Published Updated
In one or two sentences, before anything else.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
What it can do, one sentence each, and whether we actually tested it.
One API key and one credit balance for 482 models from 49 makers (its public model list on October 21, 2026). Vercel says it adds zero markup to provider token prices, including with your own keys.
Vercel's claim. Its listed prices for Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Sol and Gemini 4 Argon matched the makers' price pages on October 21, 2026.
Each Vercel team gets 5 USD of credit a month, starting with the first request, for models marked as free-tier eligible.
Per Vercel's FAQ: a payment method is required, rate limits are lower, BYOK is not allowed, and the free credit stops for good once you buy credits. Forum users report many models refused 'due to abuse' since about May 2026.
Fails a model over to its other providers, lets you order or restrict providers per request, supports model fallbacks, routing rules and 'virtual models' that carry these settings for coding agents.
Per Vercel's docs. Zack Proser counted fewer providers per open-weight model than on OpenRouter (8 against 17 for gpt-oss-120b, August 2026).
`npx vercel@latest ai-gateway setup` configures 29 agents, including Claude Code, Codex, Cursor, Cline, Kilo Code, OpenCode and Zed, and stores the key in the macOS Keychain.
Per Vercel's coding agents docs. Claude Code, Codex and Cursor get dedicated endpoints.
OpenAI Chat Completions and Responses at ai-gateway.vercel.sh/v1, and Anthropic Messages with tool use, extended thinking, cache_control passthrough, token counting and message batches.
Per Vercel's docs; Anthropic message batches are in early access and need to be enabled for your team.
Perplexity and Parallel search tools can be attached to any model on the gateway through the AI SDK.
Per Vercel's changelog; documented for the AI SDK.
Vercel's official MCP server (mcp.vercel.com) searches its docs and manages projects, deployments and logs; the AI SDK's MCP client passes any MCP server's tools to gateway models; the ai-sdk Agent Skill teaches coding agents the AI SDK.
None of these is specific to AI Gateway; Vercel's MCP tool reference says it does not create AI Gateway API keys.
Vercel says the gateway keeps no prompt or response content after the request. Per-request zero data retention routing, training opt-outs, provider allowlists and regional inference are available.
Several controls are Pro or Enterprise only; team-wide zero data retention costs 0.10 USD per 1,000 requests.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Public tests other people ran: what they gave it and how it did.
Case 1
Input
100 requests each to Claude Haiku 4.5, gpt-5.4-mini and gemini-3.6-flash from Northern Virginia, through Vercel AI Gateway and straight to the providers, on cold and warm connections.
What happened
Median time to first token through Vercel was 462 ms against 394 ms direct on Claude Haiku 4.5 and 503 ms against 412 ms on gpt-5.4-mini, and every request succeeded. One US East location and short prompts, so results elsewhere may differ.
Case 2
Input
npx vercel@latest ai-gateway setup --agent claude-code, or by hand: ANTHROPIC_BASE_URL=https://ai-gateway.vercel.sh/claude-code, ANTHROPIC_AUTH_TOKEN set to the gateway key and ANTHROPIC_API_KEY set to an empty string.
What happened
Claude Code's requests and spend then appear in the gateway's logs, and every gateway model shows in its /model picker. Spend comes out of credits, and budgets only count spend on Vercel's own credentials, not BYOK.
Case 3
Input
His own agent fleet run through Vercel AI Gateway and OpenRouter, comparing catalogs and the cheapest route for gpt-oss-120b and DeepSeek V4 Flash.
What happened
DeepSeek V4 Flash cost the same on both, but gpt-oss-120b's cheapest route was 0.10 / 0.50 USD per million tokens on Vercel against 0.03 / 0.17 USD on OpenRouter. He kept Vercel as default for Vercel-hosted agents. One user's account, with prices as of August.
Case 4
Input
A request using the team's own provider key while the AI Gateway credit balance is 0 USD.
What happened
The request failed with 'Insufficient funds'. A Vercel team member closed the issue as intended behaviour, because a BYOK request that fails falls back to Vercel's own credentials and is billed to credits.
What it does badly, with the evidence, and what we could not settle.
Weak
Billing bugs have blocked paying accounts.
Evidence: Vercel Community threads: 'billing issue (skipping all models)' with 29 USD of credits and no staff reply (April 2026); balance shown as NaN with calls refused (June 2026); 403 on every request after a 10 USD purchase, fixed two days later without a stated cause (July 2026).
