LLM gateways and OpenRouter alternatives

Where to actually run a model

117 services that will run a model for you, grouped by what they actually are. Read the category before the price: a router and a GPU cloud both look cheap and solve completely different problems.

Routers

One endpoint in front of many models, picking a host per request. This is the OpenRouter category.

Routers
Service	Pricing model	Adds on top	Free tier	OpenAI-compatible	HQ
AI/ML API	Subscription plus per-token	~30% above list price	Yes	Yes	US
Eden AI	Per call with a markup	5.5% above list price	Yes	No	FR
OpenRouter	Upstream price plus a credit purchase fee; BYO-key supported	0% on tokens; 5.5% ($0.80 min) on credit purchase; BYOK 5% above $25k/mo	Yes	Yes	US
Requesty	Markup on upstream tokens	5% above list price	Yes	Yes	NL
Vercel AI Gateway	Upstream price, BYO-key supported	0% - list price, including on bring-your-own-key	Yes	Yes	US

Serverless inference

Host the weights themselves and bill per token. Where most open-model traffic actually runs.

Serverless inference
Service	Pricing model	Adds on top	Free tier	OpenAI-compatible	HQ
AI21	Per token	-	-	Yes	IL
AionLabs	Per token	-	-	Yes	IL
AkashML	Per token	-	-	Yes	-
Alibaba	Per token	-	-	Yes	SG
Amazon Nova	Per token	-	-	Yes	US
Ambient	Per token	-	-	Yes	-
Anthropic	Per token	-	-	Yes	US
Arcee AI	Per token	-	-	Yes	US
AtlasCloud	Per token	-	-	Yes	US
Avian	Per token	-	-	Yes	US
Baidu	Per token	-	-	Yes	CN
BaseTen	Per token or per GPU-minute	None - sets its own prices	-	Yes	US
Black Forest Labs	Per token	-	-	Yes	-
Cerebras	Per token, with paid tiers for higher limits	None - sets its own prices	Yes	Yes	US
Chutes	Per token	None - sets its own prices	Yes	Yes	US
Cirrascale	Per token	-	-	Yes	-
Clarifai	Per token or per operation	None - sets its own prices	Yes	Yes	US
Claude Platform on AWS	Per token	-	-	Yes	US
Cloudflare	Neurons, with a free daily allocation	None - billed in neurons	Yes	Yes	US
Cohere	Per token	-	-	Yes	US
Cosine	Per token	-	-	Yes	-
Crucible	Per token	-	-	Yes	-
Darkbloom	Per token	-	-	Yes	-
Decart	Per token	-	-	Yes	US
Deepgram	Per token	-	-	Yes	-
DeepInfra	Per token, pay as you go	None - systematically undercuts list price	-	Yes	US
DeepSeek	Per token	-	-	Yes	CN
DekaLLM	Per token	-	-	Yes	ID
DigitalOcean	Per token	None - sets its own prices	-	Yes	-
Featherless	Flat monthly subscription	Flat subscription instead of a margin	-	Yes	US
Fireworks	Per token, plus per-GPU-hour for dedicated	None - sets its own prices	Yes	Yes	US
Fish Audio	Per token	-	-	Yes	-
Friendli	Per token or per GPU-hour	None - sets its own prices	Yes	Yes	US
GMICloud	Per token	-	-	Yes	US
Groq	Per token	None - sets its own prices	Yes	Yes	US
HeyGen	Per token	-	-	Yes	US
Inception	Per token	-	-	Yes	-
Inceptron	Per token	-	-	Yes	SE
Inferact vLLM	Per token	-	-	Yes	-
InferenceNet	Per token	-	-	Yes	US
Infermatic	Per token	-	-	Yes	-
Inflection	Per token	-	-	Yes	-
Io Net	Per token	-	-	Yes	US
Ionstream	Per token	-	-	Yes	US
Krea	Per token	-	-	Yes	US
Liquid	Per token	-	-	Yes	-
Makora	Per token	-	-	Yes	-
Mancer 2	Per token	-	-	Yes	-
Mara	Per token	-	-	Yes	US
Meta	Per token	-	-	Yes	US
Minimax	Per token	-	-	Yes	SG
Mistral	Per token	-	-	Yes	FR
ModelRun	Per token	-	-	Yes	US
Modular	Per token	-	-	Yes	US
Moonshot AI	Per token	-	-	Yes	SG
Morph	Per token	-	-	Yes	US
Near AI	Per token	-	-	Yes	-
Nebius	Per token	None - sets its own prices	Yes	Yes	NL
Nex AGI	Per token	-	-	Yes	CN
NextBit	Per token	-	-	Yes	ES
Novita	Per token	None - sets its own prices	Yes	Yes	US
Nvidia	Per token	-	-	Yes	US
Ollama	Free locally, subscription in the cloud	None - subscription	Yes	Yes	-
OpenAI	Per token	-	-	Yes	US
OpenInference	Per token	-	-	Yes	US
Parasail	Per token	None - sets its own prices	-	Yes	US
Perceptron	Per token	-	-	Yes	US
Perplexity	Per token	-	-	Yes	US
Phala	Per token	-	-	Yes	US
Poolside	Per token	-	-	Yes	-
PrimeIntellect	Per token	-	-	Yes	-
Quiver	Per token	-	-	Yes	-
Recraft	Per token	-	-	Yes	-
Reka	Per token	-	-	Yes	-
Relace	Per token	-	-	Yes	-
Replicate	Per second of compute, or per token on some models	None - billed for compute, not a token margin	Yes	No	US
Runway	Per token	-	-	Yes	US
Sail Research	Per token	-	-	Yes	US
Sakana AI	Per token	-	-	Yes	-
SambaNova	Per token	None - sets its own prices	Yes	Yes	US
Seed	Per token	-	-	Yes	SG
SiliconFlow	Per token	None - sets its own prices	Yes	Yes	SG
Sourceful	Per token	-	-	Yes	-
Stealth	Per token	-	-	Yes	-
StepFun	Per token	-	-	Yes	-
StreamLake	Per token	-	-	Yes	CN
Switchpoint	Per token	-	-	Yes	US
Tencent	Per token	-	-	Yes	CN
Tenstorrent	Per token	-	-	Yes	US
Thinking Machines	Per token	-	-	Yes	-
Together	Per token, per model	None - sets its own prices	Yes	Yes	US
Upstage	Per token	-	-	Yes	-
Venice	Subscription or per token	None - subscription	Yes	Yes	US
VoyageAI by MongoDB	Per token	-	-	Yes	-
Wafer	Per token	-	-	Yes	US
xAI	Per token	-	-	Yes	US
Xiaomi	Per token	-	-	Yes	CN
Z.AI	Per token	-	-	Yes	SG

