AI Stacks · Pricing checked November 21, 2026

LLM Gateways and Model Routing Compared: OpenRouter vs Bedrock vs Vertex vs Groq vs Replicate

One API key for many models: platform fees, free-tier limits and billing models for seven gateways and inference hosts, from their own pages on November 21, 2026, and why our studio runs on one.

Watch: Episode 1, Built This Week Ep1: How We Built an AI Video Studio in 3 Days from 22:49. Jordan explains the Replicate setup behind the Ryz Labs video studio and names fal and Bedrock as alternatives.
Subscribe on YouTubeWe build something real with AI every week

AI price-change tracker and build-cost estimator: every dated price change from the vendors' own pages, starting with Gemini 4 Carbon doubling on December 1, 2026.

The short answer

If you want every major model behind one key and one bill, use OpenRouter (5.5% fee on credits) or Vercel AI Gateway (list price, zero markup, $5 a month free). If you're in AWS or Google Cloud for compliance reasons, use Amazon Bedrock or Vertex AI. For raw speed on open models, Groq is the fastest host we've used. For image and video models, Replicate and fal let you add a new model in minutes. That's how our internal video studio works.

Pricing and limits (checked November 21, 2026)

Gateways charge in three ways: a fee on top of list price (OpenRouter), list price with no markup (Vercel, Bedrock and Vertex pass through each model's price) or per hardware-second or output (Replicate, fal). Model prices themselves are on our LLM API pricing page.

OpenRouter

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free$025+ free modelsFree models: 20 requests/min and 50/day—
Pay as you go5.5% fee on credit purchases600+ models at provider list priceNo plan rate limit on paid modelsBYOK: first $25,000/mo of inference free of fee, then 5%
Business8% fee on credit purchases600+ models—Same BYOK terms
EnterpriseFee discounts available———

Source: OpenRouter pricing, checked November 21, 2026.

Vercel AI Gateway

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free tier$5/month included creditFree-tier models onlyLower rate limits; no BYOKProvider list rates, zero markup
Paid tierPay as you go (purchased credits)All available modelsBYOK availableProvider list rates, zero markup; payment processing fees may apply

Source: Vercel AI Gateway pricing, checked November 21, 2026.

Amazon Bedrock

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Pricing modesOn-demand, batch, provisioned throughputPer-token on-demand; provisioned billed hourly per model unit—Batch 50% below on-demand for select models
Example: Gemma 4 31B (on-demand)$0.14 in / $0.40 out per 1M tokens—US East and US West regions—

Source: Amazon Bedrock pricing, checked November 21, 2026.

Google Vertex AI

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Pricing modesStandard, Priority, Batch; global vs regional endpointsCache read and write priced separatelyRegional about 10% above global in listed rowsGemini batch 50% off
Example: Gemini 4 Carbon (global)$2 in / $10 out per 1M through Nov 30, 2026Cache read $0.10$4 input from Dec 1, 2026—

Source: Google Vertex AI pricing, checked November 21, 2026.

Groq

Plan / modelPriceIncludedDocumented limits
openai/gpt-oss-120b$0.15 in / $0.60 out per 1M131K context; about 500 tokens/secFree plan: 30 requests/min, 1K/day
openai/gpt-oss-20b$0.075 in / $0.30 out131K context; about 1,000 tokens/secFree plan: 30 requests/min, 1K/day

Source: Groq pricing, checked November 21, 2026. groq.com/pricing now redirects to the homepage; prices from Groq's model docs.

Replicate

Plan / modelPriceDocumented limits
Public modelsPer second of hardware, or per outputPrivate models also bill setup and idle time
Nvidia H100$0.001525/sec ($5.49/hr)—
Example: flux-1.1-pro$0.04 per output image—

Source: Replicate pricing, checked November 21, 2026.

fal

Plan / modelPrice
Model APIsPer output: video per second, images per image or megapixel
Example: nano-banana-2$0.08 per image
Compute: H100 80GB$4.50/hr list, as low as $2.49/hr

Source: fal pricing, checked November 21, 2026.

