Skip to content

Research OpenRouter models

From a Pipelex development checkout, run make install to provision .venv, then use its development CLI:

.venv/bin/pipelex-dev openrouter list-models -f claude
.venv/bin/pipelex-dev openrouter list-models -f claude --json
.venv/bin/pipelex-dev openrouter list-models --sort context_length
.venv/bin/pipelex-dev openrouter list-models -f flux --save

It needs no hosted credentials or Pipelex boot. OPENROUTER_API_KEY is optional: when present, it is sent as a bearer token; otherwise the catalog request is anonymous. No inference calls are made.

The command reads the public models API with output_modalities=all, including image models, and follows catalog pagination. It does not use the frontend API. --category forwards OpenRouter's category filter. --filter matches a substring of the name or ID, ignoring case. --sort accepts name, created (newest first) and context_length (largest context window first); the table explicitly calls that column context tokens.

Markdown is the default. --json emits an envelope with ok, models, saved_to and skipped_model_ids, without a banner. Every model retains the catalog's supplied metadata. Expected failures emit ok: false and error on stderr in JSON mode, or Markdown on stderr otherwise. An empty lookup is a successful result with no models; attempting to save it is refused.

--save writes generated/openrouter_models.toml relative to the current directory. Category and name filters add sanitized suffixes, such as generated/openrouter_models_claude.toml. Existing files are never overwritten. JSON mode can be combined with saving; saved_to reports the generated path.

Exports are disabled draft entries. Review each entry before changing enabled to true. The exporter uses the current runtime's backend document validator, quotes model names safely and converts known token prices into USD per million tokens. Missing, negative or nonfinite prices remain absent; absent costs do not imply free inference. Non-token charges, image pricing, routing, reasoning configuration and provider restrictions require manual review. Context length is never exported as max_tokens: that field comes only from top_provider.max_completion_tokens. Tool support alone never implies structured output; only explicit structured_outputs metadata adds that capability. Export includes only models whose outputs are text or images. Unsupported output modalities are skipped and explicitly listed in skipped_model_ids in JSON or below the Markdown result; all matched models remain in the lookup. If no matched model is exportable, saving is refused.

Provider documentation remains the primary authority. OpenRouter is a cross-check and may lag a new model. When the development CLI is unavailable, fetch the public catalog directly and filter its data list:

curl -fsS 'https://openrouter.ai/api/v1/models?output_modalities=all'

Adding a hosted model remains a change to the hosted plane's own catalog; this tool never publishes models or modifies backend files automatically.