Chinese AI Model Rankings
As of 2026-10-23, Claude Opus 5.5 leads the ChinaAPI Consensus v15 at 73.7; the highest-placed Chinese model is Qwen3.8-Max at 59.9 (#12). 84 of the 107 models listed can be called through ChinaAPI.
Where the Chinese AI models ChinaAPI serves stand next to the international frontier, quoted from public leaderboards. Snapshot of 2026-10-23; roster last maintained 2026-10-23 (moves when a model or source is added). Every figure is reproduced from the board named in its column and links back to it. ChinaAPI runs no evaluation of its own. Text uses the open-source AIHOT v15 consensus algorithm on ChinaAPI's curated roster; media and scene columns retain their source-specific metrics. See how this list is built, the price of every model marked available, and the same data as JSON at /api/model-rankings.
Sources
- 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, 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
- 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
- Epoch Capabilities Index (ECI) — Epoch AI, General ECI, live board, retrieved 2026-10-04, 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
- Vendor's own evaluation card — the model's own vendor, varies; the benchmark is named in each row, self-reported, retrieved 2026-09-22
- 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, 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, 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, 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, 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, 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, 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
- 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
- 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
- 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
- Artificial Analysis Text To Image — Artificial Analysis, Elo, live board; exact row settings retained, retrieved 2026-09-13
- Artificial Analysis Editing — Artificial Analysis, Elo, live board; exact row settings retained, retrieved 2026-09-13
- Artificial Analysis Text To Video · With Audio — Artificial Analysis, Elo, live board; with audio pool; exact row settings retained, retrieved 2026-09-13
- Artificial Analysis Image To Video · With Audio — Artificial Analysis, Elo, live board; with audio pool; exact row settings retained, retrieved 2026-09-13
ChinaAPI Consensus v15
41 eligible models; available budget 52%. Missing weights are not redistributed. Admission requires 3 sources, 3 evaluation families, 3 operators, 3 specialty categories and 2 direct anchors within an 18-month release window. Insufficient evidence means unranked, not last. Scores are display values, not accuracy percentages. This uses a different candidate and evidence pool from AIHOT.
- artificial-analysis: 30.00% (unavailable) — Artificial Analysis API access is not configured; its weight remains unused.
- artificial-analysis-multilingual: 6.00% (reserved) — Reserved by v15; no comparable evidence yet
- Arena Text Style-Controlled: 5.00% (available)
- Arena WebDev: 2.40% (available)
- Mercor APEX-Agents 1.1: 4.00% (available)
- Vals Finance Agent: 3.00% (available)
- DeepSWE v1.1: 3.60% (available)
- TapTap Maker Benchmark: 3.60% (available)
- livebench-coding: 2.40% (unavailable) — LiveBench 2.0 replaces the v1 category pairs used by v15. Its new categories need a reviewed mapping; the existing weights remain unused.
- livebench-reasoning: 4.20% (unavailable) — LiveBench 2.0 replaces the v1 category pairs used by v15. Its new categories need a reviewed mapping; the existing weights remain unused.
- livebench-writing: 3.00% (unavailable) — LiveBench 2.0 replaces the v1 category pairs used by v15. Its new categories need a reviewed mapping; the existing weights remain unused.
- Arena Creative Writing: 5.00% (available)
- Arena Vision Style-Controlled: 6.00% (available)
- tau-banking: 2.00% (reserved) — Reserved by v15; no comparable evidence yet
- SimpleQA Verified: 4.00% (available)
- GPQA Diamond: 2.00% (available)
- FrontierMath v2 · Tiers 1–3: 3.60% (available)
- FrontierMath v2 · Tier 4: 1.20% (available)
- Chess Puzzles: 1.80% (available)
- Mystery Game Puzzles: 1.20% (available)
- Creative Writing v3: 3.60% (available)
- Longform Writing: 2.40% (available)
Text and multimodal models
Ordered by ChinaAPI Consensus v15. Unranked entries have no numerical position; their discovery order follows release date.
