Which AI gateway should a funded AI startup choose? The one that answers four questions well: what each customer costs you, whether it can stop one customer before the provider is called, whether it fails over when a provider has an outage, and whether traffic stays on provider accounts you own. To self-host, look at LiteLLM or Bifrost. If you already run on Cloudflare, Cloudflare AI Gateway. For one balance across every model without provider accounts of your own, OpenRouter or Requesty. If you build on Vercel and need nothing beyond its budgets, Vercel AI Gateway. For a hosted gateway and control plane that runs only on your own provider accounts, with a flat subscription and 0% markup on model spend, INFRO.
This guide is for a startup with several engineers, a product in production and customers on usage-based or seat pricing. If you are one person shipping a first product, read Do solo founders need an AI gateway? instead.
What should an AI startup need from a gateway?
A funded AI startup needs a gateway that protects gross margin and uptime, not one with the longest feature list. Five things separate the options at this stage:
- Cost per customer. Your pricing depends on what each customer costs you in model spend. The gateway should attribute every request to an end-customer id you send. AI feature unit economics shows the arithmetic.
- A cap per customer that acts before the call. One customer's runaway agent should be refused at your limit, not discovered on the invoice. Per-customer AI usage limits covers the design.
- Failover. When a provider errors or rate-limits, the request should move to another provider or key before your user sees it.
- Your own provider accounts. If a provider or cloud program has granted your company credits, or you have negotiated rates, those sit on your own account. A gateway that sends traffic on your keys keeps that usage where the credits and rates apply; one that routes through its own accounts does not.
- Low lock-in and a known price. An OpenAI-compatible endpoint you can point back at your provider, and a gateway price you can forecast: a subscription, a percentage of model spend, or an engineer's time.
Which AI gateways should startups compare in 2026?
| Gateway | Who runs it | Provider accounts | Per-customer caps | Gateway's own price |
|---|---|---|---|---|
| LiteLLM | You (open source) | Yours | Yes, customer budgets | Free license, your infrastructure |
| Bifrost | You (Apache 2.0) | Yours | Customer budgets listed | Free license, your infrastructure |
| Cloudflare AI Gateway | Cloudflare | Yours or Cloudflare's billing | Spend limits split by a user-id metadata key | Core features free; 5% on Unified Billing purchases |
| Vercel AI Gateway | Vercel | Vercel's, or yours on the paid tier | Budgets do not count BYOK spend | See Vercel's pricing |
| OpenRouter | OpenRouter | OpenRouter's, or yours | Budgets count BYOK spend only if switched on | 5.5% or 8% on purchases; BYOK 5% after $25,000 a month |
| Requesty | Requesty | Requesty's, or yours | Per key and per group | Usage plus 5% |
| INFRO | INFRO (hosted) | Only yours | From the Team plan | $79 to $599 a month; 0% on model spend |
Which gateway is best for a startup that wants to self-host?
LiteLLM is the default self-hosted answer: one OpenAI-compatible endpoint for more than 100 providers, virtual keys per user, team and project with their own budgets and rate limits, spend tracked by key, team, user and tag, and load balancing, fallbacks and retries, deployed with Docker, Helm or Terraform against a Postgres database1. It also tracks spend and sets budgets per end customer, identified by the request's user field or a header, with a default budget for customers who have none2. That covers the per-customer cap directly.
Bifrost is Apache 2.0, starts with npx -y @maximhq/bifrost or Docker, and lists automatic fallbacks between providers and models and "hierarchical cost control with virtual keys, teams, and customer budgets"3.
Both are the better pick when you have an engineer who will own the gateway: patching it, securing the database that holds your provider keys, and keeping it up, since it sits on every request.
Which gateway is best for a startup already on Cloudflare or Vercel?
Cloudflare AI Gateway offers analytics, caching and rate limiting free; guardrails are billed as Workers AI inference, and Unified Billing adds a 5% fee to purchases4. Its spend limits are dollar budgets scoped by provider, model or custom metadata; split by a user_id metadata key, each customer gets its own budget, and a request over budget gets a 429 or moves to a fallback model. They apply to BYOK requests for models with known pricing, and Cloudflare states that cost tracking is a best-effort estimate and that concurrent bursts can briefly exceed a limit5. For a startup on Cloudflare, it is the natural first gateway.
