Skip to content
Model Wayfinder
Methodology

How the recommendations work

The recommendations follow fixed rules. No AI makes the decision, and there are no paid placements.

01 · Exclusion

First, models that fail a requirement are removed: a context window too small for the chosen text length (8,000 / 32,000 / 200,000 tokens), no image input for image analysis, no processing in the requested region, or no open weights for local use.

For local use we also check whether the model fits into your hardware’s memory. Estimate for the compressed version (Q4): 0.6 GB per billion parameters plus 1.5 GB headroom.

For work or team use, models whose licence forbids commercial use are excluded.

02 · Scoring

Every rated model has a quality tier from 1 to 3 per task. The tiers come from public rankings: the top third of a ranking is tier 3, the middle third tier 2, the bottom third tier 1. For each task we use the most specific ranking available:

Task Ranking (fallback)
Writing texts LMArena Creative Writing (Text)
Summarising documents LMArena Document (Longer Query, Text)
Chatbot LMArena Text
Programming LMArena Coding (WebDev)
Translating LMArena Non-English (Text)
Maths LMArena Math
Analysing images LMArena Vision
Analysing videos LMArena Vision – there is no public ranking for video understanding yet, so image understanding serves as an approximation
Creating / editing images LMArena Text-to-Image / Image Edit
Creating videos / photo → video LMArena Text-to-Video / Image-to-Video
Transcribing audio Hugging Face Open ASR Leaderboard (average of English and multilingual position)

The LMArena rankings are based on millions of blind comparisons by real users. The Open ASR Leaderboard measures the word error rate on fixed test recordings (lower = better). If you choose a specialist field (law, medicine, business, science), the matching LMArena subject ranking counts as an additional tier in the average. Models that appear in none of these rankings stay unrated. We correct individual tiers by hand where we have better sources, and say so on the model page. The price of a typical request is placed on a logarithmic scale, because prices differ by a factor of a thousand. Both are weighted by your priority:

Priority Quality Price
Cheap 30 % 70 %
Balanced 50 % 50 %
Best quality 80 % 20 %

From 10,000 requests per month, the weighting shifts 10 percentage points towards price, because small price differences start to add up. Ties are broken by the lower price, then the larger context window.

03 · Several tasks at once

You can select several tasks. Text and analysis tasks are answered by one model, which must handle all of them; its quality is the average of its tiers for those tasks. Images, videos and transcription each get their own podium, because hardly any model does everything. Images are priced per image, videos per 10-second clip, transcription per minute of audio. These prices come from the public LiteLLM price list; where several providers or quality levels are listed, we use the middle (typical) price, not the cheapest. Open speech recognition models without a cloud offer are only recommended for local use.

04 · Benchmarks

Benchmark scores use different scales (percentages, Elo points, error rates), so they can’t simply be averaged. We convert each score into a position from 0 to 100 among all models listed here that took the same benchmark (100 = best; for error rates the lowest value is best). A position is only shown when at least 3 models can be compared. The benchmark average is the mean of these positions. “Strengths” are areas with an average position of 75 or more, “weaknesses” 30 or less. Benchmarks don’t feed into the recommendation – it relies on our sourced quality tiers.

05 · Data sources

Models and prices: OpenRouter. Open weights and hosting providers (for regions): models.dev. Licence and model size of open models: Hugging Face. Rankings: LMArena and the Open ASR Leaderboard. Prices per image, per second of video and per minute of audio: LiteLLM. Exchange rate: European Central Bank. Everything is refreshed automatically every day; if a source is down, the last known values stay in place.

06 · Ads

If we show ads, they are clearly labelled, appear only in separate areas and never influence rankings or recommendations.

07 · Limits

Quality tiers are a simplification. Prices come from OpenRouter and may differ when buying directly from the provider or in other regions. Memory figures are estimates. Models without our own rating appear in the comparison, but not in the recommendations.

Some kinds of AI aren’t covered yet, e.g. music generation, 3D models and CAD, or speech output (text-to-speech): for these there is currently no public ranking with open data that we could use. We’d rather leave them out than recommend without sources.

Privacy notes are not legal advice. Check contracts (e.g. data processing agreements) with the provider yourself.