| # | Model | Vendor | Origin | Released | On ChinaAPI | ChinaAPI Consensus v15 | Consensus evidence | LMArena Text Arena | Artificial Analysis Intelligence Index | SuperCLUE 智能指数 | Epoch Capabilities Index (ECI) | Vals Index | Vendor's own evaluation card | LMArena Text Arena · Chinese | LMArena Text Arena · Coding | Artificial Analysis · Output Speed | Artificial Analysis · GDPval-AA v2.1 | Artificial Analysis · Terminal-Bench v4.0 | Artificial Analysis · MLCR-AA | Notes | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| New model watch — released in the 30 days ending 2026-10-23; awaiting evidence, no overall rank | ||||||||||||||||||||||||||||||||
| — | Kimi K3.1 Pro | Moonshot AI | China | 2026-10-23 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | — | — | — | — | — | — | — | Released 2026-10-23, as dated in Kimi Code's official changelog (https://www.kimi.com/code/docs/en/kimi-code/whats-new.html). The official technical blog (https://www.kimi.ai/blog/kimi-k3-1-pro) identifies it as the first K3.1-series model, a post-trained version of K3 with a 1M-token context window, available in Kimi products and via kimi-k3.1-pro in the Kimi API. Full model weights are promised by 2026-11-04; that is a future release commitment, not evidence that weights are already public. The official Moonshot AI Hugging Face organization returned no K3.1 repository at this read. Highspeed is described as the same model capabilities at approximately 2-3x speed, so it is recorded as a serving variant rather than a second ranked model. No exact K3.1 Pro row is present in the 16 available consensus metrics checked on 2026-10-23; K3 and K2.x scores are not inherited and vendor-reported results do not supply independent evidence. Artificial Analysis is unavailable to consensus without its API key; absence here does not claim that no other evaluator has tested it. Listed in the public ChinaAPI pricing catalogue, verified 2026-10-24. | ||||||||||||
| — | MiniMax M3.1 Flash | MiniMax | China | 2026-10-19 | yes | — | Unranked: insufficient families. 3 sources, 2 families, 3 operators; 9.4% budget coverage.Source observations
Compare an earlier model: MiniMax-M3Same evaluation protocol only; configurations are stated for each entry. No overall rank is inferred.
| — | — | — | — | 53.78 ±1.07 | — | — | — | — | — | — | — | Announced 2026-10-19 in MiniMax's official release post (https://minimax.cn/blog/minimax-m31f). Official weights are available at https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash, verified 2026-10-23, under the MiniMax Community License. Native multimodal model for coding and agentic work; not an image generator. Only exact-build independent observations enter consensus; vendor-reported benchmark scores are not substituted. Listed in the public ChinaAPI pricing catalogue, verified 2026-10-24. Open weights · License: MiniMax Community License · Weights released: 2026-10-23 · Verified: 2026-10-23 | ||||||||||||
| ChinaAPI Consensus v15 | ||||||||||||||||||||||||||||||||
| 1 | Claude Opus 5.5 | Anthropic | International | 2026-09-22 | yes | 73.7 | 15 sources, 8 families, 6 operators; 48.8% budget coverage. Sensitivity rank range 1–2 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1504 ±9 | 58 | — | 167 (164 - 172) | 70.18 ±1.00 | — | 1571 ±29 | 1538 ±18 | — | 67 | 60 | — | Released 2026-09-22 per Anthropic's announcement (Introducing Claude Opus 5.5), the first model in the Claude 5.5 family. Artificial Analysis printed no output speed for its Claude Opus 5.5 (max with fallback) row on 2026-09-22 (the column reads --), so no speed figure is quoted; the speeds of its lower-effort rows are not substituted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 2 | Gemini 4 Argon | International | 2026-09-30 | yes | 73.5 | 12 sources, 7 families, 5 operators; 36.8% budget coverage. Sensitivity rank range 1–4 across 56 recomputation scenarios (not a confidence interval).Source observations