Vercel AI Gateway suits teams on Vercel and the AI SDK. Its BYOK page notes that BYOK is useful for "using credits provided by the AI provider", but BYOK is on the paid tier and needs purchased AI Gateway credits, a failed request on your credentials is retried on Vercel's system credentials and billed to that balance, and spend through your own credentials "isn't counted in budgets", so a budget cannot cap BYOK spend6. If per-customer caps on your own keys matter, plan for that.
Which gateway is best for a startup without provider accounts?
OpenRouter gives one balance for a very large catalog. Its fee is charged when you buy credits, 5.5% on Standard and 8% on Business; BYOK carries no fee up to $25,000 of list-price inference a month and 5% after7. When your keys fail or are rate limited it falls back to shared OpenRouter capacity unless you switch that off, and BYOK spend counts toward budgets only if you turn on "Include BYOK spend"8. Stripe announced on 19 August 2026 that it has agreed to acquire OpenRouter9.
Requesty has a free plan of 200 requests a day on free models and a pay-as-you-go plan described as "what your application spends, plus 5%", with bring-your-own-keys, automatic fallbacks, spend limits per key and group budgets10.
Both are the better pick before you have provider accounts or when you want to try many models quickly. A percentage of model spend grows with your usage, which is the line to model against your gross margin.
Is Helicone still an option for startups?
Not for new teams. Mintlify announced on 3 March 2026 that it is acquiring Helicone and moving it to maintenance mode, with security patches, bug fixes and new models, and help for every customer to migrate11. Its sign-up page reads "Sign ups are disabled"12. Helicone alternatives covers the move.
INFRO, and when it is not the right choice
INFRO is a hosted gateway and control plane that runs only on provider accounts you connect. Your providers bill you for inference, directly; INFRO charges a software subscription with 0% markup on model spend. Every request records its route, provider, tokens, latency and estimated provider cost at list rates, and can be grouped by model, project, key or an end-customer id you attach, on every plan. Budgets at organization, project, member and key level are checked before the provider is called. Named routes fail over to another connection when a provider times out or returns a 5xx before the response starts. Each app, agent and teammate gets its own gateway key, scoped to projects, models and providers, IP-limited and rotatable, so no raw provider key ships in your app.
What a funded startup should know about the plans: Developer is $35 a month for one seat; Startup is $79 a month for 3 seats, 10 projects and 500k requests a month; Team is $199 a month for 10 seats and adds end-customer spend caps (a refused request returns a 402 with the time the cap resets), a per-customer usage export for your own billing, and negotiated provider rates in cost estimates. Every plan starts with a 14-day trial, card required. Estimated provider cost and the INFRO subscription are separate numbers and are never added together.
INFRO for Startups gives early AI startups 12 months of the Startup plan without charge: 50 seats per batch, for live products or ones launching within 60 days, raised under $2M, 15 people or fewer. Batch 01 takes applications from 2 November to 30 November 2026. The program asks that you route production traffic through INFRO within 14 days, take one 15-minute call, and approve a logo and short case study; the year-two price is locked at $39.50 a month. The AI startups page covers the product side.
INFRO is not the right choice if you need to self-host without an enterprise agreement (use LiteLLM or Bifrost), want a gateway at zero subscription cost (Cloudflare or a self-hosted gateway), have no provider accounts and want one balance for every model (OpenRouter or Requesty), or need per-customer caps while staying on the Startup plan's price (LiteLLM, Bifrost and Cloudflare offer them without a plan step). INFRO compared has each product side by side.
How should a startup choose an AI gateway this week?
- Write down whose accounts carry the traffic, and whether you hold provider credits or negotiated rates you need to keep using.
- Decide whether you need a per-customer cap now, or only per-customer cost.
- Model the gateway's price at next year's spend: a percentage of model spend, a flat subscription, or an engineer's time.
- Run a day of real traffic through the top two behind a flag, and pull one provider key to see what failover does. From direct OpenAI calls to a gateway is the runbook.
Summary
For a funded AI startup, the best AI gateway is the one that shows cost per customer, caps it before the call, fails over during an outage and keeps traffic on your own accounts. Self-hosters have LiteLLM and Bifrost; Cloudflare teams have a free starting point; teams without provider accounts have OpenRouter and Requesty; and teams that want a hosted control plane on their own keys at a flat price can use INFRO. The broader comparison covers the enterprise options.