Your own keys still need Vercel credits.
Evidence: GitHub issue #11280: BYOK requests fail at a 0 USD balance, closed as intended on August 10, 2026. Vercel's docs say a failed BYOK request can fall back to Vercel's credentials and is billed to credits.
The free tier is narrow.
Evidence: Vercel's FAQ: a card is required, only free-tier eligible models, lower rate limits, no BYOK, and the free credit ends once you buy credits. A May 2026 forum thread reports 403 'restricted due to abuse' errors on many models.
Fewer hosts per open-weight model than OpenRouter.
Evidence: Zack Proser (August 2026): 8 providers for gpt-oss-120b against 17 on OpenRouter, and a cheapest route of 0.10 / 0.50 USD against 0.03 / 0.17 USD per million tokens.
Little support on the Hobby plan.
Evidence: In a July 2026 forum thread, Vercel's help bot said Hobby teams cannot open support cases for AI Gateway billing or access problems.
Not for pure-BYOK setups or teams that need a reliable free tier.
Evidence: BYOK requires a positive credit balance and can be billed to credits on failure, and the free credit is limited to some models and has been restricted. Teams that only want a proxy in front of their own keys have free options such as Cloudflare AI Gateway.
Uncertain
Minimum top-up and refunds.
Why unsure: Vercel's pricing docs, FAQ and changelog state neither a minimum credit purchase nor a refund policy.
Free-tier limits and model list.
Why unsure: The docs say free-tier rate limits are lower and may change, but publish no numbers or fixed list.
Vercel's 20 ms overhead claim.
Why unsure: Vercel gives no method. ComputeSDK measured about 40 to 95 ms over direct calls from US East, and Edgee, a competitor, 164 to 193 ms across 43 regions.
Payment methods other than cards.
Why unsure: The docs describe a payment method on the team and invoicing for Enterprise, and say nothing about PayPal, crypto or local wallets.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | Vercel Inc., San Francisco, US; AI Gateway generally available since August 21, 2025 | Source: vercel.comchecked |
| Price vs list price | Provider list price with no token markup, also with BYOK | On October 21, 2026: Claude Opus 5.5 at 4 / 20 USD, GPT-6 Sol at 2 / 10 USD and Gemini 4 Argon at 2 / 10 USD per million tokens, the same as the makers' prices. Source: vercel.comchecked |
| Platform fee | None on tokens or BYOK; card processing fees are the buyer's; optional extras such as team-wide zero data retention at 0.10 USD per 1,000 requests | Source: vercel.comchecked |
| Free tier | 5 USD of credit a month per Vercel team; card required; free-tier eligible models only; ends once you buy credits | Source: vercel.comchecked |
| Minimum top up, expiry and refunds | No minimum stated; bought credits expire one year after purchase; no refund policy in the docs | Source: vercel.comchecked |
| Payment methods | Payment method on the Vercel team, with optional auto top-up; invoices for Enterprise | Source: vercel.comchecked |
| Models | 482 entries from 49 makers: 323 language, 39 video, 38 image, 28 embedding, 14 speech, 12 transcription and others | Source: ai-gateway.vercel.shchecked |
| API formats | OpenAI Chat Completions, OpenAI Responses, Anthropic Messages (with token counting and batches in early access); tool calling, structured outputs, prompt caching | Source: vercel.comchecked |
| MCP | Official Vercel MCP server (mcp.vercel.com) for docs, projects, deployments and logs, with no AI Gateway tools; AI SDK MCP client passes MCP tools to gateway models | Source: vercel.comchecked |
| Agent Skills | Official skills on skills.sh (ai-sdk, vercel-cli, vercel-deploy and others) and a Vercel plugin; none specific to AI Gateway; Anthropic Skills API not documented | Source: vercel.comchecked |
| Coding agents | One-command setup for 29 agents, including Claude Code, Codex, Cursor, Cline, Kilo Code, OpenCode and Zed; manual guide for Roo Code | Source: vercel.comchecked |
| SDKs | AI SDK for TypeScript and Python; guides for LangChain, LiteLLM and Pydantic AI; llms.txt | Source: vercel.comchecked |
| Privacy and retention | No prompt or response content kept after the request; metadata logged, routing details kept 30 days; per-request zero data retention | Source: vercel.comchecked |
| Status | AI Gateway component on vercel-status.com; one AI Gateway incident in the last 50 (July 15, 2026, 30 minutes) | Source: vercel-status.comchecked |
Answer first, then why, an example, the limits and what to do next.
