Liquid Inference vs Vercel AI Gateway

Vercel AI Gateway is a hosted gateway. It sells each model at the list price of the provider, with no fee on tokens, and picks the provider by recent uptime and latency. Liquid Inference is an exchange. Providers quote their own prices, and each request starts an auction among the offers that meet your rules.

Both services take the OpenAI and Anthropic formats. The facts about Vercel AI Gateway come from its own pages, checked on 10 November 2026. The links are in Sources.

Feature by feature

Feature	Vercel AI Gateway	Liquid Inference
Price	The provider's list price	Each provider quotes its own price
Provider choice	Recent uptime and latency. You can set `order`, `only` or `sort`	An auction: the lowest charge plus wait cost wins
Fee	None on tokens. You pay the card fees, and some controls cost extra	5% of the charge, or a floor on cheap models. No top-up fee
Cap for each request	No	Yes. `cap_usd`; the provider pays any usage above it
Spend limits	Partly. Soft: the request that crosses a budget completes	Yes. Exact: the cap of each job in progress counts
Preview with no charge	No. A price list for each provider	Yes. The charge and cap of each offer, and why others are out
Record of each request	Yes. `gateway.cost` in each response, and a lookup by id	Yes. A receipt: rates, fee, cap, charge split and the offers left out
No retention by default	Partly. Vercel keeps no prompts, but routing ignores provider retention. Team ZDR costs $0.10 for each 1,000 requests	Yes. Only offers that keep no prompt compete
Open to sellers	No. A fixed list of providers	Yes. By application; sellers post their own offers
Own cloud keys	Yes. No fee on the paid tier. A failure goes again on Vercel's keys and your credit	Partly. Beta. The exchange fee at the list price. You choose market fallback or fail
Media	Yes. Image, video, speech, transcription, embeddings and rerank	Yes. Image, video, speech, transcription and embeddings

What the exchange does that Vercel AI Gateway does not

Prices from competition. Each provider quotes its own price for each model, and the auction buys the lowest cost. See Price and wait.
A hard cap for each request. Send exchange.cap_usd, or the x-liquid-cap-usd header. See Purchasing rules.
Budgets that hold. The exchange holds the price cap of each job before the first token, so a budget is exact. See Guardrails.
A dry run. See which offers compete and what the request can cost before you buy. See Purchasing rules.
Retention rules by default. Only offers that promise not to keep your prompt compete, unless you allow retention. See Data handling.

When Vercel AI Gateway is the better choice

You buy at list price and want no fee on tokens.
Your app runs on Vercel and uses the AI SDK, and you want one bill.
You need a model or a provider that the exchange does not list.

Switch in one line

 client = OpenAI(
-    base_url="https://ai-gateway.vercel.sh/v1",
+    base_url="https://router.inference.ai.exchange/v1",
     api_key=...,
 )

Then change each model to a name that GET /v1/models lists. The AI SDK works with the exchange through its OpenAI-compatible provider. See Vercel AI SDK.

Questions

Is a budget on the exchange exact?

Yes. A budget counts settled charges and the price cap of each job in progress. The ledger enforces it, so concurrent requests cannot pass it. See Guardrails.

What happens when my own key fails?

You choose for each cloud. With market, the default, the exchange moves the job to the next offer before any output, and charges it as a market job. With fail, the job ends. See Bring your own key (BYOK).

Sources

Checked on 10 November 2026.

Product and BYOK: AI Gateway, BYOK
Routing: Provider filtering and ordering
Fees: Pricing, FAQ
Budgets: Budgets
Request record: Logs
Retention: Zero data retention
Models and media: Models and providers