AI plugins on an API gateway
INFRO vs Kong AI Gateway
Kong AI Gateway brings AI traffic onto Kong Gateway, one of the most widely deployed API gateways: a model and provider abstraction over the major providers, load balancing and failover between them, semantic caching, token and cost-based rate limits, prompt guards, PII sanitizing, cloud guardrail integrations and MCP tooling — all as policies on the same runtime that already carries a company's APIs. If your platform team runs Kong today, adding AI traffic to it is a configuration change rather than a new vendor, and that is a strong position.
- You already run Kong. Your APIs, your authentication plugins, your observability and your on-call already live there, and AI traffic becomes one more set of routes under the same operating model. A second gateway in front of your models would be scope you do not need.
- You need to run the whole gateway yourself — data plane in your own environment, or control plane and data plane both on-premises. Kong is built for that. INFRO's plans are hosted; self-hosted and hybrid deployment come only with an enterprise agreement.
- Plugin depth is the requirement. Kong's plugin ecosystem — custom plugins, request and response transformations, cloud guardrail integrations with AWS, Azure and Google — goes far beyond what INFRO exposes, and INFRO has no plugin system at all.
- You govern AI traffic and ordinary API traffic with one policy model, and splitting the two would cost you more in consistency than it would buy in AI-specific features.
- You need it this week. Kong is generally available; INFRO is in private early access.
This section is not a courtesy. If one of these describes you, pick Kong AI Gateway — a customer who adopts the wrong layer on our advice is worse for us than one we never had.
Where the two actually diverge.
- An API gateway that learned AI, or a control plane built for it
- Kong's unit is the route and the plugin: AI capabilities are policies you attach to traffic that passes through a general-purpose gateway. INFRO's unit is the organization — projects, environments, gateway keys, provider connections and the people who own them — and every request carries an estimated provider cost attributed to a project, key, model, route or end user. Both are coherent designs. The difference shows when someone outside the platform team asks a question: a finance lead asking what the support assistant cost last month, or a team lead asking which models their project may call, gets an answer in INFRO's console without a query against gateway logs.
- Where the cost number comes from, and what checks it
- Kong can enforce spend: its AI Rate Limiting Advanced plugin counts cost from the input and output prices you set per model in the AI Proxy configuration, and returns 429 at the limit. That is a capable mechanism, and it is a price you maintain in gateway config. INFRO keeps the vendor's published rate per model, lets you enter the rates your contracts carry from the Team plan, records a request with no known price as unpriced rather than free, and on the Business plan checks the month's estimate against OpenAI's and Anthropic's own cost APIs and states the variance. Budgets then act on that number at organization, project, member and key level.
- What each tier includes
- Kong's pricing page lists role-based access control, single sign-on and audit logging as Enterprise only; several AI plugins, semantic caching among them, are a paid add-on on its self-serve tier; the self-serve tier also limits how many distinct models its AI plugins proxy; and cost-based rate limiting is documented as part of the AI Gateway Enterprise offering. Read their pricing page for the details. In INFRO, budgets, model allowlists, the audit log and the response cache are on every plan, members and roles from the Startup plan, negotiated rates and end-customer caps from Team, and single sign-on, SCIM, guardrails, the MCP gateway and reconciliation on Business.
- Not either-or
- Many companies that run Kong will keep running it. INFRO's gateway speaks the OpenAI API, so a Kong route can send AI traffic to INFRO as an upstream: Kong keeps the API estate and its edge policies, and INFRO adds provider connections, cost attribution, budgets and reconciliation behind it. That is two hops instead of one, which is a real cost, and whether it is worth paying depends on how much of the second list you need.
Side by side.
Our column carries dashes too. A comparison where one side is all checks is one nobody believes, and rightly.
| Capability | Kong AI Gateway | INFRO |
|---|---|---|
| One API across many providers | yesmodel and provider entities | yesthe providers you connect |
| Routing, retries and failover | yesload balancing across providers | yes |
| Your own provider accounts | yescredentials in gateway config | yesconnected per environment |
| Image, video, and audio | yesimage, speech, transcription and video routes | yesone request shape each |
| Self-hosting | yeshybrid or fully self-hosted | partialenterprise agreement |
| Plugin ecosystem and custom plugins | yestheir specialism | nono plugin system |
| Guardrails and PII redaction | yesprompt guards, PII sanitizer, cloud guardrails | partialchat patterns plus prompt classifiers; Business plan |
| Response caching | yessemantic cache; paid add-on on Plus | partialexact match, chat completions; every plan |
| Spend limits in dollars that refuse a request | partialcost-mode rate limits; AI Gateway Enterprise | yes402; org, project, member, key; every plan |
| Estimates at your negotiated rates | partialper-model cost in plugin config | yesstored per organization; Team plan |
| Reconciliation against the provider's own bill | no | partialOpenAI and Anthropic, Business plan; variance stated |
| Roles, SSO and audit log | partialEnterprise tier | yesaudit every plan; SSO and SCIM on Business |
| Available today | yes | noprivate early access |
Our reading of each product's public documentation as of October 2026. We do not quote competitors' prices — theirs change and a wrong number about someone else's business is worse than none. Think a row is wrong? Tell us and we will fix it.
Work out what the difference is worth
Estimate what your current model mix costs at your providers' list rates and the plan it needs, or have us read your real usage and tell you where we would and would not help.