Vercel AI Gateway includes 5 USD of free credit a month per Vercel team, but only for some models. You need a payment method on file, rate limits are lower, your own keys are not allowed, and the free credit stops for good once you buy credits.
Limits
Since about May 2026, forum users report 403 errors saying free credits are restricted 'due to abuse' on many models.
Vercel AI Gateway costs the same as the official APIs for the models checked, not less. On October 21, 2026 it listed Claude Opus 5.5 at 4 / 20 USD, GPT-6 Sol at 2 / 10 USD and Gemini 4 Argon at 2 / 10 USD per million tokens, the makers' own prices.
Why
Vercel says it adds no markup; you pay card processing fees, and some models run below-list promotions.
Vercel AI Gateway is usually cheaper on the same route, because it adds no fee while OpenRouter adds 5.5% when you buy credits. OpenRouter can still win on open-weight models: Zack Proser found gpt-oss-120b's cheapest route at 0.03 / 0.17 USD per million tokens on OpenRouter against 0.10 / 0.50 USD on Vercel in August 2026.
Why
OpenRouter had 17 providers for that model and Vercel 8.
Vercel AI Gateway is run by Vercel Inc., a large US company, and its docs say how prompts are handled. Vercel says it does not train on prompts and keeps no prompt or response content after the request; zero data retention routing is available per request.
Limits
The practical risk is billing: forum users report paid credits that stopped working for days, and Hobby teams say they could not open a support case.
Vercel AI Gateway has no MCP server of its own, but Vercel publishes an official MCP server for its platform and an MCP client in the AI SDK. The server at mcp.vercel.com searches docs and manages projects, deployments and logs; the AI SDK's MCP client passes any MCP server's tools to models on the gateway.
Limits
Passing remote MCP servers inside an API request is not documented, and message batches do not support MCP servers.
Vercel AI Gateway accepts your own provider keys with no fee, but only while you have bought credits. A BYOK request fails at a 0 USD balance, and a BYOK request that fails can fall back to Vercel's credentials and be billed to your credits.
Why
Vercel closed GitHub issue #11280 about this as intended behaviour on August 10, 2026.
This page is built from Vercel's AI Gateway pages, docs, public model list and status feed, the official price pages of Anthropic, OpenAI and Google, three public latency and price comparisons, and user reports on the Vercel Community forum and GitHub, all checked on October 21, 2026. Vercel's claims are labelled as its own.
Vercel AI Gateway is one API and one prepaid credit balance for models from many makers. Vercel's landing page says 400+ models; its public model list returned 482 entries on October 21, 2026, from 49 makers: 323 language models, 39 video, 38 image, 28 embedding, 14 speech and 12 transcription models, among others, from Claude, GPT and Gemini to Veo, Seedance, Kling and Flux.
You call it with an AI Gateway API key from any server, or with Vercel's OIDC tokens on Vercel deployments; your app does not have to run on Vercel. The gateway fails a model over to its other providers, lets you order or restrict providers per request, supports model fallbacks, and logs each request's provider, latency, tokens and cost. It became generally available on August 21, 2025.
Vercel says the gateway adds zero markup to provider token prices, including when you bring your own keys. For the models checked, its prices matched the makers':
| Model, per million input / output tokens | Vercel AI Gateway | Official price | Official source |
|---|---|---|---|
| Claude Opus 5.5 | 4 / 20 USD | 4 / 20 USD | Anthropic |
| Claude Sonnet 5.5 | 2 / 10 USD | 2 / 10 USD | Anthropic |
| GPT-6 Sol | 2 / 10 USD | 2 / 10 USD | OpenAI |
| Gemini 4 Argon | 2 / 10 USD | 2 / 10 USD until November 30, 2026; 4 / 20 USD from December 1 | |
| Gemini 3.8 Flash | 0.75 / 3.75 USD | 0.75 / 3.75 USD until December 31, 2026 |
Vercel's prices come from its public model list. Card processing fees are yours, some extras are billed from credits (team-wide zero data retention costs 0.10 USD per 1,000 requests), and bought credits expire after one year. Vercel's landing page showed 26 below-list promotions on October 21; one, Gemini 3.8 Flash at half price, equals Google's own price until December 31.