| — | — | — | — | 69.24 ±1.09 | — | — | — | — | — | — | — | Announced 2026-09-30 in Google's Gemini 4 Argon launch post (https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). Google describes an initial rollout to trusted cyber defenders through the Fairwind Program; broader developer and consumer availability was still forthcoming. This announcement date does not imply general API availability. Listed in the public ChinaAPI pricing catalogue, verified 2026-10-24. | |||||||||||||
| 3 | GPT-6 Astra | OpenAI | International | 2026-09-03 | yes | 72.7 | 16 sources, 9 families, 7 operators; 52.4% budget coverage. Sensitivity rank range 1–9 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1477 ±7 | 53 | — | 167 (163 - 171) | 65.02 ±1.10 | — | 1510 ±22 | 1543 ±13 | 58 | 52 | 59 | 35.0 | Released 2026-09-03 per OpenAI's announcement (GPT-6 Astra: A new generation of intelligence). LMArena added gpt-6-astra-max to its text board on 2026-09-11 per its leaderboard changelog; the row stands on 9,156 votes in its 2026-10-02 dataset publish. The Chinese category carried no row for it as of 2026-09-22; in the 2026-10-02 publish it does, on 674 votes. On sale here since 2026-09-05. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 4 | Claude Fable 5.1 | Anthropic | International | 2026-09-01 | yes | 72.6 | 13 sources, 7 families, 5 operators; 43.8% budget coverage. Sensitivity rank range 3–5 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1501 ±6 | 53 | — | 165 (162 - 169) | — | — | 1570 ±21 | 1531 ±13 | 65 | 62 | 52 | 71.1 | Released 2026-09-01 per Anthropic's announcement (Introducing Claude Fable 5.1 and Claude Mythos 5.1); on sale here since 2026-09-02. LMArena's row is the max-effort build, standing on 11,800 votes in its 2026-10-02 dataset publish; the same row on the Chinese category stands on 874 votes. Epoch's fit as read on 2026-10-04 scores it 164.82 and Claude Fable 5 162.22. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 5 | Claude Fable 5 | Anthropic | International | 2026-06-09 | yes | 71.7 | 14 sources, 8 families, 5 operators; 45.8% budget coverage. Sensitivity rank range 4–10 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1504 ±4 | 50 | — | 162 (160 - 166) | — | — | 1555 ±12 | 1552 ±7 | 61 | 55 | 42 | 64.4 | Artificial Analysis marked its Claude Fable 5 (with fallback) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. LMArena relabelled this row from claude-fable-5 to claude-fable-5-high between 2026-09-06 and 2026-09-18: the Internet Archive's copies of the board show the same model key (claude-fable-5-text) with its vote count running on from 27,189 to 30,057, so the entry follows the new label on all three LMArena text boards. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 6 | Gemini 3.7 Flash | International | 2026-08-13 | yes | 66.7 | 13 sources, 8 families, 5 operators; 43.4% budget coverage. Sensitivity rank range 5–14 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1488 ±5 | 39 | — | 157 (155 - 160) | — | — | 1546 ±15 | 1518 ±8 | 281 | 44 | 14 | 15.0 | Released 2026-08-13. LMArena's row stands on 21,539 votes in its 2026-10-02 dataset publish. Artificial Analysis marked its Gemini 3.7 Flash (high) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | |||||||||||||
| 7 | GPT-6.1 Sol | OpenAI | International | 2026-09-29 | yes | 66.4 | 13 sources, 7 families, 5 operators; 42.8% budget coverage. Sensitivity rank range 3–13 across 56 recomputation scenarios (not a confidence interval).Source observations
