How these rankings are built

Where the list comes from, what each figure means, and how often it is updated.

Back to the rankings

Image / Video benchmark scope

Image and Video cite public preference evaluations from Arena and Artificial Analysis. Operators are deduplicated: multiple task boards from one operator are not multiple independent operators or comprehensive capability certification.

Text to image, image editing, text to video and image to video have separate views. Artificial Analysis video boards with audio form separate pools. Arena audio conditions follow source entries; we do not claim a uniformly silent test. Each board orders independently, without combining tasks, audio conditions or quality settings.

Cells preserve source entries, retrieval dates, captured sample counts and intervals. Preliminary is a source flag, not our rating. Missing samples are not zero; overlapping intervals do not establish a significant lead.

Third-party prices and generation times have separate tabs and do not affect quality coverage or order. Prices retain units and configurations. Time names the provider and test conditions: a three-day median of successful requests, including download but excluding retries. Neither is a ChinaAPI quote or endpoint test, nor evidence that our models have not drifted.

How sources are chosen

Only first-party leaderboards anyone can open: the same address, the same number, checkable by anyone. Every figure records the page it was read from and the date it was read.

Nothing is scraped at request time. The data lives in a file in the repository, and an update is a Pull Request that moves the snapshot date, so what changed is visible in the diff.

No third party's composite score or aggregate ranking is reproduced wholesale. Their selection, arrangement and weighting are their own work; we quote only checkable original figures, one row at a time.

What the figures mean, and how the table is ordered

Text uses AIHOT’s open-source v15 algorithm: fixed-weight comparisons of jointly evaluated models, ordered to minimize overall disagreement. Our curated models and verified inputs differ from AIHOT’s, so scores and ranks differ.

Fixed budgets: broad evaluations 30%, human preference 10%, specialized evaluations 60%. Missing data is not zero and its weight is not redistributed. Available sources and weights are listed below.

Eligibility requires at least 3 sources, 3 evaluation families, 3 operators, 3 specialty categories and 2 direct anchor models, within an 18-month release window. Models with insufficient evidence remain visible without a rank.

Each source selects a model entry by fixed configuration priority, not its highest score. Kemeny consensus determines rank; the 0–100 display score uses adjacent-reversal costs and fixed anchors, not accuracy. Filters preserve overall ranks.

Sensitivity ranges recompute after removing a source or operator, shifting weights by ±20%, or ignoring intervals. They are not statistical confidence intervals.

A Terminal-Bench score depends on the model and on the agent scaffold that drove it, so the figure has to say which agent produced it. This page cites the column Artificial Analysis runs with one harness across every model — mini-SWE-agent v2.4.6, 66 tasks, pass@1 averaged over 3 repeats per task — and the harness is named in that column's source note. The benchmark's own board was cited before by taking each model's best row, which could land different models on different harnesses; with one harness for the whole column, the comparison is between the models themselves rather than each model plus its best scaffold.

MLCR-AA, cited in the Medical view, comes from the same operator as the primary board (Artificial Analysis) but is not a component of it — Artificial Analysis marks it a standalone evaluation outside the Intelligence Index, so it is a genuinely independent line of evidence, unlike the two columns above. The underlying benchmark is Wisedocs' open MLCR: multi-document synthesis over synthetic medical records of roughly 25,000–64,000 tokens, reconstructing chronology, causality, treatment patterns and claim relevance, with questions graded across six tiers. It measures reasoning over long records, not medical capability in general, and it is not medical advice.

A vendor's own figure appears only where no independent board covers the model at all. It is labelled as such and never orders the table.

Where a board's row label differs from the model name we serve — a preview build, a pro build, a different resolution tier — the board's own label is printed under the cell. We do not quietly reassign someone else's score.

The “Released” column is the vendor's announced release date for the model, not the date it went on sale here: it comes from the vendor's own release announcement, and model names that carry a date suffix (such as -260628) use the date the suffix encodes. Dates recorded by third-party boards such as Epoch only corroborate and are not used directly. When no substantiated date can be found the cell stays “—” rather than a guess.

Scope

The first edition covers text and multimodal, image generation and video generation. Speech synthesis and transcription are left out: the public boards this page accepts do not cover them, and a blank is better than an invented position.