What broke when we built with it

Replicate: one key for every model, but you still debug each one

The Ryz Labs video studio from our first episode exists because of this idea:

avoid, like, high subscription fees over a large population of people and instead leverage, like, API keys as a consumption layer so that just drives down the cost of the use case, and you just kinda pay for what you use.

— Jordan Metzner, Episode 1, at 04:15

Adding a model was as simple as copying Replicate's sample API call and schema into the Replit agent (at 24:04). But every model behaves differently, and in the live demo some of them failed:

sometimes I have to go back and forth, make some testing… some of these models failed, so I'm gonna have to go back and, you know, kind of understand why they failed.

— Jordan Metzner, Episode 1, at 24:54

The cost model is simple, and so is the catch. As Sam put it: "every time we generate a video, we're paying replicate" (at 25:28).

Groq: fast, but the open model wasn't enough

Jordan demoed gpt-oss-120b on Groq writing ten marketing emails and called it "incredibly fast". The output wasn't perfect, so he re-ran it with a "no dashes" instruction, at about four to five seconds a run (at 19:00). On a heavier job, a finance dashboard reading P&Ls and balance sheets, the open model "wasn't enough" and he moved to Gemini (at 06:59).

Routing: nobody does it well yet

I don't think any any of these LLMs have done a good job of, I would say, like, routing… Amazon has Bedrock which is kind of like this, like, LLM kind of routing platform.

— Jordan Metzner, Episode 23, at 26:47

We have no build of our own with OpenRouter, Vercel AI Gateway, Vertex AI or fal (Jordan named fal only as a Replicate alternative), so they appear in the price table only.

What our guests reported

The healthcare company ARYA reaches Anthropic only through Bedrock and Gemini only through Vertex AI, which "make it possible for us to build HIPAA compliant workflows" (at 11:22). The cost of that choice is waiting on the cloud: "I remember Bedrock, like, initially, they didn't have it. Like, we we pushed them to get a particular version" (at 12:40). The same team saw a bill "about six x" what they expected after a change in how long their Lambda held the Anthropic connection open (at 14:33).

How to choose

  • Prototyping across many LLMs: OpenRouter free models (50 requests a day) or Vercel AI Gateway's $5 monthly credit.
  • Production on one cloud, regulated data: Bedrock or Vertex AI, and check that the exact model version you need is available there.
  • Image and video generation in an app: Replicate (per second or per output) or fal (per output). Put a per-user cap in your app, because every generation is billed.
  • Cheap, fast open models: Groq, where gpt-oss-120b was $0.15 / $0.60 per million tokens.

FAQ

How much does OpenRouter charge?

OpenRouter passes models through at provider list price and charges a 5.5% fee on credit purchases on pay as you go (8% on Business). Bring-your-own-key use is free for the first $25,000 of list-price inference a month, then 5%. Free models are limited to 20 requests a minute and 50 a day (November 21, 2026).

Does Vercel AI Gateway add a markup?

Vercel's documentation says AI Gateway uses provider list rates with zero markup. The free tier includes $5 of credit a month on free-tier models; paid use runs on purchased credits, and payment processing fees may apply.

What is the difference between Replicate and fal?

Both host many image, video and open models behind one API. Replicate bills public models by hardware time per second (an H100 is $0.001525 a second) or per output, for example $0.04 per Flux 1.1 Pro image. fal bills per output, such as $0.08 per Nano Banana 2 image or per second of video.

Why use Bedrock or Vertex AI instead of calling the model directly?

Usually for compliance and procurement. A healthcare guest on our show said they only reach Anthropic through Bedrock and Gemini through Vertex AI, because their cloud contracts let them build HIPAA-compliant workflows. The trade-off they mentioned was waiting for new model versions to appear on Bedrock.

Is Groq the same as Grok?

No. Groq is an inference company that hosts open models such as gpt-oss on its own chips; Grok is xAI's model family. On November 21, 2026, Groq listed gpt-oss-120b at $0.15 per million input tokens and $0.60 output.

Sources