Hyperscaler clouds

Managed model services inside a hyperscaler, bought with existing cloud commitments.

Hyperscaler clouds
Service	Pricing model	Adds on top	Free tier	OpenAI-compatible	HQ
Amazon Bedrock	Per token, plus provisioned throughput	None - list price	-	No	US
Azure	Per token, plus provisioned units	None - list price	-	Yes	US
Databricks	Per token or per DBU	None - list price	-	Yes	US
Google	Per token or per character	None - list price	-	No	US
Google AI Studio	Per token, free tier then paid	None - list price	Yes	Yes	US

GPU clouds

Rent the hardware and run your own server. Cheapest at scale, most work to operate.

GPU clouds
Service	Pricing model	Adds on top	Free tier	OpenAI-compatible	HQ
CoreWeave	Per GPU-hour	None - undercuts list price	-	No	US
Crusoe	Per token or per GPU-hour	None - sets its own prices	-	Yes	US
Hyperbolic	Per token or per GPU-hour	None - GPU rental	Yes	Yes	US
Lambda	Per token or per GPU-hour	None - GPU rental	-	Yes	US
Modal	Per second of GPU time	None - billed for compute, not tokens	Yes	No	US
RunPod	Per second of GPU time	None - billed for GPU time	-	No	US

Proxies & observability

Sit in front of your own provider keys to add caching, logging, guardrails and spend limits.

Proxies & observability
Service	Pricing model	Adds on top	Free tier	OpenAI-compatible	HQ
Helicone	Per logged request, free tier then subscription	0% on tokens - billed per logged request	Yes	Yes	US
Portkey	Per request, free tier then subscription	0% on tokens - billed per request	Yes	Yes	US

SDKs & self-hosted

Libraries and self-hosted proxies that normalise provider APIs without taking a cut.

SDKs & self-hosted
Service	Pricing model	Adds on top	Free tier	OpenAI-compatible	HQ
LiteLLM	Free and open source; paid enterprise tier	0% - self-hosted, bring your own provider keys	Yes	Yes	US

Some of these services have left the category

Unify, Kluster.ai, Predibase, Martian, Hyperbolic and Anyscale no longer offer what they are still widely recommended for. See what replaced each one.

Frequently asked

What is the difference between a provider and a gateway?

A provider trains and publishes a model's weights. A gateway serves those weights over an API. OpenAI is both. DeepSeek publishes weights that dozens of gateways serve, often at very different prices for identical output.

What are the main alternatives to OpenRouter?

Requesty, Vercel AI Gateway and AI/ML API compete most directly as routers. LiteLLM and Portkey solve the same problem without taking a token margin - you bring your own provider keys. Together, Fireworks, Groq, DeepInfra and Novita host open models directly and are usually cheaper per token, but cover fewer models. Note that several services still listed as OpenRouter alternatives elsewhere have left the category: Unify shut its router down, Kluster.ai is offline, Martian became an interpretability lab, and Hyperbolic pivoted to GPU rental.

Do gateways charge more than going direct?

The spread is wider than most people assume. Vercel AI Gateway adds 0% and passes through list price. OpenRouter also adds 0% on tokens, monetising through a 5.5% fee ($0.80 minimum) when you buy credits. Requesty adds 5%, Eden AI 5.5%, and AI/ML API roughly 30% over list. LiteLLM and Portkey take nothing on tokens at all because you bring your own keys. Serverless hosts set their own prices entirely and are frequently cheaper than the model's own lab for open-weight models.

Is a router worth it if it adds a fee?

Often, because the price spread between hosts of the same model is larger than any router's fee. The same open-weight model can vary more than tenfold in input price across the hosts serving it, so routing to the cheapest healthy endpoint usually saves more than the margin costs. Each model page lists every host side by side so you can check the spread yourself.