All OpenRouter (OpenRouter) models available through the API

Read the OpenRouter catalog as a collection of model routes with serving and routing metadata. Distinguish concrete identifiers, variants and moving aliases before building an evaluation.

Model families at OpenRouter

OpenRouter brings different model developers and serving arrangements into one catalog. Group candidates by the task your application needs rather than assuming the catalog is a single family with uniform behavior. Text generation, embeddings, media input and media output should begin as distinct eligibility checks.

The catalog records input and output modalities, tokenizer information and supported request parameters for its model entries. Official documentation.

Use these attributes to reject incompatible candidates early. Then inspect the actual endpoint and feature guide for a mandatory requirement such as structured output or tool use. A catalog checkbox should lead to a focused test of the required behavior, not to a broad claim that every feature works identically across all routes.

Keep the model developer separate from the serving provider. If you compare a direct API with an OpenRouter route for a similarly named model, record both endpoint identities and account terms. The model family may be related while the operating controls and diagnostic path differ.

All models

Copy an API identifier instead of a display label. The published record provides the source context for its capabilities and prices. If the record lacks a field necessary for your decision, leave it unresolved or investigate the official endpoint data before selecting it.

Export a shortlist with the exact identifier, host, required capabilities and source date. Add separate columns for observed output quality and usage. This makes later catalog refreshes possible without overwriting the evidence produced by your own evaluation.

Cheapest / strongest / longest-context

A broad catalog makes it tempting to treat price and context as universal quality scores. They are measurable selection attributes, but neither proves that a candidate will satisfy the application. Define an accepted result and measure cost per accepted task for the final shortlist.

For a coding task, require a patch that passes the relevant checks. For grounded document work, require a supported answer and a clear missing-evidence state. For structured extraction, validate the schema and field meaning. These task-specific criteria give a strongest recommendation a concrete meaning that a general catalog ranking cannot supply.

When several candidates tie on a recorded attribute, preserve the tie rather than inventing precision. Use the next relevant requirement or an actual evaluation to decide. A transparent shortlist with uncertainty is more useful than an unsupported universal winner.

Deprecations and replacements

Inspect model status and any documented endpoint retirement before updating production. A provider can change a serving option while the underlying family remains available elsewhere. Decide whether an alternative host is acceptable before treating that availability as a seamless replacement.

Search stored presets, fallback lists and scheduled jobs for the retired route. It is easy to update the visible primary model while an old fallback remains dormant until an incident. Exercise the fallback in a controlled local harness and run a small authorized verification when deploying its replacement.

Keep historical pages clearly marked and retain the migration rationale. A replacement should be selected on compatible behavior and operating conditions, not only a similar name. Preserve the old fixture so a response-format or tool-behavior regression remains detectable.

Naming and versions

Latest-family slugs beginning with a tilde resolve to a current concrete model, and the response model field identifies the selected version. Official documentation.

Choose a moving alias only when the application accepts that continuity tradeoff. A regression suite that needs a fixed target should use a concrete slug and an explicit upgrade process. If an alias is appropriate, monitor its returned model identity and rerun the relevant fixtures when the target changes.

A free variant identifies a documented free route, while the separate free router can choose among eligible models. Official documentation.

Do not infer the availability of a variant by appending a suffix to an arbitrary model name. Verify it in the current catalog. Keep pricing and capability assumptions attached to the actual variant, because the unsuffixed route is a different selection.

Review route pricing, free access conditions and OpenRouter versus direct calls before finalizing the integration.

Use the AI API cost calculator to turn the model and workload you are considering into an estimate.

Last verified 12 Sep 2026 · Source ↗

Frequently asked questions

Sources

• OpenRouter quickstart ↗
• Models and pricing schema ↗
• Authentication ↗
• Credit and rate limits ↗
• Errors and debugging ↗
• Provider routing ↗
• Free model variants ↗
• Free router ↗
• Workspace budgets ↗
• Prompt caching ↗
• BYOK ↗
• Provider logging ↗
• Batch quickstart ↗
• Latest aliases ↗

Last verified 12 Sep 2026 · Source ↗