Vercel Pricing: Plans, Fluid Compute and AI Costs

Vercel hosts web apps and the serverless functions behind them. For AI products it also sells AI Gateway, which passes model tokens through at the provider's list price. On 31 October 2026 Hobby was $0 for non-commercial use and Pro was $20 a month with $20 of usage credit. Vercel runs your agent's code, not a model on a GPU. Most of the bill is usage, not the plan fee.

Some links on this page go through /go/ and may earn us a fee; it never changes our order.

Facts at a glance

For where Vercel fits against GPU clouds and token APIs, see the AI hosting overview and the interactive hosting compare.

What Vercel is, and is not, for AI

Vercel is a place to run your app, not a model. A typical AI product on Vercel has a Next.js front end, an API route or function that receives a request, and a call to a model API such as one through AI Gateway or a provider you pay directly. The GPU lives elsewhere. If you need a GPU you control, read Runpod, Lambda Cloud or Modal.

The reason this page exists next to GPU hosts is that agent apps are billed by the function time they spend waiting for a model, and Vercel's Fluid compute changes that arithmetic.

Plans

On Pro, the monthly fee includes $20 of usage credit that can be spent across resources. The docs say the plan is billed monthly by the number of deploying team seats, paid add-ons and any on-demand usage. A new team starts with one included paid seat that can deploy and manage the team, and viewers cannot configure or deploy. Beyond the credit you pay as you go, uncapped by default, though you can set a spend amount, which the docs show with a $200 per billing cycle default.

Hobby is hard-capped: its usage limits apply, and you cannot buy extra. A commercial product belongs on Pro.

Usage prices we read

From vercel.com/pricing on 31 October 2026. Rates are "starting at" and vary by region and product, so confirm for your region.

The page also lists data transfer, blob storage, cache reads and writes and other items. Hobby has lower included amounts and no extra purchases.

Fluid compute and why it matters for agents

The pricing page describes dynamic, scalable functions now available with Fluid compute under Active CPU pricing, and the docs say you pay for active CPU time and provisioned memory time. The practical meaning: while your function waits for a model to answer, it is not using CPU, so you pay for the memory held and not for CPU.

For an AI agent, that is the right shape. Most of an agent request is waiting on a model. A function that spends 20 seconds waiting and 200 milliseconds computing is billed for 200 ms of active CPU, plus 20 seconds of provisioned memory.

Fluid compute is on by default for new projects, per the limits page.

What you pay in practice

The volumes are examples we chose, and the numbers use the "starting at" rates above.

Example one: a small AI feature with 200,000 requests a month, 0.3 seconds of active CPU each, and 2 GB of memory held for 15 seconds per request.

• Active CPU: 200,000 × 0.3 s = 60,000 s, about 16.7 hours. After 4 included hours, 12.7 hours × $0.128 is about $1.63.
• Provisioned memory: 200,000 × 15 s × 2 GB = 6,000,000 GB-s, about 1,667 GB-hours. After 360 included, 1,307 GB-hours × $0.0106 is about $13.85.
• Invocations: 200,000 is inside the 1M included.
• Total usage about $15.50, which the $20 credit covers. The month costs the $20 plan fee, assuming one seat and no other usage.

Example two: a heavy agent service with 2 million runs a month, 1 second of active CPU each, 4 GB held for 60 seconds.

• Active CPU: 2,000,000 s, about 556 hours. 552 × $0.128 is about $70.66.
• Memory: 2,000,000 × 60 × 4 = 480,000,000 GB-s, about 133,333 GB-hours. After 360, 132,973 × $0.0106 is about $1,409.
• Invocations: 2M, 1M over, about $0.60.
• Total usage about $1,480 on top of the plan. Memory time dominates, so long waits are what cost money. Trimming wait time, or moving the long-running part to a queue and a worker, pays off.

Example three: model tokens. Through AI Gateway you pay the provider's list price with no markup, so 40 million input and 8 million output tokens on gpt-oss-120b at Groq's $0.15 and $0.60 would be $6.00 plus $4.80, about $10.80, whatever route carries them.

