Skip to content
Providers

Who actually serves your request

One schema in front of all of them. Prices below are what each route costs us: a direct provider's published rate, or an OpenRouter rate with their 5.5% credit fee in it. We add no spread of our own to either.

Most of this catalogue is reached through OpenRouter

84% of our 338 enabled routes, 284 of them: go to OpenRouter, another gateway, rather than to a direct contract with the company that made the model. That is what lets this catalogue list models we hold no account for.

Those routes cost more than the number OpenRouter publishes, and the rates here say so. We do not buy tokens from them, we buy their credit, and a dollar of it costs $1.055, so a route they list at $10.00 per million costs us $10.55, and $10.55 is what the catalogue charges. The 5.5% is theirs and it is inside the rate you can see. We add no spread of our own to either kind of route.

Two consequences worth having before you route production traffic. For a model that only reaches you through OpenRouter, we cannot be cheaper than OpenRouter. Their fee sits under ours, and what you are buying is the breadth of one catalogue on one balance. And a brokered route is one more company between you and the model, with its own terms and its own outages. Bringing your own key for a vendor takes us out of that path for their models, and costs nothing.

OpenAI

The broadest frontier line-up, and the wire format every other provider ended up copying.

Live
Models
15
From
Free
Max context
1.0M
Uptime30 days
100.0% up0.0% degradedReachability probes: 99.98% of 5658, 100% of the window observed.
Frontier reasoningTool callingStructured output
Anthropic logo

Anthropic

Long-context work and agentic tasks that have to stay coherent over many steps.

Live
Models
6
From
$5.00
Max context
1M
Uptime30 days
100.0% up0.0% degradedReachability probes: 99.98% of 5658, 100% of the window observed.
Long contextAgentic workPrompt caching
Groq logo

Groq

Open-weight models served fast enough that latency stops being the thing you design around.

Live
Models
4
From
$0.080
Max context
131K
Uptime30 days
100.0% up0.0% degradedReachability probes: 99.98% of 5658, 100% of the window observed.
Lowest latencyCheap open weightsHigh throughput
DeepInfra logo

DeepInfra

The cheapest way to run large open-weight models at volume.

Live
Models
11
From
Free
Max context
1M
Uptime30 days
99.9% up0.1% degradedReachability probes: 99.88% of 5658, 100% of the window observed.
Lowest priceLarge open modelsEmbeddings
Together AI logo

Together AI

The widest open-weight catalogue, including models nobody else bothers to host.

Live
Models
13
From
Free
Max context
1M
Uptime30 days
99.9% up0.1% degradedReachability probes: 99.91% of 5658, 100% of the window observed.
Widest open catalogueFine-tuningDedicated endpoints
openrouter logo

OpenRouter

Another gateway, used here as a provider. Most of the catalogue is reachable through OpenRouter rather than through a direct contract with the model's maker, which is how Multigrid can list models it has no account for. Where we hold a direct route, that one is priced and tried on its own merits.

Live
Models
284
From
$0.032
Max context
2M
Uptime30 days
100.0% up0.0% degradedReachability probes: 99.98% of 5658, 100% of the window observed.
Widest catalogueModels we hold no direct key for
  • Succeeded
  • Failed on a degraded day
  • Failed on a bad day (< 50%)
  • Nothing measured yet

Bar height carries the state as well as colour, and an empty slot is an hour we did not observe. Both signals, full width, on the status page.

What “bring your own key” means here

A provider marked that way is one this deployment holds no platform key for, so we cannot bill its usage to your Multigrid balance. Add your own key on the provider keys page and every model it serves becomes available immediately, and that traffic carries no fee from us at all, because you are paying the provider directly.

ProviderWire formatBase URLModels routed
OpenAIopenaihttps://api.openai.com/v115
Anthropicanthropichttps://api.anthropic.com/v16
Groqopenaihttps://api.groq.com/openai/v14
DeepInfraopenaihttps://api.deepinfra.com/v1/openai11
Together AIopenaihttps://api.together.xyz/v113
OpenRouteropenaihttps://openrouter.ai/api/v1284

Where a model has several routes, the cheapest is tried first. Each route is ranked by its own completion price, so the order is per model rather than a standing ranking of providers. Your own keys come before all of them, and a request can override the ordering with provider: { sort: "price" | "latency" | "throughput" } the last two are ranked from our own measured traffic, and providers we have not measured keep their catalogue position rather than being assumed slow. A provider showing zero has a key but no routes yet.

Run inference yourself? What it takes to list your models here, including the reasons you might not want to yet.

One key instead of six

Every provider above is a separate account, billing relationship and key if you go direct. Signing up here takes no card and commits you to nothing, and if you already hold contracts with any of them, bring those keys and we charge nothing at all on that traffic.