Vercel AI Gateway Review (2026)
No markup on tokens and $5 of free credits a month per team. The details worth knowing are that BYOK is paid-tier only, and zero data retention is a metered add-on at $0.10 per 1,000 requests.
Rating
Starting Price
$0 plus provider list price
Free Plan
Yes
SDKs & Frameworks
3
Deployment
3
Best For
Teams already deploying on Vercel who want provider-agnostic model access at list price with no markup, and who do not need governance features that carry a per-request charge.
Last Updated:
10 Things You Should Know About Vercel AI Gateway
- 1 AI Gateway charges no markup and no platform fee on tokens - you pay provider list price
- 2 Every Vercel team account receives $5 per month in included AI Gateway Credits from first use, refreshing every 30 days
- 3 Charges are deducted from prepaid AI Gateway Credits, with auto top-up available
- 4 Platform seats are free on Hobby, with Pro starting at $20 per developer seat per month
- 5 BYOK works for any provider in the catalogue with no markup, but is paid-tier only
- 6 Failed BYOK requests fall back to system credentials and are charged against your balance
- 7 Team-wide zero data retention and provider allowlists each cost $0.10 per 1,000 successful requests
- 8 Credit purchases can include payment-processing fees
Pros & Cons
Pros
- ✓ No markup on tokens at all, which is the decisive structural difference against OpenRouter's percentage fee
- ✓ $5 a month in free credits refreshing every 30 days is a genuine standing allowance rather than a one-off trial
- ✓ Platform seats are free on Hobby, so a solo developer pays nothing for the gateway itself
- ✓ BYOK carries no markup either, so bringing your own provider keys costs nothing extra
- ✓ Natural fit if your application already deploys on Vercel, with no new vendor relationship
Cons
- ✕ BYOK is paid-tier only, so the cheapest path to using your own keys requires a $20 per seat plan
- ✕ Failed BYOK requests fall back to system credentials and are charged against your credit balance, which is easy to miss
- ✕ Zero data retention is a metered add-on at $0.10 per 1,000 successful requests rather than a default
- ✕ Provider allowlists are metered at the same rate, so basic governance carries a per-request cost
- ✕ Credit purchases can carry payment-processing fees on top
- ✕ No evaluation or meaningful LLM observability - it is a routing and billing layer
Features
Zero markup, which is the thing that matters
AI Gateway charges no markup and no platform fee on tokens. You pay the provider’s list price, drawn from prepaid AI Gateway Credits.
That puts Vercel in the right half of this category structurally, alongside Cloudflare AI Gateway and self-hosted LiteLLM, and against OpenRouter, which takes 5.5% of everything you spend with no volume discounts.
The difference compounds indefinitely. At $50,000 a month of provider spend, that percentage is $2,750 every month for the same service.
The free allowance is also genuinely standing rather than a trial: $5 per month in AI Gateway Credits per team account, refreshing every 30 days, with platform seats free on Hobby. A solo developer pays nothing for the gateway itself.
The BYOK detail worth internalising
Failed BYOK requests fall back to system credentials and are charged against your credit balance.
Bring-your-own-key works for any provider in the catalogue with no markup, which is good - but it is paid-tier only, requiring the $20 per developer seat Pro plan.
And when a BYOK request fails, the gateway does not simply error. It retries using Vercel’s own credentials and bills you.
In normal operation that is a helpful reliability feature. If your key is misconfigured, rate-limited or expired, it means traffic you believed was hitting your own provider account is silently drawing down prepaid credits instead.
Alert on BYOK failure rate. This is a discover-it-on-the-invoice problem otherwise.
Privacy and governance are metered
This is the genuinely unusual pricing decision here.
| Add-on | Price |
|---|---|
| Team-wide zero data retention | $0.10 per 1,000 successful requests |
| Provider allowlists | $0.10 per 1,000 successful requests |
Most competitors treat data retention policy as a configuration setting, not a metered feature.
