Check model pricing

Prices come from the live API, not memory. fal is the default backend and Replicate is the supported alternative. Preferred models live in references/preferred-image-models.json at the plugin root.

Check a price

Run from the plugin root (FAL_KEY and REPLICATE_API_TOKEN are in Infisical):

with-secrets python3 scripts/model_price.py google/nano-banana-2.1 fal-ai/birefnet/v2
with-secrets python3 scripts/model_price.py --search klein
with-secrets python3 scripts/model_price.py --backend replicate black-forest-labs/flux-schnell

How to read the result:

• unit: images is per image and megapixels is per MP (1024x1024 is about 1.05 MP, 512x512 about 0.26 MP). compute seconds depends on run time.
• placeholder: true: the API's unit_price of 1 is not a price. These models bill by tokens (GPT Image, Nano Banana 2.1, some others). page_price holds the text from the fal model page. Use the panel's per-image figure from the reference file, or work it out from that text.
• Replicate: the API has no price field (api_price is always null). page_prices is scraped from the model page and picks up unrelated figures, so treat it as a hint and read https://replicate.com/OWNER/MODEL for the real price.
• error: HTTP 404: the id is wrong or retired. Use --search.

fal generation responses carry no cost, so any total you report is an estimate from list prices. Say so.

Review the library and update preferred models

Do this when asked, or when the reference's updated date is more than about two months old and a real job depends on it.

1. Snapshot the catalogue with Tech-Graphics-Benchmark: with-secrets uv run python scripts/fal_catalogue.py writes catalogue/fal-YYYY-MM-DD.jsonl. Diff it against the previous snapshot for new endpoints.
2. Shortlist for the use case at about $0.05 per image or less, plus one reference model above that.
3. Try candidates small, with skill ai-eval/informal-image-eval if it needs a proper round. A few images on the actual job is enough for a provisional entry.
4. Edit the reference: change default, budget or premium, write what was seen in evidence, set status, and bump updated.

Rules for the reference:

• Evidence is per use case. Don't carry a text-fidelity result over to illustration or characters.
• status is evaluated, or provisional for a handful of tries. An agent read is not human annotation, so say which it was.
• Never replace a default on price alone. It needs at least one look at real output.