REST call. Compliance-minded teams on Pro or Enterprise plans get Zero Data Retention (ZDR) routing at no extra charge per request, or for $0.10 per 1,000 requests if they want it enforced team-wide across every call automatically — worth knowing if your legal team is working through something like an EU AI Act compliance checklist for B2B SaaS buyers.

One practical limit worth flagging before committing a long-running agent architecture to Vercel AI Gateway: if you're calling it from a Vercel Function, that function inherits Vercel's own execution-duration ceiling — 300 seconds by default, extendable to 800 seconds (generally available) or 1,800 seconds (beta) on Pro and Enterprise. That covers most chat and RAG workloads comfortably, but a genuinely long multi-hour agent run needs Vercel Workflows layered on top, not a single function call.

Cloudflare AI Gateway: Pricing, Features, and Limits

Cloudflare takes a different starting position: the core gateway — dashboard analytics, response caching, and rate limiting — is free on every Cloudflare plan, including the free Workers plan, with no separate per-call gateway fee. What you pay for is everything layered around it. Persistent request logs are capped at 100,000 per month across all your gateways on the free Workers plan, rising to 10 million per gateway on Workers Paid; Logpush volume beyond the included 10 million requests a month on Workers Paid costs $0.05 per million.

The bigger pricing decision is whether to turn on Unified Billing, which lets you pay for third-party model usage — OpenAI, Anthropic, and the rest of Cloudflare's supported provider list — directly through your existing Cloudflare invoice instead of managing separate billing relationships with each provider. It's a convenience feature, priced as one: a flat 5% surcharge on credits purchased through Unified Billing, with the underlying provider inference pricing passed through unmarked-up.

On features, Cloudflare's provider list is the widest of the three on paper — OpenAI, Anthropic, Google AI Studio, Groq, Mistral, Cohere, Perplexity, DeepSeek, Cerebras, xAI, Workers AI, plus AWS Bedrock, Azure OpenAI, and Replicate — reflecting its position as a neutral network layer rather than a model vendor. Cross-provider failover exists via Dynamic Routing, but unlike Vercel's or OpenRouter's default-on automatic retry, you configure the fallback rules yourself. Guardrails (content-safety filtering) and DLP scanning are both free features, though Guardrails bills separately as Workers AI token-based inference based on prompt and response length, and DLP without a Zero Trust subscription is limited to two predefined profiles (Financial Information, and Social/Insurance/National Identifier Numbers).

The other structural difference is latency positioning: Cloudflare AI Gateway runs as a proxy across Cloudflare's edge network — more than 330 locations globally — rather than inside a specific serverless runtime. Community-reported overhead for the proxy hop itself is small, generally in the 10–60ms range depending on how close your traffic originates to a Cloudflare point of presence, though most public benchmarks measure this against a mock upstream rather than a live model response, which tends to overstate how much the gateway itself matters once real provider latency of 500ms–3 seconds is added in.

OpenRouter: Pricing, Features, and Limits

OpenRouter is structured less like a platform feature and more like a marketplace: it resells access to model inference from dozens of providers, and its business model is a fee on the credits you buy rather than a fee baked into the platform itself. Add funds by card through Stripe and OpenRouter takes 5.5% of the purchase (a $0.80 minimum applies, which is why a small $5 top-up effectively costs 16% in fees, not 5.5%); pay with cryptocurrency through Coinbase and the fee drops to 5%. Once credits are in your account, you spend them at the provider's own per-token rate.

Bring Your Own Key changes that math only partially: OpenRouter waives its fee on the first 1 million BYOK requests per month, then charges 5% of what the same model and provider would normally cost on everything after that. If you're already spending enough to justify BYOK on another gateway, it's worth modeling whether your traffic actually clears that 1-million-request line before assuming OpenRouter's BYOK path is free in practice.