At a million requests a month that is $100 for ZDR and another $100 for allowlists - not large in absolute terms, but an odd shape. You are paying per request for a privacy guarantee and for basic governance, and the cost scales with traffic rather than sitting as a flat platform fee.
If either is a compliance requirement rather than a preference, model it explicitly. Credit purchases can also carry payment-processing fees on top.
A correction to something you may have read
Worth stating plainly, because we hit it while researching this page.
One comparison groups LLM Gateway, OpenRouter, Vercel, Cloudflare, Eden AI, Portkey and LiteLLM together as passing provider token rates through without per-token markup.
That is accurate for most of that list and wrong for OpenRouter, which charges a 5.5% credit fee on all purchases with no volume discounts, dropping to 0.5% in BYOK mode. We verified that against OpenRouter’s own documentation.
Grouped claims in comparison tables tend to flatten exactly the distinction you are trying to draw. This is the fourth time on this site a third-party comparison has contradicted a directly verified fact - the others being Latitude’s licence, Vellum’s status, and Phoenix-to-AX migration.
Scope
A routing and billing layer, not an evaluation or observability platform.
You get model access across hundreds of models, credit management and the governance add-ons. You will still want tracing from Langfuse or Logfire and an evaluation framework alongside - the same architecture LiteLLM, Cloudflare and Kong all imply.
Vercel or Cloudflare?
Very similar economics, different platform gravity.
- Cloudflare - core features free on every plan, 100,000 logs free and $5 for a million, runs at the edge, unified billing across OpenAI, Anthropic and Google added in 2026.
- Vercel - $5/month credits, free Hobby seats, BYOK with no markup on paid tiers.
If your application deploys on Vercel, use Vercel. If you are on Cloudflare, use Cloudflare. Neither is worth adopting the underlying platform for on its own, and both beat OpenRouter on economics once you are past prototyping.
Should you use it?
Use Vercel AI Gateway if you already deploy on Vercel, you want provider-agnostic access at list price, and you do not need ZDR or allowlists at high volume.
Don’t use it if you need BYOK without a paid seat, or governance features that carry a per-request charge are a compliance requirement at scale.
Bottom line: correct economics on the part that scales, with three cost details that are easy to miss. Watch BYOK failures, model the metered add-ons, and it is a strong default for teams already on the platform.
Pricing model, credit allowance, BYOK behaviour and add-on rates verified against Vercel’s documentation and third-party analyses on 3 August 2026. A third-party claim that OpenRouter charges no markup conflicts with OpenRouter’s own documentation and has been corrected here. This is a researched directory entry - we have not yet instrumented this gateway with our reference application.
Pricing Plans
Hobby
$0
- Platform seats free
- $5 per month in AI Gateway Credits, refreshing every 30 days
- No markup and no platform fee on tokens
Pro
$20 per developer seat/mo
- Usage-based add-ons
- BYOK available (paid tier only)
- Same zero markup on tokens
Add-ons
Metered
- Team-wide zero data retention at $0.10 per 1,000 successful requests
- Provider allowlists at $0.10 per 1,000 successful requests
SDKs & Frameworks
Deployment
Eval Methods
Billing Unit
Our Verdict
Vercel AI Gateway takes no markup on tokens, which puts it in the right half of this category immediately. You pay provider list price from prepaid credits, every team gets $5 a month in credits that refresh every 30 days, and platform seats are free on Hobby - so a solo developer pays nothing for the gateway itself. Against OpenRouter's 5.5% with no volume discounts, that difference compounds indefinitely. Three details deserve attention before you commit. BYOK is paid-tier only, so using your own provider keys requires the $20 per seat plan. Failed BYOK requests fall back to system credentials and are charged against your balance, which is the kind of thing you discover on an invoice rather than in documentation. And zero data retention is not a default - it is a team-wide add-on metered at $0.10 per 1,000 successful requests, with provider allowlists priced the same way. Charging per request for privacy and for basic governance is an unusual choice and worth modelling if either matters to you.
