> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

---
title: "APIs, Inference Platforms & Gateways"
url: https://daily.dev/agentic-ai-hub/apis-inference-platforms-gateways/
description: "Four rough tiers, each with different trade-offs on model access, speed, price, and control."
lastUpdated: "2026-07-22"
---

Four rough tiers, each with different trade-offs on model access, speed, price, and control.

## First-party model APIs

Buy directly from the lab: **OpenAI**, **Anthropic**, **Google (Gemini)**, plus **Mistral**, **Cohere**, **xAI**, **DeepSeek**, and others. You get the newest models first, best feature support (tools, structured outputs, caching, batch), and the vendor's own reliability, at list price, with lock-in to that provider's SDK/quirks.

## Aggregators / routers

[**OpenRouter**](https://openrouter.ai/) puts one API and one bill in front of hundreds of models across dozens of providers, with automatic fallback and price/latency-based routing. It's best for trying many models without many contracts and for resilience via multi-provider fallback. The trade-off is a hop in the middle (a small latency/trust dependency), and some cutting-edge provider features may not pass through. [**Hugging Face Inference Providers**](https://huggingface.co/docs/inference-providers) is a similar OpenAI-compatible router across a dozen-plus backends and hundreds of models.

## Fast-inference / open-model hosts

These serve open-weight models (Llama, Qwen, DeepSeek, Mistral, gpt-oss, etc.) with heavy optimization, often OpenAI-compatible, usually priced per-token:

| Host | Best at |
| :---- | :---- |
| [Groq](https://groq.com/) | Low-latency token streaming (LPU hardware, [AI Hardware & Accelerators](/agentic-ai-hub/ai-hardware-accelerators/)) |
| [Cerebras](https://www.cerebras.ai/) | High inference speeds via wafer-scale ([AI Hardware & Accelerators](/agentic-ai-hub/ai-hardware-accelerators/)) |
| [Together AI](https://www.together.ai/) | Broad open-model catalog: serving, fine-tuning, and GPU clusters |
| [Fireworks AI](https://fireworks.ai/) | Fast serving, function calling, and fine-tuning, with a production focus |
| [Baseten](https://www.baseten.co/) | Deploying *your own* models (open Truss/BEI stack) with autoscaling: more "bring your model" than a fixed menu |
| [SambaNova Cloud](https://sambanova.ai/) | Very fast token serving on RDU hardware; a Groq/Cerebras peer for latency-sensitive open models |
| [Replicate](https://replicate.com/) | Any-model, pay-per-second hosting, useful for non-LLM (image/audio/video) too |
| [Deepinfra](https://deepinfra.com/) | Low-cost per-token serving of a broad open-model catalog |
| [Novita](https://novita.ai/) · [Hyperbolic](https://hyperbolic.ai/) | Budget open-model inference and GPU rental |

## Cloud provider platforms

| Platform | What it is |
| :---- | :---- |
| [AWS Bedrock](https://aws.amazon.com/bedrock/) | Multi-model (Anthropic, Meta, Amazon Nova, etc.) inside AWS with IAM/VPC/guardrails |
| [Google Vertex AI](https://cloud.google.com/vertex-ai) | Gemini plus open/partner models on GCP |
| [Azure AI Foundry / Azure OpenAI](https://azure.microsoft.com/) | OpenAI and others with Azure enterprise controls |

Best at enterprises wanting models under existing cloud contracts, compliance, data residency, and networking. Gotcha: newest models sometimes land here later than on first-party APIs, and there's more setup.

## Gateways / LLM proxies

A **gateway** is your own routing/control layer in front of any of the above: one internal API that handles key management, provider fallback, rate limiting, caching, cost tracking, and logging. Options: **LiteLLM** (open-source proxy/SDK that normalizes 100+ providers to the OpenAI format, and is very common), **Portkey**, **Cloudflare AI Gateway**, **Kong AI Gateway**, and **Helicone** (observability-first proxy, [LLMOps: Evals, Observability & Guardrails](/agentic-ai-hub/llmops-evals-observability-guardrails/)). Trade-off: another hop and single-point-of-failure to run, in exchange for provider-independence and central governance. Most teams past prototype should run a gateway so no application code is welded to one vendor.

> Selection heuristic: prototype on a **first-party API or OpenRouter**; put a **gateway (LiteLLM/Portkey)** in front once you have real traffic; move latency-critical open-model paths to **Groq/Cerebras/Fireworks/Together**; adopt **Bedrock/Vertex/Azure** when compliance or cloud-contract gravity demands it.