A unified AI model gateway from Vercel. It lets you call models from multiple providers — OpenAI, Anthropic, Google and others — through a single API key and a single endpoint. Its defining trait is that no gateway markup and no platform fee are added on top of token costs: you pay the provider’s list price, and the same applies when you bring your own provider key (BYOK). Since general availability in August 2025 it has broadened across modalities — text, image, video, audio and embeddings — and now supports routing the spend of coding agents such as Claude Code and Codex through the gateway.
Key Features
- One key, hundreds of models: No need to set up an account and key per provider. A single key and endpoint reach a wide catalog of models, and switching providers is often just a matter of changing the model string
- Easy migration from existing SDKs: Works with AI SDK v5 / v6 and is compatible with the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages APIs. In many cases existing code runs after swapping the base URL and key
- No-markup billing and BYOK: Tokens are billed at the provider’s list price. You purchase credits up front and charges are deducted from the balance, with optional auto top-up when the balance falls below a threshold. BYOK is also free of gateway markup
- Automatic failover during outages: If a request to one provider fails, it is automatically retried against another. With BYOK, a failure of your own credentials falls back to system credentials, and that fallback usage is charged against your credit balance
- Observability and spend management: Request counts, latency and per-provider spend are visible in the dashboard. Custom Reporting adds tags and user IDs for aggregation, and Trace Drains forward OpenTelemetry traces to your own monitoring tool
- Security settings for business use: Zero Data Retention (routing only to providers that do not retain or train on prompt data), a provider allowlist, and settings to disallow prompt training can all be managed centrally at the team level
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Free tier | $0 | Free credits limited to a subset of models. Rate limits are lower than the paid tier and return a 429 when exceeded |
| Pay-as-you-go | Credit purchase (any amount) | Usage billed at model list price, with no markup or platform fee. BYOK available, with higher rate limits |
| Enterprise | Contact sales | Invoiced billing (no payment processing fees) and volume discounts |
Main add-on charges (off by default):
| Add-on | Price |
|---|---|
| Custom Reporting (writes) | $0.075 per 1,000 writes |
| Custom Reporting (queries) | $5 per 1,000 queries |
| Team-wide provider allowlist | $0.10 per 1,000 successful requests |
| Team-wide zero data retention | $0.10 per 1,000 requests |
Note that per-request provider filtering (the only parameter) and per-request ZDR carry no additional cost. On pay-as-you-go, payment processing fees are the customer’s responsibility.
Pricing is as of August 2026. Check the official site for the latest pricing.
Pros & Cons
✅ Pros
- Because there is no markup on tokens, adding a middle layer across multiple providers does not increase cost
- Frees you from per-provider key and billing management, with spend visible in one dashboard
- Existing OpenAI / Anthropic SDK code carries over almost as is, so switching cost is low
- Automatic failover reduces the risk of depending on a single provider
- Requirements common in business use, such as ZDR and provider allowlists, are met through settings rather than code
⚠️ Cons
- The free tier is limited in available models and rate limits, so serious use effectively requires purchasing credits up front
- The prepaid balance model means you need auto top-up or monitoring to avoid interruptions when credits run out
- Advanced capabilities (Custom Reporting, team-wide ZDR and allowlist, Trace Drains) are billed separately and scale with request volume
- Placing a gateway in the path introduces a component that can itself become a single point of failure
- Availability timing for the newest models and support for niche parameters may differ from connecting to a provider directly
Comparison with Similar Services
| Criteria | Vercel AI Gateway | OpenRouter | Cloudflare AI Gateway | Direct provider connection |
|---|---|---|---|---|
| Provider | Vercel | OpenRouter | Cloudflare | Each vendor |
| Billing | Prepaid credits, no markup on tokens | Prepaid credits, no markup on inference (5.5% payment fee on credit purchase) | Runs through your own provider contracts (the gateway is a proxy) | Paid separately to each provider |
| Key management | One key across providers, BYOK supported | One key across providers, BYOK supported | Uses your own provider keys | Managed per provider |
| Failover | Yes | Yes | Fallback configurable | Must be built yourself |
| Strengths | Existing SDK compatibility, centralized spend and security settings | Breadth of model coverage and community | Edge-side controls such as caching and rate limiting | Fastest access to the newest features |
Who Is It For
- Teams using several AI providers at once, where key and billing management has become cumbersome
- Developers who compare and switch models frequently during development
- Operators who want to reduce the risk of one provider’s outage taking down the whole service
- Companies that need compliance requirements such as ZDR and provider restrictions enforced through settings
- Anyone who wants to consolidate the spend of coding agents such as Claude Code and Codex in one place with a cap
- Teams already running applications on Vercel who want the AI layer in the same dashboard
Summary
Vercel AI Gateway sidesteps the usual trade-off — that adding a middle layer across providers raises costs — with a no-markup billing design. Migration from existing OpenAI / Anthropic SDKs is straightforward, and it takes on operational concerns such as failover, spend management and ZDR in one place. Since the free tier covers only a subset of models, a practical path is to buy a small amount of credit, verify fees and behavior against a real workload, and then move production traffic over.