How ModelsAtlas Works

Data sources, pricing normalization, capability signal methodology, and freshness approach — explained transparently.

Data sources

Pricing normalization

All pricing on ModelsAtlas is normalized to USD per 1 million input tokens. This is the standard unit used by OpenAI, Anthropic, Google, and most other providers. Output token pricing is shown separately as "USD per 1M output tokens."

Per-call fees

Some models charge additional per-call fees (image processing, audio generation). These are shown separately from per-token pricing. When a model has a per-call fee, it is labeled explicitly and the pricing column shows "per-call" instead of a per-token rate.

OpenRouter vs. direct provider pricing

OpenRouter adds a markup layer on top of provider pricing. Calling Anthropic's API directly is typically 5–15% cheaper than through OpenRouter. ModelsAtlas shows OpenRouter pricing because it is the most accessible entry point for developers integrating multiple providers. If you are building a production system with a single provider, check their direct pricing — it will usually be lower than what is shown here.

Capability signals

The capability scores shown on model pages (reasoning, coding, creative) are directional estimates derived from model metadata, not official benchmark results. They should be used for quick filtering, not as authoritative rankings.

The scoring methodology uses a weighted combination of:

• Context window size — larger contexts correlate with more capable base models
• Reasoning parameter presence — models with dedicated reasoning parameters score higher on the reasoning axis
• Tool-calling support — function calling, structured outputs, response_format parameters add points to coding and agentic axes
• Modality flags — image, audio, and video support add points to the creative axis
• Description analysis — keyword matching in model descriptions (e.g., "code", "reasoning", "creative") adds contextual signal
• Price-based adjustment — lower-priced models with strong capability signals get a small bonus (value-for-money signal)

These scores are bounded between 72–98 to avoid unrealistic extremes. They are recalculated every time the catalog refreshes, so they will shift as model metadata updates.

For official benchmark data, consult LMArena (for community-reported results) or provider documentation (for internally-validated scores).

Model ranking approach

Catalog ordering

The default model catalog ordering is relevance-based, prioritizing well-known providers (OpenAI, Anthropic, Google) and models with complete metadata. Users can sort by newest, alphabetical, price-low, price-high, or context-high.

"Best for" pages

The "Best for" pages rank models by use case. Rankings are derived from capability signals, pricing, and context window data. Each "best for" page includes a fit score (0–100) combining these factors, plus editorial commentary on the top 3 picks.

Comparison pages

Curated comparison pages (/compare) are hand-written with editorial verdicts. Auto-generated comparison pages (/compare/{provider}/{modelA}/vs/{provider}/{modelB}) use a structured axis comparison with winner flags but no editorial content.

Freshness approach

The catalog uses Next.js ISR (Incremental Static Regeneration) with a 6-hour revalidation window. This means:

• Pricing and model data refresh every 6 hours automatically
• New models appear within 6 hours of being added to OpenRouter
• Price changes are reflected within the same 6-hour window
• Individual pages can be revalidated on-demand via API call

The Phase 5 freshness daemon (GitHub Actions cron) will add a weekly changelog digest documenting major changes — new model launches, significant price reductions, deprecations, and new capability flags.

Uniqueness and quality standards

All generated pages (comparison pages, "best for" rankings, alternatives pages) are checked against a shingled Jaccard uniqueness threshold of ≥30%. This means any two generated pages must share fewer than 30% of their 3-shingled n-grams. Pages that fall below this threshold are flagged for editorial review.

Comparison pages include unique editorial content — a primary verdict, capability axis explanations, and FAQ entries — that distinguishes them from templated listings.

Thin or near-duplicate content is a quality risk that Google and AI citation systems both penalize. The uniqueness validator is run as part of the content quality audit.

What ModelsAtlas does not do

• Official benchmarks — ModelsAtlas does not run benchmarks. Capability scores are metadata-derived estimates.
• Latency testing — Response time depends on server load, deployment region, and network conditions. These are too variable to track.
• Fine-tuning recommendations — Fine-tuning requires deep integration with your training data and use case. ModelsAtlas tracks base model pricing, not fine-tuning costs.
• Provider reliability ratings — Uptime and service quality are tracked informally through user feedback, not systematically.