OpenAI API
OpenAI
First-party
OpenAI-compatible API
First-party Chat Completions and Responses APIs for OpenAI models.
Hosts
First-party APIs, inference clouds that serve open models, and gateways that route across them. Each row shows what it serves, what it charges, and whether it speaks the OpenAI API, which lets any harness point at it the way it points at hardware you own.
Open the plannerRows were last checked on 2026-09-25. Catalogs and API shapes move; confirm pricing and endpoint support against each operator’s current docs before you wire anything up.
Filters
35 of 35 hosts
The OpenAI-compatible marker is what lets a harness treat a host like a local endpoint. See which tools speak that protocol in the harness directory.
OpenAI
OpenAI-compatible API
First-party Chat Completions and Responses APIs for OpenAI models.
Anthropic
Messages API is the native surface. Anthropic documents an OpenAI SDK compatibility layer at api.anthropic.com/v1 for testing and comparison, but says it is not a long-term or production-ready path (no prompt caching, strict tool schemas ignored).
OpenAI-compatible API
Gemini Developer API with an OpenAI-compatible endpoint option.
Google Cloud
OpenAI-compatible API
Google Cloud’s enterprise platform formerly called Vertex AI: Gemini, partner models such as Claude, and Model Garden open models. Catalog is Google-managed; arbitrary self-upload is outside what we mark as BYO here.
Microsoft Azure
OpenAI-compatible API
Formerly Azure AI Foundry. Azure-hosted OpenAI models plus directly sold DeepSeek, Kimi, Grok, Llama, Mistral, and gpt-oss models behind Chat Completions and Responses APIs. Managed-compute deployment of catalog models exists; arbitrary custom-weight upload is not counted as BYO here.
xAI
OpenAI-compatible API
Grok models behind an OpenAI-compatible chat completions surface.
DeepSeek
OpenAI-compatible API
Hosted DeepSeek models. Open-weight checkpoints are published separately for self-hosting; this row covers the first-party API only.
Moonshot AI
OpenAI-compatible API
Kimi models via an OpenAI-compatible platform API. The platform moved from platform.moonshot.ai to platform.kimi.ai.
Alibaba Cloud
OpenAI-compatible API
Qwen and related Model Studio models, including open-weight Qwen variants in the catalog.
Zhipu AI
OpenAI-compatible API
GLM models on the Z.ai platform with an OpenAI-compatible API (base URL api.z.ai/api/paas/v4).
Mistral AI
OpenAI-compatible API
Mistral open and commercial models through an OpenAI-compatible API, managed from Mistral Studio.
Amazon Web Services
OpenAI-compatible API
Multi-provider catalog on AWS. Native Bedrock APIs plus OpenAI Chat Completions/Responses and Anthropic Messages-compatible endpoints (bedrock-runtime and bedrock-mantle). Open-weight catalog tends to trail specialist hosts by a version or two.
Cohere
OpenAI-compatible API
Command and Embed models on Cohere’s native API, plus an OpenAI-compatible Compatibility API at api.cohere.ai/compatibility/v1.
MiniMax
OpenAI-compatible API
MiniMax text and speech models on the MiniMax platform API. Pay-as-you-go pricing moved into the platform docs.
ByteDance / Volcengine
OpenAI-compatible API
Volcengine ModelArk hosts Doubao and related ByteDance models, with an OpenAI SDK-compatible endpoint. Pricing is in CNY for the China region; international customers use BytePlus ModelArk.
Baidu
OpenAI-compatible API
Qianfan platform for ERNIE plus open models such as DeepSeek and Qwen, with an OpenAI-compatible API at qianfan.baidubce.com/v2.
Thinking Machines Lab
OpenAI-compatible API
Tinker is a fine-tuning (LoRA) and sampling API over open models plus Thinking Machines’ own Inkling models. Serverless inference for Inkling and Inkling-Small is in beta, with OpenAI- and Anthropic-compatible API docs. Prices are per million tokens, split by prefill, sample, and train.