Where OpenRouter clearly leads is catalog breadth: 500+ models across dozens of providers, including a rotating lineup of free models — as of mid-2026, drawn from providers like Meta, Google, NVIDIA, Alibaba's Qwen, Cohere, and Poolside — that let you prototype without a credit card. Free-model rate limits step up based on account status: 50 requests a day with no credits purchased, 1,000 a day once you've added at least $10, and OpenRouter is upfront that these aren't meant for production traffic. Automatic fallback is on by default — if a provider errors out, OpenRouter routes to the next one transparently — and "latest" model aliases let you point at a flagship model without redeploying code every time a provider ships a new version.

One nuance shows up in independent latency testing: OpenRouter's architecture holds the first response byte until the first token is actually ready, so its time-to-first-byte and time-to-first-token are identical on every request — it doesn't stream headers early the way some other gateways do. In one small, single-location benchmark (n=5 runs — not a large statistical sample), OpenRouter's median cold time-to-first-token came in close to Vercel AI Gateway's (1,123ms vs. 1,099ms) but noticeably behind on warm requests (986ms vs. 822ms). Treat any single latency benchmark as directional rather than definitive — provider mix, region, and time of day all move these numbers by hundreds of milliseconds — but it's a reasonable data point if raw speed matters more to your use case than catalog breadth.

Feature-by-Feature Breakdown

Pricing and the Real Cost of "Zero Markup"

All three vendors advertise pass-through, no-markup pricing on the model calls themselves, and on that specific claim, all three are telling the truth — the per-token rate you pay for, say, a GPT or Claude model is generally the provider's own list price regardless of which gateway you route through. Where the real cost differences show up is in the fee layered on top of getting money into the system. Vercel charges nothing extra on credits at any tier. Cloudflare only charges if you opt into Unified Billing (5% on credit purchases) — skip it, and you're paying providers directly with no Cloudflare fee at all. OpenRouter's fee is the least avoidable of the three: it applies to every credit purchase (5.5% via Stripe, 5% via crypto) unless you're on the BYOK path and staying under 1 million requests a month. For a team spending roughly $2,000/month on inference, that's about $110/month in OpenRouter fees that a Vercel setup or a fee-free Cloudflare setup wouldn't charge — worth modeling explicitly rather than assuming "no markup" means "no total cost for the gateway."

Model Catalog Breadth

If breadth is the deciding factor, OpenRouter's 500+ models across dozens of providers is the largest catalog of the three, and the one most likely to already carry a niche open-weight model the day it ships. Cloudflare's named-provider list (OpenAI, Anthropic, Google AI Studio, Groq, Mistral, Cohere, Perplexity, DeepSeek, Cerebras, xAI, Workers AI, Bedrock, Azure OpenAI, Replicate) is close behind in providers, though Cloudflare doesn't publish a comparable total model count. Vercel's catalog runs to "hundreds of models" from mainstream frontier and open-weight providers — narrower in raw count than OpenRouter, but every model in it carries the same zero-markup pricing guarantee, which matters more than raw catalog size to some buyers.

Failover and Latency

Failover behavior differs in an easy-to-miss way: Vercel and OpenRouter both retry to another provider automatically, out of the box, when a request fails. Cloudflare's Dynamic Routing does the same job but isn't automatic by default — you configure which providers back each other up. That's more setup work, but it also means Cloudflare won't silently reroute a request somewhere you didn't explicitly approve, which some security-conscious teams treat as a feature rather than friction. On raw speed, don't expect a clean, permanent winner: the one independent benchmark we found comparing Vercel and OpenRouter directly (see Methodology) showed Vercel modestly ahead on warm requests and roughly tied cold, in a small sample from a single test location; Cloudflare's edge-proxy overhead is reported separately, in the 10–60ms range, and isn't a like-for-like comparison with the other two. Provider response time — not gateway overhead — remains the dominant factor in end-to-end latency for all three.