Complaints on the Vercel Community forum are about billing state. In April 2026 a user with 29 USD of credits got "billing issue (skipping all models)" on every request, with no staff reply. In June 2026 another saw the balance shown as NaN and calls refused. In July 2026 a user got 403 on every request after buying 10 USD of credits; the help bot said Hobby teams cannot open AI Gateway billing cases, and staff fixed it two days later without naming a cause. A May 2026 thread reports free credits refused "due to abuse" on many models.
GitHub issue #11280 on vercel/ai reports that requests with a team's own provider keys fail with "Insufficient funds" at a 0 USD balance. A Vercel team member closed it on August 10, 2026 as intended, because a failed BYOK request falls back to Vercel's credentials and is billed to credits.
Vercel's official MCP server at mcp.vercel.com searches its docs, manages projects and deployments, and reads logs, with setup for Claude Code, Codex CLI, Cursor, VS Code, OpenCode and others. Its tool reference has no AI Gateway category and says it does not create AI Gateway API keys. For MCP in your own app, Vercel documents the AI SDK's MCP client (@ai-sdk/mcp), which passes any MCP server's tools to gateway models. Passing remote MCP servers inside an API request is not documented, and message batches do not support them.
Vercel's skills.sh directory lists official Agent Skills such as ai-sdk, vercel-cli and vercel-deploy, plus a Vercel plugin for coding agents. None is specific to AI Gateway, and the Anthropic Messages docs do not mention Anthropic's Skills API.
npx vercel@latest ai-gateway setup detects installed agents, shows a diff before writing, stores the key in the macOS Keychain and configures 29 of them, including Claude Code, Codex, Cursor, Cline, Kilo Code, OpenCode and Zed; Roo Code has a manual guide. The API speaks OpenAI Chat Completions, OpenAI Responses (tools, structured outputs, reasoning, prompt caching) and Anthropic Messages (tools, extended thinking, cache_control, token counting, batches in early access). Perplexity and Parallel web search work with any model through the AI SDK. There are AI SDK packages for TypeScript and Python, guides for LangChain, LiteLLM and Pydantic AI, and an llms.txt.
Vercel Inc. of San Francisco raised a 300 million USD Series F at a 9.3 billion USD valuation on September 30, 2025. Vercel says it does not train on prompts and the gateway keeps no prompt or response content after the request; it logs metadata and keeps routing details for 30 days. Zero data retention routing, provider allowlists and regional inference are available, several only on Pro or Enterprise. In the status feed's last 50 incidents, from May 23 to October 12, 2026, one touched AI Gateway: a 30-minute error for teams with 2FA on July 15. The docs state no minimum top-up and no refund policy.
Skip it if you want to run only on your own provider keys without prepaying Vercel, need a free tier that reliably reaches flagship models, or are on the Hobby plan and cannot wait days for a billing fix. Teams shopping between many hosts of open-weight models will find fewer providers per model than on OpenRouter. For teams already on Vercel or the AI SDK, it is one of the few relays that adds nothing to list prices.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
ZenMux is an LLM gateway, operated under its terms by NexaMind Singapore Pte. Ltd., that sells Claude, GPT, Gemini and many Chinese models through OpenAI, Anthropic and Vertex-style endpoints at the makers' list token prices. It adds automatic credits for bad or slow answers, which it calls insurance, and 20 to 200 USD monthly coding plans.
WaveSpeedAI is an API platform, run by a Singapore company and a Hong Kong company named in its terms, that sells 1,000+ image, video, audio and 3D models plus an LLM service on one balance. It matches fal on Kling 3.0, charges above Google's list for Veo 3.1 Fast and accepts Alipay and WeChat Pay.
Requesty is a London-based LLM gateway, run by Requesty Ltd and seed-funded by 20VC in 2025, that sells OpenAI- and Anthropic-compatible access to more than 600 models at list price plus a 5% top-up fee. It stands out for routing policies, spend controls and an EU endpoint in Frankfurt.
Replicate is an API for running AI models, owned by Cloudflare since a $57.4 million acquisition closed on December 1, 2025. It sells 100+ official models such as Veo 3.1, Nano Banana Pro and Kling 3.0 per output, and 50,000+ community models by the second of GPU time.
Poe API is Quora's developer API for the bots in its Poe chat app, opened on July 31, 2025, with OpenAI-compatible endpoints and an Anthropic-compatible one for official Claude bots. Calls spend the same points as a Poe subscription, so there is no separate API bill; current Claude bots cost about 1% more than Anthropic's list price.