Skip to content
ShareQuota

How this compares

Four routers and the option of calling the providers directly. Each row says which one to pick and why — including the rows where the answer is not this one.

Pick by what you want

Eight things people actually come here deciding between, and the honest answer for each — including the four rows where the answer is another tool.

Comparison
400+ models, and nothing to installOpenRouterFive providers and a process to run is a worse answer to that question than a hosted endpoint with no install at all.
Hundreds of providers, MCP, token compression, a desktop appOmniRouteTheir breadth and feature surface are far past this one, and it is free and MIT too.
One shared gateway for a whole team, on a VPS9routerShareQuota is a single-user local proxy. Its management API is loopback-only by design.
Exactly one provider, and no plans for a secondTheir API, directlyA routing layer earns its keep at the second provider. At the first it is one more thing to run.
To use the Claude Pro or ChatGPT Plus plan you already pay forShareQuotaOpenRouter sells API credits and cannot take a subscription login. This signs in with OAuth and uses the plan you have.
One key across Claude Code, Codex and Gemini CLI at onceShareQuotaFour dialects on one endpoint, and the CLI configs are switched over reversibly — backed up before the first write, restored on disable.
The lightest thing that will do the jobShareQuotaOne Go binary, three direct dependencies, no runtime, no container, no account.
Nothing about your requests to leave your machineShareQuotaThere is no server in the middle. Requests go straight to the provider; state is a SQLite file in your data directory.

Competitor facts read from their own sites on 2026-08-05. They change; this page is checked against them.

Five providers, and why that is the number

It is the question this product gets most, so here is the arithmetic rather than a slogan. Those five are the ones with subscriptions worth pooling — and OpenRouter is one of them, which is how the long tail is reached. What the number buys is depth: the binary is 157,679 lines of Go under 3,015 race-enabled tests, and the translator layer is pinned to golden fixtures rather than to hope. Breadth without tests is a list, not a feature.

See the architecture

There is no cost row on this page

Because a price claim has to be checkable. Every rate is published per million tokens beside the provider’s own published rate, so the difference is arithmetic you can do rather than a percentage to trust. Bring your own accounts and you pay your providers directly, exactly as now.

Pricing