AI access & gateway glossary

Common terms in AI access and gateways, defined in one line each — so you can see the real differences behind the marketing.

AI gateway

A unified access layer between applications and large language models — one endpoint to many models, with routing, auth, quota, observability, and data controls handled centrally. Read more →

AI access layer

The layer that abstracts how you connect to and manage LLMs, so your application faces one stable interface.

Official channel

Access through the model provider's official API — not reverse-engineered, not diluted — for full, traceable capability.

Full-capability model

The original model at full strength — not downgraded, swapped, or context-trimmed; the opposite of a degraded model.

Token

The basic unit by which LLMs process text; cost and usage are typically measured in tokens.

Context window

The maximum input-plus-output tokens a model handles per call; quietly shrinking it degrades long-document tasks.

Token governance & quota

Allocating usage limits, keys, and permissions by team or project, with overage blocks and visible usage to keep cost under control.

Independent node

A dedicated node that doesn't rely on shared or pooled accounts; stability comes from monitoring and recovery, not luck.

Observability

Visibility into calls, usage, and cost that's reconcilable — for troubleshooting, reimbursement, and audit.

Data boundary

The line defining how customer data is handled — never captured, sold, or used for training, typically written into the contract.

API relay

A third-party service that forwards requests between you and the model; cheap relays often cut prices with reverse-engineering and shared accounts. Compare →

Reverse-engineered API

Accessing a model through reverse-engineering rather than official authorization; stability and compliance are both questionable.

Model degradation

Quietly routing requests to a weaker or cheaper model so real capability falls short of what's claimed.

Model dilution

Passing off a substitute or a stripped-down configuration as the flagship model.

API compatibility (OpenAI-compatible)

Following OpenAI-style API conventions, so you integrate by changing base_url while your existing SDK mostly stays the same.

Delivery

More than connecting — validation, prompts, quotas, review, and rollout that fit the model into your business.