NVIDIA
OpenAI-compatible API
Microservice containers and hosted endpoints for NVIDIA-optimized and partner open models. Hosted endpoints are free for prototyping via the NVIDIA Developer Program; production self-hosting needs NVIDIA AI Enterprise (from $4,500 per GPU per year).
Cerebras
OpenAI-compatible API
Positions on wafer-scale chips for very high tokens-per-second on supported models.
Public pay-as-you-go catalog is down to two open models, GPT OSS 120B and Qwen 3.8 27B, on Cerebras hardware. BYO custom weights are outside the public inference product as checked.
Groq
OpenAI-compatible API
LPU inference marketed for low latency on a fixed open-model catalog.
OpenAI-compatible chat and audio endpoints for Groq’s hosted model list. Llama 3.1 8B and Llama 3.3 70B are still served but now priced through sales; Qwen 3.8 27B (preview) replaced Qwen 3.6 27B.
SambaNova
OpenAI-compatible API
Custom dataflow (RDU) chips marketed for high tokens-per-second on large open models.
Small catalog of open models (MiniMax M3 and M2.7, DeepSeek V3.x, gpt-oss-120b, Gemma 4 31B, Llama 3.3 70B) on SambaNova hardware. Docs describe running inference on your own fine-tuned checkpoints.
Together AI
OpenAI-compatible API
Serverless open-model inference plus dedicated endpoints and fine-tuning. Serverless list now includes Kimi K3, Qwen3.8-2.4T-A95B, and DeepSeek V4 Pro 0813; Nemotron 3 Ultra is no longer offered.
Fireworks AI
OpenAI-compatible API
Fast open-model inference with deployment options for custom fine-tunes. Every serverless model also has a pricier Priority mode, and US-only serverless models cost 1.5x base since September 1, 2026.
DeepInfra
OpenAI-compatible API
Broad open-model catalog with OpenAI-compatible and native HTTP APIs.
Baseten
OpenAI-compatible API
Model deployment platform; OpenAI-compatible wrappers are common on deployed Truss apps.
Nebius
OpenAI-compatible API
Open-model serverless and dedicated endpoints on Nebius cloud with an OpenAI-compatible API. Custom model weights on dedicated endpoints are in beta and available on request.
Novita AI
OpenAI-compatible API
Open-model serverless API with GPU deployment options.
Replicate (Cloudflare)
OpenAI-compatible API
Strong on image, video, and audio models alongside LLMs; Replicate is now part of Cloudflare and its catalog is being integrated into Workers AI. OpenAI-compatible HTTP is available for a subset of language models.
Cloudflare
OpenAI-compatible API
Serverless open-model inference on Cloudflare’s network, billed in Neurons ($0.011 per 1,000; 10,000 free per day) with per-token equivalents published. Cloudflare plans to add Replicate’s catalog and custom-model support.
OpenRouter
OpenAI-compatible API
One OpenAI-compatible key across many provider backends and open-model hosts. Model prices pass through at provider rates; credit purchases carry a 5.5% fee, and BYOK is free up to $25K/month of usage, then 5%.
Hugging Face
OpenAI-compatible API
One Hugging Face token routed across partner hosts such as Groq, Cerebras, Together, Fireworks, and Novita. OpenAI-compatible chat endpoint with :fastest and :cheapest provider policies, no markup on provider rates, and a small free monthly credit.
Vercel
OpenAI-compatible API
Routes through Vercel’s AI Gateway to multiple providers at provider list price with no token markup (including BYOK), plus shared observability. Some team-wide add-ons (allowlist, ZDR, reporting) bill extra.
BerriAI
OpenAI-compatible API
Self-hosted (or proxied) gateway that translates one OpenAI-compatible surface onto many providers. You run the proxy; usage cost is whatever backends you attach.
Cloudflare
OpenAI-compatible API
Proxy layer in front of provider APIs for caching, logging, and rate controls. Core features are free; optional Unified Billing adds a 5% fee on credit purchases.
Palo Alto Networks
OpenAI-compatible API
AI gateway with routing, guardrails, and an OpenAI-compatible control plane. Acquired by Palo Alto Networks (closed May 29, 2026) and being folded into Prisma AIRS; the free Developer and $49/month Production plans are still listed.