Vercel AI Gateway vs OpenRouter: which to pick in 2026

Updated August 15, 2026. We revisit this comparison as the market moves.

Vercel AI Gateway and OpenRouter answer the same question, how to reach many models through one key without paying a markup on tokens, and they answer it from opposite directions. Vercel folds a gateway into the deployment platform your frontend may already live on, OpenRouter runs the largest independent model marketplace. Because both pass provider list prices straight through, the real differences sit in catalogs, retention guarantees, plan gates and where your data is processed, so that is what this guide compares, with every claim linked to the vendor's own pages as verified on August 11, 2026. One disclosure before we start: we build Opper, an AI gateway in the same category, and it appears below as a clearly labeled third column rather than a hidden thumb on the scale.

Vercel AI Gateway vs OpenRouter at a glance

All figures are vendor-stated and current, sources linked throughout the article.

How is pricing different when neither adds a markup?

Both vendors say the same thing about tokens. Vercel's pricing page states that AI Gateway "charges no markup and no platform fee on tokens", you prepay credits and spend them at the provider's list price, and OpenRouter's FAQ says it passes through underlying provider pricing "without any markup". The difference is the on-ramp. OpenRouter charges 5.5% with a $0.80 minimum on card credit purchases, or 5% for crypto. Vercel charges no platform fee but notes you are "responsible for any payment processing fees that may apply" on top-ups, and only enterprise teams paying by invoice avoid those entirely.

Bring-your-own-keys is where the models diverge more. Vercel's BYOK terms carry no markup or fee at all, but BYOK requires the paid tier, and failed BYOK requests are retried on Vercel's own credentials and charged to your balance. OpenRouter prices BYOK with a monthly allowance: it is free up to $25,000 per month of list-price usage on pay-as-you-go ($200,000 on enterprise), with 5% of the equivalent OpenRouter cost deducted from credits above that.

Free tiers follow the same pattern of generosity with strings attached. Every Vercel team gets $5 of credits every 30 days, which the current pricing doc restricts to a subset of models with per-model rate limits below the paid tier, and the monthly free credit no longer applies once you purchase credits. OpenRouter's free usage runs through its free model variants instead, capped at 50 requests per day until you have bought $10 of credits, 1,000 per day after, at 20 requests per minute. Opper follows the same no-markup principle with a fee on credit purchases, and since fee schedules change faster than blog posts, the live numbers stay on the pricing page.

Who has more models, and who routes them better?

OpenRouter publishes concrete numbers on its homepage, 400+ models from 70+ providers, and that catalog plus its ecosystem remains the strongest reason to choose it. Vercel says "hundreds of models" across text, image and video plus embeddings, without publishing an exact total. For reference, Opper routes 700+ models including image, voice and video models, and you can put any two head to head on the compare pages.

Routing philosophies differ in ways that matter under load. Vercel's gateway automatically retries requests on other providers when one fails, and you can configure model fallbacks on top. OpenRouter goes further by default, load-balancing across providers with price as the priority, deprioritizing providers with recent outages, with fallbacks on by default and fine-grained controls on top, sort by price, throughput or latency, :nitro and :floor suffixes, or an explicit provider order. If routing control is your main criterion, OpenRouter's surface is the most configurable of the three.

Observability is the reverse trade. OpenRouter returns detailed usage information with every response, native-tokenizer token counts and cost in credits, with a /generation endpoint for async retrieval. Vercel includes usage, latency and spend monitoring in its dashboard, but deeper export is metered: Custom Reporting bills per write and per query, and Trace Drains, a Pro and Enterprise feature, bill on delivered traces and trace data volume.

What happens to your prompts?

Vercel has the strongest default retention story on paper. Its ZDR page states the gateway "does not retain prompts, outputs, or sensitive data" and that user data is "immediately and permanently deleted after requests are completed". Read the fine print though, because gateway retention and provider retention are different things. Enforcement that keeps your traffic on providers with zero-retention agreements is Pro and Enterprise only, free per request, $0.10 per 1,000 requests team-wide, and model-level exceptions exist: Vercel's own page notes that anthropic/claude-fable-5 supports ZDR on no provider, with prompts retained for 30 days and not used for training. BYOK keys are skipped under ZDR unless you mark them compliant yourself.