Observability, Caching, and Compliance

Vercel and Cloudflare both give you first-party observability without extra setup — Vercel's dashboard shows per-generation cost, latency, and token counts; Cloudflare's shows request-level analytics and logs, with the free plan capped at 100,000 stored logs a month. OpenRouter's dashboard is comparatively lighter on per-request detail, focused on usage and spend by API key. Cloudflare is alone in offering response caching as a built-in gateway feature — useful if your traffic has any repeated-prompt pattern worth short-circuiting before it reaches the model provider at all. On compliance, Vercel's ZDR routing (free per-request, $0.10/1,000 requests team-wide, Pro/Enterprise) is the most explicit of the three about which providers have contractually agreed not to retain or train on your prompts; Cloudflare's DLP scanning catches sensitive data patterns in requests rather than controlling provider-side retention; OpenRouter doesn't offer a gateway-level ZDR control, leaving data-retention terms to whatever underlying provider you're routed to has agreed to. If AI vendor risk is a live topic in your board updates, it's worth reading this alongside a broader look at AI governance platforms built for the EU AI Act era, since gateway-level ZDR is one control among several a governance review will typically ask about.

Pros and Cons

Vercel AI Gateway

Cloudflare AI Gateway

OpenRouter

Who Should Choose Which

The honest answer is that all three are reasonable defaults for different starting points — the right pick usually falls out of infrastructure you've already committed to, not a feature you're missing.

• You're deploying on Vercel and already use the AI SDK: Vercel AI Gateway is the least-friction option — one dashboard for deploys and model spend, zero markup, and BYOK when you're ready for it. Plan around Vercel Functions' duration limits if you're building long-running agents, or reach for Vercel Workflows instead.
• You're already routing traffic through Cloudflare, or want an edge-first proxy in front of multiple providers: Cloudflare AI Gateway gets you caching, rate limiting, and DLP scanning for free, and Unified Billing collapses provider invoices into one if that matters to your finance team. Budget time to configure Dynamic Routing yourself — it isn't automatic.
• You need the broadest model catalog, are evaluating many models for a specific task, or don't want platform lock-in: OpenRouter's catalog and free tier make it the fastest way to compare models side by side. Model the 5.5%/5% credit fee into your cost projections before committing production spend, and check whether your usage actually clears the 1M-request BYOK threshold.
• You're running a Vercel-hosted frontend and a Cloudflare-fronted API layer at the same time: nothing stops you from using both. Teams increasingly treat gateways as swappable infrastructure rather than a permanent choice, which is part of the point of using one in the first place.

Frequently Asked Questions

Methodology

Pricing and feature details for all three gateways were checked directly against each vendor's own documentation in August 2026: Vercel's AI Gateway docs and pricing reference, Cloudflare's AI Gateway pricing reference, and OpenRouter's quickstart and FAQ pages. Latency figures come from a third-party, publicly published benchmark (see References) that timed DNS, TCP, TLS, time-to-first-byte, and time-to-first-token separately across interleaved requests from the same test machine; the Vercel-vs-OpenRouter figures cited here are explicitly a small sample (n=5, single location) and should be read as directional, not a guarantee of what you'll see on your own infrastructure. Cloudflare's edge-proxy overhead figures come from separate community-reported measurements rather than the same controlled benchmark, since we couldn't find a single test that measured all three gateways identically at the time of writing — a gap worth flagging rather than papering over. We didn't run our own production load test against any of the three; treat every figure here as a documented, source-checked starting point for your own evaluation, not a substitute for testing against your actual traffic pattern.

References & Sources

• Vercel — AI Gateway documentation
• Vercel — AI Gateway pricing reference
• Vercel — Vercel Functions limits (max duration)
• Cloudflare — AI Gateway pricing documentation
• OpenRouter — Frequently asked questions (fees, BYOK, free-tier limits)
• OpenRouter — Quickstart documentation
• Dev Community (smakosh) — OpenRouter vs. Vercel vs. LLMGateway performance benchmark

Ken Hayashi

Technology consultant with 10+ years in the Japanese tech industry. Specializing in SaaS evaluation, workflow automation, and B2B tool integration.