Model ranking & recommendations

On Discover → AI Models, models are not only listed alphabetically by default: MediaSnap can rank them using quality, speed, stability, and price signals so you can quickly find a good fit.

What you see in the product

Sort by (Balanced, Cheapest, Fastest, Best Quality)

Use the Sort by control above the model grid:

Changing the sort reorders the same cards; it does not hide models. Your other filters (type, category, search) still apply to visibility.

Badges on cards

Some models may show recommendation badges:

• Best overall — Top score under the balanced weights (good default pick when you are unsure).
• Best value — Strong combination of quality and low relative price (quality per credit / dollar).
• Fastest — Highest speed score in the catalog among models on the page.
• Premium quality — Quality score is in the high tier (fine for finals or polished work).

Badges describe how a model compares to the rest of the active catalog on the site, not a guarantee of results for every prompt.

Who decides the ranking?

There is no manual ordering of the list and no third-party AI picking winners in real time.

• MediaSnap (catalog / admin) sets each model’s quality, speed, and stability scores and its price fields in the database—these reflect how we describe and price the model in the product.
• The site’s software then computes a ranking score from those values (plus fixed rules for each “Sort by” mode) and sorts or labels models accordingly.

If optional success-rate fields are filled from real generation data in the future, those can influence stability in the formula; until then, rankings follow catalog scores and price only.

Scores in the database

Each model can store quality, speed, and stability scores (0–100). Older data may still use a 1–5 scale; the system maps those to a 0–100 scale for ranking.

How it works (technical)

Ranking score

For a given mode (balanced, quality, speed, budget), the score is:

ranking_score = wQ × quality_norm + wS × speed_norm + wSt × stability_norm + wC × cost_norm

• quality_norm, speed_norm, stability_norm — Normalized from ai_models.quality_score, speed_score, stability_score. Missing values default to a neutral midpoint (50). Whole numbers 1–5 are treated as legacy and scaled to 20–100.
• cost_norm — 0–100, from the inverse of effective price across the current result set: cheaper models get a higher cost score. Effective price is api_sale_per_call if set and positive, otherwise api_cost_per_call. Models without a usable price get the lowest cost contribution in that set.

Default weights (balanced)

Other modes (canonical)

• Quality — Higher weight on quality; others reduced.
• Speed — Higher weight on speed.
• Budget — Higher weight on cost (favors cheaper models in the set).

Aliases accepted in the API: best_quality → quality, fastest → speed, cheapest → budget.

Optional: live success rate

If ai_models.ranking_success_rate (0–100) is populated, stability used in the formula blends catalog stability with success rate so rankings can reflect real-world outcomes once you wire updates from jobs or analytics. See website/database/add_ai_models_ranking_optional_stats.sql.

API

• GET /api/models-ranked.php?mode=balanced — Returns active models with ranking_score, normalized subscores, and ranking_mode, sorted by score descending. Optional: media_type (e.g. image, video, and the same audio/character aliases as elsewhere).
• GET /api/models/ranked.php — Same implementation (path may depend on server URL rewriting).

The generator /api/models.php responses also include ranking fields enriched with balanced weights without changing the existing SQL sort order (so dropdowns stay stable).

Implementation files

• website/includes/ai-model-ranking.php — Formulas, weights, normalization, and helpers.
• website/app/api/models-ranked.php — Ranked list endpoint.

Engineering reference: documentation/19-ai-model-ranking.md (repo root).

Related

• AI Models — Browsing the catalog and cards.
• Choosing the Right Model — What to consider when picking manually.