Similar Tools
Kong AI Gateway
Organisations already standardised on Kong for API management, where bringing LLM traffic under existing governance matters more than LLM-specific features.
Martian
Nobody choosing a gateway today. The RouterBench work remains worth reading if you are evaluating routing approaches generally.
LiteLLM
Teams with engineering capacity that want provider-agnostic routing with no per-token markup, particularly at volumes where a percentage fee on a managed gateway becomes significant.
Cloudflare AI Gateway
Teams already on Cloudflare who want caching, rate limiting and cost visibility at the edge with no markup, and who have separate tooling for evaluation.
Frequently Asked Questions
Is it genuinely zero markup?
Yes on tokens, which is the part that scales. AI Gateway charges no markup and no platform fee - you pay the provider's list price, drawn from prepaid AI Gateway Credits. That is the structurally correct model and it puts Vercel alongside Cloudflare AI Gateway and self-hosted LiteLLM rather than alongside OpenRouter, which takes 5.5% of everything you spend with no volume discounts. At $50,000 a month of provider spend that difference is $2,750 every month. Where Vercel does charge is elsewhere - seats on Pro, and metered add-ons for zero data retention and provider allowlists - so model those separately rather than assuming zero markup means zero cost.
What is the BYOK trap?
Failed BYOK requests fall back to system credentials and are charged against your credit balance. That is the detail to internalise. Bring-your-own-key works for any provider in the catalogue with no markup, which is good, but it is paid-tier only, and when a BYOK request fails the gateway does not simply error - it retries using Vercel's own credentials and bills you for it. In normal operation that is a helpful reliability feature. If your key is misconfigured, rate-limited or expired, it means traffic you believed was going through your own provider account is silently drawing down prepaid credits instead. Alert on BYOK failure rate rather than discovering it on an invoice.
Why is zero data retention a paid add-on?
We do not know the reasoning, but the pricing is unusual enough to flag. Team-wide zero data retention costs $0.10 per 1,000 successful requests, as do provider allowlists. Most competitors treat data retention policy as a configuration setting rather than a metered feature. At a million requests a month that is $100 for ZDR and another $100 for allowlists, which is not large in absolute terms but is an odd shape - you are paying per request for a privacy guarantee and for basic governance. If either is a compliance requirement rather than a preference, model it explicitly, because it scales with traffic rather than being a flat platform fee.
How does it compare with Cloudflare AI Gateway?
Very similar economics, different platform gravity. Both take no markup on tokens, both have generous free allowances, and both are best if you already run the platform. Cloudflare gives core features free on every plan with 100,000 logs a month and $5 for a million, runs at the edge, and added unified billing across OpenAI, Anthropic and Google in 2026. Vercel gives $5 a month in credits, free Hobby seats, and BYOK with no markup on paid tiers. If your application deploys on Vercel, use Vercel. If you are on Cloudflare, use Cloudflare. Neither is worth adopting the underlying platform for on its own.
Does it replace an observability tool?
No. It is a routing and billing layer, not an evaluation or observability platform. You get model access, credit management and the governance add-ons, and you will still want tracing from Langfuse or Logfire and an evaluation framework alongside. That is the same architecture LiteLLM, Cloudflare AI Gateway and Kong all imply, and it is a reasonable separation rather than a gap - a gateway that also tried to be an eval platform would likely be mediocre at both.
I read that OpenRouter also has no markup. Is that right?
No, and this is worth correcting because we encountered the claim while researching this page. One comparison groups LLM Gateway, OpenRouter, Vercel, Cloudflare, Eden AI, Portkey and LiteLLM together as passing provider token rates through without per-token markup. That is accurate for most of that list and wrong for OpenRouter, which charges a 5.5% credit fee on all purchases with no volume discounts, dropping to 0.5% in BYOK mode. We verified OpenRouter's fee against its own documentation. Treat grouped claims in comparison tables with care - they tend to flatten exactly the distinction you are trying to make.