The free plugin ships four models: Claude 4.5 Haiku, Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite and GPT-4.1 mini. Pro adds three more to the same AI Model dropdown, one per provider. This page lists what each one costs, what its registry id is, and when the free four are the better call.
The three models Pro adds Pro
Models Pro registers
| Feature | Registry id | Est. cost / 1,000 images | What it is |
|---|---|---|---|
| Claude 4.5 Sonnet Anthropic. The most expensive model in the dropdown. | claude-sonnet-4-5 | ~$9.00 | Listed in the dropdown as Claude 4.5 Sonnet (Nuanced, brand-safe descriptions). About 4.5 times Claude 4.5 Haiku per image, which is the free Anthropic option. |
| Gemini 3.1 Pro Google. A preview model — see the caveat below. | gemini-3.1-pro-preview | ~$2.50 | Listed in the dropdown as Gemini 3.1 Pro (State-of-the-art multimodal reasoning). About 4.5 times Gemini 3.1 Flash-Lite, which is the default model on a fresh install. |
| GPT-5 mini OpenAI. The cheapest of the three. | gpt-5-mini | ~$1.20 | Listed in the dropdown as GPT-5 mini (Cost-efficient reasoning for structured tasks). Under twice GPT-4.1 mini, the free OpenAI option. |
The registry id is the stable handle, not the display name. It is what the plugin stores in infyp_optimization_settings.ai_model and what PATCH /wp-json/infyp/v1/settings accepts. Display names get renamed between releases; ids do not.
Pro sets no default model
Activating a license does not change which model runs. The default stays gemini-3.1-flash-lite-preview — the free plugin’s INFYP_DEFAULT_MODEL — until you pick something else in the dropdown. Nothing gets more expensive because a license was entered.
How the seven models compare
With Pro active the dropdown holds seven models, grouped by provider. Every entry carries a blended per-million-token price in the model registry, which is where the cost column below comes from.
Every model in the AI Model dropdown
| Feature | Tier | Registry id | Est. cost / 1,000 images |
|---|---|---|---|
| Claude 4.5 Haiku Anthropic. Fast, cost-effective analysis. | Free | claude-haiku-4-5 | ~$2.00 |
| Claude 4.5 Sonnet Anthropic. Nuanced, brand-safe descriptions. | Pro | claude-sonnet-4-5 | ~$9.00 |
| Gemini 3.1 Flash-Lite Google. Most cost-effective frontier model — the default on a fresh install. | Free | gemini-3.1-flash-lite-preview | ~$0.56 |
| Gemini 3.5 Flash-Lite Google. Newer generation, cost-effective. Added in free 1.0.6. | Free | gemini-3.5-flash-lite | ~$0.85 |
| Gemini 3.1 Pro Google. State-of-the-art multimodal reasoning. | Pro | gemini-3.1-pro-preview | ~$2.50 |
| GPT-4.1 mini OpenAI. Cost-effective image analysis. | Free | gpt-4.1-mini | ~$0.70 |
| GPT-5 mini OpenAI. Cost-efficient reasoning for structured tasks. | Pro | gpt-5-mini | ~$1.20 |
The arithmetic behind that column is deliberately simple. Each model stores one blended price per million tokens, and a single optimization — image in, short JSON out — runs at roughly 1,000 tokens. A thousand images is therefore about a million tokens, so the price per million reads directly as dollars per 1,000 images.
Cost numbers are estimates
Token counts move with image resolution and prompt length, and the prices live in the plugin source rather than being fetched from the providers — check them against the provider’s own pricing page before you budget a large run. The API usage dashboard shows what your library actually logged.
Pro makes three AI calls per image, not one
With a valid license, a full optimization also writes the caption and the description, and each of those is a separate call to the same model. Budget roughly three times the figures above for any run that uses the full optimize path. See captions and descriptions.
Two of the Gemini models are preview releases
Two ids end in -preview: gemini-3.1-flash-lite-preview, the free default, and gemini-3.1-pro-preview, the Pro one. That suffix is part of the id sent to Google, not a label the plugin adds. claude-sonnet-4-5 and gpt-5-mini carry no such suffix.
Preview access is the provider’s to grant. It comes with tighter rate limits than a general-availability model, and it can be refused outright on a project that has no billing set up. The plugin passes the refusal through rather than falling back to another model.
What a refused request looks like
| Feature | Message the plugin shows |
|---|---|
| HTTP 403 Access to the model was refused. | Invalid API key or insufficient permissions. (HTTP 403) |
| HTTP 429 Rate limit hit on the provider side. | Rate limit exceeded. Please wait and try again. (HTTP 429) |
Test a preview model on one image before a bulk run
A 403 does not stop a bulk run — every remaining image fails the same way and gets counted as a failure. Optimize a single image first, confirm it comes back, then start the run.
When a bigger model is worth the cost
Output varies between runs, so a side-by-side on your own images beats any general claim about which model writes better alt text. A dry run generates without persisting anything, which makes the comparison free of consequences — though not free of cost, since the AI call still happens.
# Generate without persisting, so the same image can be run twice.
# Switch the AI Model dropdown between runs — there is no --model flag.
wp infyp optimize 42 --dry-run --user=admin--dry-run is single-image only, and every write command needs --user. The same comparison works over REST with dry_run on POST /images/{id}/optimize. Both are covered in the WP-CLI reference and the REST API reference.
Mixing cheap and Pro models on the same site
The AI Model dropdown is global. One setting drives every path: single-image runs, the browser bulk runner, server-side jobs, WP-CLI and REST. There is no per-run override — POST /jobs takes no model parameter and wp infyp optimize has no --model flag. Switching between runs is the workaround.
Run the easy work on a cheap model
Leave the AI Model on Gemini 3.1 Flash-Lite, the cheapest of the seven. Bulk-optimize the backlog of straightforward images — product shots, blog hero images, library photos.
Switch to a Pro model for the hard ones
Filter the image table down to the problem images — low scores, complex subjects, key landing pages. Switch the model to Claude 4.5 Sonnet, Gemini 3.1 Pro or GPT-5 mini, then bulk-optimize that filtered subset.
Switch back
Set the dropdown back to the cheap model. Later single-image runs, bulk runs and jobs pick it up from the same setting.
Watch the dropdown if real-time upload is on
Real-time upload optimization reads the same setting, so every new upload runs on whatever model is selected right now. Pick the model you want fresh uploads to use day to day, and switch back after a catch-up run on an expensive model.
Next steps
AI Settings metabox
Where the AI Model dropdown lives, alongside language and the prompt add-on.
API usage and costs
What your library actually spent, per model and per provider, instead of an estimate.
Bulk Optimization (Pro)
The browser bulk runner — where a model switch costs the most, because it multiplies by every selected image.