7 active research sources, what they measure, and how much matching evidence they contribute to Penny Owl.
Penny Owl brings independent AI rankings together before comparing seller prices. Of our 7 listed sources, 7 contribute matching evidence to the models displayed here: 29 Brain, 30 image and 24 video models.
More matches ranks higher. The total adds distinct matching models across the three categories. Each exact model counts once per category, even when an organization has several boards or configurations. Ties share a rank. The bars show coverage order, not benchmark quality or reliability; sources with no matches appear afterward without a rank.
Coverage is one part of the picture
Human preferences, general capability tests, frontend coding and technical video tests answer different questions. A specialist board can be excellent for your job while covering fewer models here. Read the strengths and limits below. Evidence dates stay attached to the results; retained results can be older than the latest catalog update.
Current catalog snapshot with individual evidence dates.
These are coverage rankings for this catalog. Click a source’s official page to investigate its full results; its coverage can be much wider than our matching subset.
#1 · Design Arena
Why here: 72 matching models across 3 categories. More matches ranks higher; equal totals share a rank.
What it measures: Design Arena uses crowdsourced comparisons for frontend design and coding, image generation and video generation.
Good for: Useful for visual and creative work, including websites, interfaces and generated media.
Keep in mind: Its Brain evidence is specialist frontend/design performance, not a universal intelligence test. Media comparisons may not declare the same resolution or audio settings as a seller’s price.
Matching evidence checked: . A later read failed; earlier evidence is retained with its original dates.
Brain · included · Source review: Specialist frontend coding/design human preference: Overall Frontend (Text-to-HTML), Website/UI/Game/Data Visualization/3D tasks. Full 171 ordered published configurations audited; 157 eligible identified configurations / 142 exact versions enter normalization; 14 unrevealed or unpinned aliases excluded. Historical named results retained. Best documented effort per exact version contributes once; one DesignArena organization vote. Earlier results retained after a failed read. Reviewed board
Images · included · Source review: General image ranking: 82 entries. The identified Riverflow router is excluded from the model comparison. Earlier results retained after a failed read. Reviewed board
Videos · included · Source review: General video preference: 47 entries. Positions are derived from Elo scores; tested audio and resolution settings are not declared. Earlier results retained after a failed read. Reviewed board
#2 · Artificial Analysis
Why here: 70 matching models across 3 categories. More matches ranks higher; equal totals share a rank.
What it measures: Artificial Analysis combines language-model capability tests with separate human-preference image and video boards.
Good for: Useful for comparing general capability and media preferences in one research platform.
Keep in mind: Its intelligence index and media preference boards measure different tasks. Reasoning effort, video audio and exact model settings still need matching.
Brain · included · Source review: Intelligence Index: 194 verified score configurations for 123 exact versions. Estimated or unbound records excluded; tested effort settings vary. Reviewed board
Images · included · Source review: Complete 166-entry general text-to-image ranking. Reviewed board
Videos · included · Source review: Current 28-entry text-to-video ranking with audio. Only versions with verified native audio enter our comparison. Reviewed board
#3 · Arena
Why here: 69 matching models across 3 categories. More matches ranks higher; equal totals share a rank.
What it measures: Arena ranks responses and generated media using human pairwise preferences.
Good for: Useful for understanding which outputs people prefer in text, image and video comparisons.
Keep in mind: Preference depends on the users, prompts and tested settings. A high overall position does not guarantee the best result for your particular job.
Matching evidence checked: . A later read failed; earlier evidence is retained with its original dates.
Brain · included · Source review: Overall text preference: 413 configurations, published October 2. Repeated effort settings count once per exact version. Earlier results retained after a failed read. Reviewed board
Images · included · Source review: Complete current 81-entry text-to-image ranking. Search, preview and resolution settings are recorded. Earlier results retained after a failed read. Reviewed board
Videos · included · Source review: General text-to-video ranking: 48 entries. Silent, retired and unverified-audio versions are excluded. Earlier results retained after a failed read. Reviewed board
#4 · Rapidata Benchmark
Why here: 40 matching models across 2 categories. More matches ranks higher; equal totals share a rank.
What it measures: Rapidata Benchmark uses human judgments for versioned prompt sets. Penny Owl matches its image and SVG coding results.
Good for: Useful for image preferences and the visual quality of generated SVG graphics.
Keep in mind: SVG coding is a specialist Brain task. Separate physics or camera video tests are not interchangeable with a general native-audio video ranking.