The collection includes Chinese models sold here, verified new models under observation, and international reference models. Inclusion does not mean availability on ChinaAPI; open weights do not imply a permissive license. Models with insufficient evidence remain unranked. International models provide context for the standing of Chinese models.

Disclaimer

This page is not our own evaluation. We took no part in designing, running or scoring any of these tests.

Positions change as the sources update. This page holds only for the snapshot date shown at the top.

A public benchmark measures performance on the tasks it tests, which may not match your own. Choose a model by testing it on your own use case.

The names, trademarks and content of each source belong to their respective owners; this page only cites and explains them.

Update cadence

There is no scheduled job and no automatic sync. Data changes land only through a Pull Request: the sources are read again, the figures and retrieval dates are updated, the snapshot date moves forward, and it ships with the next release once reviewed.

That means the page can lag a board's latest refresh. We would rather be checkable and dated than live and unreviewed.

The “Maintained” date above the table follows the roster alone: it moves when a model or a cited source is added, and a figures-only re-read leaves it untouched — that is the snapshot date's job. The date is tied to the roster counts, and the build fails when they disagree, so it cannot quietly go stale.

Public sources cited on this page

  • LMArena Text Arena
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (text_style_control, category overall), published 2026-10-02, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • Artificial Analysis Intelligence Index
    Artificial Analysis · Intelligence Index · Intelligence Index v4.3.2, read 2026-09-22; its figures are not comparable with the v4.3 figures this page quoted from its 2026-09-12 read. v4.3.1 refreshed the pairwise judge panels of AA-Briefcase and GDPval-AA, and v4.3.2 moved GDPval-AA to v2.1 (Elo scale anchored to DeepSeek V4.1 Flash (max) at 1600, fitted with a Crowd-BT model) and refitted AA-Briefcase v1.1 with Crowd-BT, so rows moved without any model changing. v4.3 had already swapped τ³-Banking out of the index for AutomationBench-AA and Terminal-Bench v2.1 for v4.0. The leaderboard's default Status: Current filter hides rows Artificial Analysis marks deprecated; they still render with Status set to All and are re-read from that view like any other row (CA-948). A trailing * is reproduced as the board prints it: the board's data marks that score as estimated (intelligenceIndexIsEstimated), i.e. not every evaluation in the index had been run for that model · retrieved 2026-09-22
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • SuperCLUE 智能指数
    SuperCLUE · 总分 (overall) · 2026-07 evaluation, published 2026-08-06. The homepage reads this edition from its data file (/data/generalboard/2026年7月.xlsx, sheet 总排行榜, last modified 2026-09-11), which also carries rows evaluated after publication, each with its own evaluation publication date (the latest 2026-09-11); the page's Latest Update line refers to a different sub-board · retrieved 2026-10-04
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Epoch Capabilities Index (ECI)
    Epoch AI · General ECI · live board · retrieved 2026-10-04
    Licence: CC BY 4.0, as labelled on the page (https://creativecommons.org/licenses/by/4.0/)
  • Vals Index
    Vals AI · Accuracy (GDP-weighted finance, coding, legal and tax) · V2.2, page updated 2026-10-22; complete current cohort read 2026-10-23. Finance Agent v2.1 replaces v2 in this index; figures are not comparable with the V2.1 snapshot. Missing exact builds are not carried forward. · retrieved 2026-10-23
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Vendor's own evaluation card
    the model's own vendor · varies; the benchmark is named in each row · self-reported · retrieved 2026-09-22
    vendor-reported, never used for ordering
  • LMArena Text-to-Image Arena
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (text_to_image, category overall), published 2026-09-24, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • LMArena Image Edit Arena
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (image_edit, category overall), published 2026-09-30, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • LMArena Text-to-Video Arena
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (text_to_video, category overall), published 2026-09-22, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • LMArena Image-to-Video Arena
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (image_to_video, category overall), published 2026-09-22, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • LMArena Text Arena · Chinese
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (text_style_control, category chinese, prompts in Chinese), published 2026-10-02, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • LMArena Text Arena · Coding