AI Gateway

Vercel's AI Gateway docs say it charges no markup and no platform fee on tokens. You pay the provider's list price pay as you go, with purchased credits, and the paid tier gives zero markup including when you bring your own key. Every team gets a free tier and a paid tier. . Our OpenRouter alternatives page compares gateways, and whether a gateway keeps or drops prompts is covered in its data terms.

Limits

The duration limit matters for agents. A run that goes longer than 800 seconds on Pro needs a different design: a queue and a worker, or a host without that limit. See Render, whose services are not bound by a function timeout in the same way..

Data and regions

Vercel describes a globally distributed platform with compute-capable regions where your code can run close to your data. Pick the region nearest your database and model provider, since cross-region hops add latency. The pages we read do not describe retention of request bodies for functions.

Cutting the bill

Most savings come from the memory line, because waiting costs memory time.

1. Stream responses. Returning tokens as they arrive does not shorten the model call, but it lets the user see progress, so you can set tighter timeouts and abandon stalled calls sooner.
2. Cap model time. Set a timeout on every model call and a maximum output length. A runaway generation holds memory for as long as it runs.
3. Right-size memory. Provisioned memory is billed per GB-hour. A function that asks for 4 GB and uses 300 MB pays for 4 GB. Check the setting for each route.
4. Move long jobs off functions. If a task takes minutes, queue it and run it on a worker, then return a job id. That also avoids the duration limit.
5. Cache repeat answers. A cached response does not run the function. The pricing page notes that a cached function response does not run and does not incur an invocation or any duration.
6. Use one model route. Send traffic through AI Gateway so token spend sits in one place.
7. Watch bots. Public AI endpoints attract scrapers. The firewall lists 1M allowed rate-limit requests a month included, and bot protection has a basic managed ruleset. Put a rate limit on any route that calls a paid model.

Vercel against two common alternatives

The Workers figures come from the platform pricing page: 10 million requests and 30 million CPU milliseconds included on the Paid plan, then $0.30 per extra million requests. Render's instance prices are covered on Render.

Teams and previews

Pro's billing counts deploying seats, and viewers are free but cannot deploy. That matters for agencies and product teams, where designers and reviewers need to look at preview deployments without paying for a seat. The page lists free viewer seats on Pro. A good habit is to keep the deploying seat count to the people who really ship, and to give everyone else viewer access. Previews are useful for agent apps because you can test a prompt change on a real URL before it reaches production. Remember that a preview that calls a paid model also spends real tokens, so use a cheap model and a spend cap there.

When to leave

Signs that Vercel is no longer the cheapest place:

• Memory time is the biggest line, and your runs are long. A flat-priced instance on Render may cost less.
• You need a persistent process, such as a long-lived agent holding state or a websocket server. A normal service suits that better than a function.
• You need a GPU. Vercel does not offer one, so you add a second host regardless.
• Your usage sits so far above Pro's credit that an Enterprise contract or a different platform is cheaper.

Moving is easiest if you keep the model calls behind your own function and the state in a database you control, and if your build does not rely on Vercel-only APIs.

Who should pick Vercel

Alternatives

• Render: long-running services at flat monthly instance prices, from $7.
• Cloudflare Workers: Free plan with 100,000 requests a day, Paid from $5 a month; see Cloudflare Workers AI for models alongside it.
• AWS Bedrock: models inside AWS.
• Modal: serverless Python with GPUs.
• Hugging Face Inference: dedicated model endpoints.

Getting started

1. Connect a Git repository and deploy on Hobby to test. Move to Pro before launching a commercial product.
2. Pick a region near your database and model provider.
3. Keep model calls on a server route, never in the browser, so keys stay secret.
4. Set a spend limit and an alert on usage.
5. Log duration and token counts per route, so the expensive routes are visible.

For the agent build itself, see how to build an AI agent.

Our take

Vercel is a fine host for the app layer of an AI product, and Fluid compute fits agent traffic better than older per-millisecond models. It is not an AI host in the GPU sense, so judge it against Render and Cloudflare Workers, not against Runpod. The cost trap is long waits: memory time grows with every second a function holds a request, as our heavy example shows. Check the 800-second limit before you design long agent runs on it. These notes rest on Vercel's pricing page and docs, and we ran no deploys for this page.

Related pages

• Hosting hub and AI hosting overview
• Private AI hub
• How to build an AI agent