llmrelay vs Vercel AI Gateway

Claude Opus 4.7, per million tokens

Read from the Vercel AI Gateway model page for anthropic/claude-opus-4.7 and the gateway pricing docs on 26 July 2026.

Zero markup is not the same as cheap

This is the whole comparison. "Zero markup" answers the question "is the gateway taking a cut?" It doesn't answer "am I paying less than list?" Vercel's honest answer to the second question is no, and they don't claim otherwise.

Our answer is $2.50 per million input tokens on the same model ID, because we buy upstream capacity in bulk and resell at half of official list rather than passing list through unchanged.

What Vercel bills separately

Tokens are clean. Some surrounding features are metered per request, which matters at volume:

• — Team-wide provider allowlist: $0.10 per 1,000 successful requests
• — Team-wide zero data retention: $0.10 per 1,000 requests
• — Custom reporting: $0.075 per 1,000 writes, plus $5 per 1,000 queries
• — Payment processing fees on credit purchases are yours to absorb

Run 10 million requests a month with allowlist and ZDR on and that's $2,000 in request fees on top of tokens. Small at low volume, real at scale. Note also that Vercel describes its rates as mirroring upstream providers, so they move whenever the provider moves — the $5.00 is a reflection, not a commitment.

Where Vercel AI Gateway is the right choice

• — You already deploy on Vercel and want inference on the same invoice as hosting.
• — You're building with the AI SDK and want the gateway's provider switching to just work.
• — You want built-in observability without wiring your own logging.
• — You have existing enterprise Anthropic keys and want BYOK routing at no token cost.

That last one is a genuinely good deal: if your company already negotiated Anthropic pricing, BYOK through Vercel adds no token markup at all.

Where llmrelay wins

If you don't have negotiated Anthropic pricing — which describes most indie developers and small teams — BYOK gives you nothing and you're on list. Halving that is worth more than SDK ergonomics.

The AI SDK works fine against us, incidentally. It speaks OpenAI-compatible, so you keep the SDK and change the base URL.

Migration

// Before (Vercel AI Gateway)
import { gateway } from '@ai-sdk/gateway';
const model = gateway('anthropic/claude-opus-4.7');

// After (llmrelay, same AI SDK)
import { createOpenAI } from '@ai-sdk/openai';
const llmrelay = createOpenAI({
  baseURL: 'https://api.llmrelay.dev/v1',
  apiKey: process.env.LLMRELAY_API_KEY,
});
const model = llmrelay('claude-opus-4-7');

A note on token counts

Anthropic states that Claude 4.7 and later use a newer tokenizer producing roughly 30% more tokens for the same text. Same rate card everywhere, more tokens per document than 4.6-era estimates. Build cost models on 4.7 tokenization.