    LMArena (arena.ai) · Arena score (Elo) · Hugging Face dataset lmarena-ai/leaderboard-dataset (text_style_control, category coding), published 2026-10-02, revision 46919c467f7f. Each rating is rounded half-up to a whole number and its interval is rating_upper minus rating, as arena.ai prints them; the dataset carries no Preliminary badge, so each citation records its row's vote count instead, and row names follow the dataset, which can trail a rename on arena.ai · retrieved 2026-10-05
    Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), as labelled on the dataset card (https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset)
  • Artificial Analysis · Output Speed
    Artificial Analysis · Median output tokens/s · live board; measured by Artificial Analysis against the vendor's own API, not through ChinaAPI. Rows Artificial Analysis marks deprecated are read with the leaderboard's Status filter set to All (CA-948) · retrieved 2026-09-22
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis · GDPval-AA v2.1
    Artificial Analysis · Agentic Real-World Work Tasks, (Elo-500)/2000 (%) · GDPval-AA v2.1, live board, read 2026-09-22. 220 agentic task-completion tasks with file outputs, ranked pairwise by a panel of three frontier LLM judges; since v2.1 the Elo scale is anchored to DeepSeek V4.1 Flash (max) at 1600 and fitted with a Crowd-BT model, so its figures are not comparable with the v2 figures this page quoted from its 2026-09-12 read. The column prints clamp((Elo-500)/2000). It carries 10% of the Agents category inside Intelligence Index v4.3.2, so this column is a component of the AA Intelligence Index shown as an individual source, not an independent operator. Rows Artificial Analysis marks deprecated are read with the leaderboard's Status filter set to All (CA-948). The column is hidden until the leaderboard's Intelligence group is expanded · retrieved 2026-09-22
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis · Terminal-Bench v4.0
    Artificial Analysis · Agentic Coding & Terminal Use (%) · Terminal-Bench 4.0 (66 tasks), run by Artificial Analysis with the mini-swe-agent harness, pass@1 averaged over 3 repeats per task; read 2026-09-22. When this column replaced the benchmark's own board on 2026-09-07 the methodology page named the harness as mini-SWE-agent v2.4.6; on 2026-09-21 it and the evaluation page name mini-swe-agent without a version, and every figure quoted before then read back unchanged. One harness for every model, so the column compares models rather than model-plus-best-scaffold — which is why it replaced the benchmark's own board, whose rows are agent × model and covered one Chinese model out of 34 (CA-788, overturning CA-734). It carries 10% of the Coding category inside Intelligence Index v4.3.2, so this column is a component of the AA Intelligence Index shown as an individual source, not an independent operator. Rows Artificial Analysis marks deprecated are read with the leaderboard's Status filter set to All (CA-948). The column is hidden until the leaderboard's Intelligence group is expanded · retrieved 2026-09-22
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis · MLCR-AA
    Artificial Analysis · Medical Long Context Reasoning (%) · MLCR-AA, live board, read 2026-09-22. Artificial Analysis marks it a standalone evaluation that is NOT part of Intelligence Index v4.3 — unlike the GDPval-AA v2 and Terminal-Bench v4.0 columns, this one is not a component of the AA Intelligence Index shown as an individual source, only the same operator. The underlying benchmark is Wisedocs' open MLCR (Wisedocs-AI/medical-long-context-reasoning): synthetic medical records of roughly 25,000-64,000 tokens, graded across six tiers from locating a single fact to expert-level clinical synthesis. It measures multi-document reasoning over long records, not medical capability in general. The main leaderboard carries no column for it; the figures were first read off the evaluation page's chart (2026-09-07) and since 2026-09-22 from the per-model pages' payload, which carries every model the chart can show (CA-956) · retrieved 2026-09-22
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis Text To Image
    Artificial Analysis · Elo · live board; exact row settings retained · retrieved 2026-09-13
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis Editing
    Artificial Analysis · Elo · live board; exact row settings retained · retrieved 2026-09-13
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis Text To Video · With Audio
    Artificial Analysis · Elo · live board; with audio pool; exact row settings retained · retrieved 2026-09-13
    That page states no reuse licence, so this page cites it as a source and links back to it.
  • Artificial Analysis Image To Video · With Audio
    Artificial Analysis · Elo · live board; with audio pool; exact row settings retained · retrieved 2026-09-13
    That page states no reuse licence, so this page cites it as a source and links back to it.

Data file

All of this page's data is available in machine-readable form: /api/model-rankings (2026-10-23)