Brain · included · Source review: Specialist SVG graphics coding human preference; unweighted standings. Full 63 published eligible configurations audited; 48 identified Brain configurations / 41 exact versions enter normalization; 15 SVG-only/unrevealed/unpinned identities excluded. Best documented effort per exact version contributes once; one Rapidata organization vote. Reviewed board
Images · included · Source review: General image preference: 41 active scored entries. Disabled and unscored results are excluded. Reviewed board
Videos · not-comparable · Source review: General video rankings are listed as coming soon. Physics and camera tasks are excluded. Reviewed board
#5 · Epoch AI
Why here: 25 matching models across 1 category. More matches ranks higher; equal totals share a rank.
What it measures: Epoch AI’s Epoch Capabilities Index combines results across capability benchmarks into a general model index.
Good for: Useful for seeing a broader capability picture instead of relying on one test.
Keep in mind: The underlying evaluations can use different settings and include results reported by developers. A general index can understate a specialist model’s strengths.
Brain · included · Source review: General Capabilities Index: 270 rows. Positions are derived from published ECI scores; benchmark settings vary. Reviewed board
Images · not-applicable · Source review: LLM capabilities; no general image-generation ranking. Reviewed board
Why here: 24 matching models across 1 category. More matches ranks higher; equal totals share a rank.
What it measures: LiveBench evaluates language models with tasks designed for objective scoring and reduced test contamination.
Good for: Useful for checking reasoning, coding and other capabilities against verifiable answers.
Keep in mind: Penny Owl uses the matched overall results it has verified. Test releases and model configurations differ; this is not a media-generation preference board.
Brain · included · Source review: Overall LiveBench quality: 66 configurations. The June 25 test release includes later model evaluations. Reviewed board
Images · not-applicable · Source review: LLM evaluation; no general image-generation ranking. Reviewed board
Why here: 17 matching models across 1 category. More matches ranks higher; equal totals share a rank.
What it measures: Vals AI evaluates language models on practical tasks across professional domains. Penny Owl imports its verified Vals Index results.
Good for: Useful for looking beyond generic chat toward work such as coding, finance and legal tasks.
Keep in mind: The mix and weighting of tasks matter. Penny Owl’s saved index evidence is not a claim that every newer Vals benchmark or release has been imported.
Brain · included · Source review: Vals Index v2.1: 43 models, updated October 2. Claude fallback settings are recorded; ambiguous Qwen Max versions stay separate. Reviewed board
Images · not-applicable · Source review: LLM and industry tasks; no general image-generation ranking found. Reviewed board
Videos · not-applicable · Source review: LLM and industry tasks; no video-generation ranking. Reviewed board
How Penny Owl combines model rankings
Each organization gets one equal vote for an exact model. Positions in the matching comparison set are normalized so first is 0 and last is 100, then averaged. Results backed by fewer organizations are pulled toward the middle, so thin evidence cannot win by luck; lower adjusted scores rank higher. At least two matching organizations are needed. Multiple boards from the same organization do not create extra votes. Missing, incomparable and unverified versions are skipped. This page’s coverage order and seller prices do not change a model’s quality rank.
The existing daily process updates connected ranking boards. Failed ranking reads retain earlier evidence and its original dates; other boards need review. Opening this page is not a successful source check. Read the automatic and manual source map for the connected readers.
The leaderboard source coverage JSON contains the same order, counts, matched model identities, source reviews and definitions. The AI reading index helps search engines and AI assistants interpret and cite the data.
Common questions about AI leaderboards
How are the leaderboard sources ranked?
Sources with more matching displayed models come first. Each exact model counts once per category, even when a source has several boards or configurations. Equal totals share a rank. Sources with zero matches are listed afterward without a rank.
Does first place mean the most trustworthy leaderboard?
No. This order measures contribution to Penny Owl’s displayed models, not reliability, benchmark quality or popularity. Read each source’s strengths, limits and evidence dates before using its results.
Why can a respected benchmark have zero matches?
Its tests may cover different model versions, specialist tasks, historical models or incompatible audio settings. Zero means no matching evidence in this displayed catalog, not that the benchmark is bad or that it evaluates no models.
Does this source order change the combined model rankings?
No. Each matching organization still gets one equal vote. Penny Owl normalizes matching positions, averages organization votes and pulls thinly supported results toward the middle. At least two organizations are required for a combined model rank. Price, source coverage and the source order on this page do not change those votes.
Are all sources checked successfully every day?
No. Connected boards are checked by the existing refresh process. Failed ranking reads retain earlier evidence and its original dates. Other boards need manual review. Loading this page or exporting its JSON does not recheck a source.