OpenRouter's posture is transparency rather than enforcement. Its privacy policy states it does not use your inputs or outputs for model training and that media files are not persisted beyond what routing requires, with carve-outs for abuse detection, security, billing and legal compliance. Downstream, provider data policies vary widely, and OpenRouter displays them but does not route on them by default, its docs say plainly that routing rules do not change based on provider retention policies, though you can disallow providers that train on prompts account-wide and filter by data policy per request. The same policy notes personal data may be transferred to servers in the US or other countries outside the European Economic Area.

On residency the positions are clear cut. OpenRouter now documents a concrete EU option: enterprise customers can request in-region routing via eu.openrouter.ai, where prompts and completions are processed within the European Union, enterprise status and an explicit request required. Vercel publishes no EU hosting option anywhere in its security and compliance section, its controls are provider-level rather than geographic. Keep one distinction in mind with any gateway: where the gateway processes a request and where the model actually runs are separate questions, so residency needs both the gateway leg and the inference leg pinned. That combination is Opper's home turf, the platform runs in AWS Stockholm with EU routes end to end and metadata-only retention by default, documented with a public sub-processor list on the compliance page, and the wider European field is covered in our European AI gateways guide.

When should you pick Vercel AI Gateway?

Pick Vercel when your team already deploys there. The gateway authenticates with API keys or Vercel OIDC tokens, credits and monitoring live in the dashboard you already use, and the default zero-retention posture is genuinely good. One myth to retire: it is not locked to Next.js or the AI SDK, the docs list OpenAI Chat Completions, OpenAI Responses and Anthropic Messages surfaces at ai-gateway.vercel.sh/v1, callable from any language. The real anchor is the account and the plan, you need a Vercel team, and ZDR enforcement, provider allowlists and trace export sit behind Pro and Enterprise, so the further you are from the Vercel platform, the less of the product you can reach.

When should you pick OpenRouter?

Pick OpenRouter for breadth and ecosystem. Its 400+ models from 70+ providers arrive with the most granular routing controls in this comparison, nearly every AI tool ships an OpenRouter integration, and it maintains an official provider for the Vercel AI SDK, so choosing OpenRouter does not mean giving up Vercel's tooling. Teams with enterprise contracts get the EU routing endpoint on request. One thing to watch during procurement: OpenRouter is reportedly in acquisition talks with Stripe, unannounced as of this writing. If you are weighing OpenRouter against Opper specifically, we keep a direct head-to-head on the OpenRouter alternative page.

Where does Opper fit as the third option?

Opper is our product, so weigh this section accordingly, but the gap it fills is visible in the table above: European hosting as the default rather than an enterprise negotiation. The gateway runs in AWS Stockholm with EU routes end to end, stores metadata only, no prompts, unless you opt into tracing on the Control Plane plan, and routes 700+ models including image, voice and video. Every call is metered with cost, latency and token accounting, uptime is published on a public status page, and aliases with automatic fallback handle the failover story. Switching is a base-URL change because the gateway speaks the OpenAI, Anthropic and Gemini SDK surfaces on one endpoint:

import os
from openai import OpenAI

client = OpenAI(
  base_url="https://api.opper.ai/v3/compat",
  api_key=os.environ["OPPER_API_KEY"],
)

For how the whole category shakes out beyond these three, our ten-gateway comparison covers the full field.

Vercel AI Gateway vs OpenRouter FAQ

Pick the gateway that matches where you run

If your app deploys on Vercel and Pro-tier controls cover your compliance needs, Vercel AI Gateway is a clean choice, and if catalog breadth with configurable routing matters most, OpenRouter has earned its ecosystem. If the deciding factor is European hosting with no markup on inference and observability built in, create a free Opper account and make your first call in minutes, browse the 700+ models it routes to, or check the numbers on pricing yourself.