| — | — | — | — | 63.94 ±1.07 | — | — | — | — | — | — | — | Released 2026-09-29 per OpenAI's announcement (https://openai.com/index/introducing-gpt-6-1-sol/). This is a distinct model from GPT-6 Sol, whose row and scores are retained separately. LiveBench's Max Effort row is used consistently for Overall, Coding and Agentic Coding. On sale here since 2026-10-07. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 8 | Gemini 3.8 Flash | International | 2026-09-02 | yes | 65.9 | 16 sources, 9 families, 7 operators; 52.4% budget coverage. Sensitivity rank range 7–11 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1495 ±5 | 41 | — | 157 (155 - 160) | 55.95 ±1.14 | — | 1536 ±14 | 1530 ±8 | 297 | 46 | 20 | 21.7 | Released 2026-09-02 per Google's announcement (Introducing Gemini 3.8 Flash and 3.8 Flash Cyber). LMArena's row stands on 26,298 votes in its 2026-10-02 dataset publish. Epoch, which carried no row for it as of 2026-09-05, lists it as Gemini 3.8 Flash (dated 2026-09-02) in the 2026-09-22 read. Confirmed on sale in the public catalogue read on 2026-09-12. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | |||||||||||||
| 9 | Claude Opus 5 | Anthropic | International | 2026-07-24 | yes | 65.9 | 14 sources, 8 families, 5 operators; 45.8% budget coverage. Sensitivity rank range 4–9 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1490 ±4 | 51 | — | 163 (160 - 167) | — | — | 1548 ±10 | 1533 ±6 | 54 | 60 | 49 | 55.6 | LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 10 | GPT-5.6 Sol | OpenAI | International | 2026-07-09 | yes | 63.1 | 15 sources, 9 families, 6 operators; 49.4% budget coverage. Sensitivity rank range 9–14 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1484 ±4 | 47 | — | 162 (159 - 165) | — | — | 1538 ±12 | 1534 ±7 | 73 | 54 | 40 | 26.1 | OpenAI announced a limited partner preview of the GPT-5.6 series (Sol, Terra and Luna) on 2026-06-26 and made the models generally available on 2026-07-09, the post the preview page links as the launch; release_date follows the launch. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 11 | Claude Sonnet 5.5 | Anthropic | International | 2026-09-28 | yes | 62.0 | 13 sources, 7 families, 5 operators; 42.8% budget coverage. Sensitivity rank range 6–12 across 56 recomputation scenarios (not a confidence interval).Source observations
| — | — | — | 165 (162 - 169) | 70.85 ±0.93 | — | — | — | — | — | — | — | Released 2026-09-28 per Anthropic's announcement (https://www.anthropic.com/claude-sonnet-5-5). LiveBench's xHigh Effort row leads its Max Effort row on Overall (77.8 versus 75.7); all three LiveBench columns here quote the same xHigh row. Vals uses max effort and includes provider fallbacks in its default score; that configuration is separate from LiveBench's. On sale here since 2026-10-08. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 12 | Qwen3.8-Max | Alibaba | China | 2026-08-02 | no | 59.9 | 13 sources, 8 families, 5 operators; 43.4% budget coverage. Sensitivity rank range 11–17 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1482 ±5 | 40 | 71.48 | 157 (155 - 159) | — | — | 1538 ±15 | 1524 ±8 | 38 | 55 | 19 | 19.4 | The 2026-08-02 build of Qwen3.8-Max (Artificial Analysis dates it 2026-08-03), which replaced Qwen3.8-Max-Preview (retired 2026-08-05). Its release_date follows Epoch, which re-dated this build from 2026-07-19, the Preview's announcement at WAIC, to 2026-08-02 between its 2026-08-25 and 2026-09-03 releases; Model Studio listed qwen3.8-max on 2026-08-02 as well. Alibaba Cloud's update notices of 2026-09-02 (www.aliyun.com/notice/118616 and the international Model Studio notice Update Notice for Qwen3.8-Max Models) state that from 10:00 Beijing time on 2026-09-05, subject to the actual change time, the qwen3.8-max model name automatically moved to the qwen3.8-max-0902 snapshot with billing unchanged. The name on sale here therefore now serves the Qwen3.8-Max-0902 row, and Model Studio lists no snapshot ID for this build, so it can no longer be called there. Every figure on this row belongs to this build: Artificial Analysis's Qwen3.8 Max (slug qwen3-8-max-0803), Epoch's Qwen 3.8 Max (dated 2026-08-02), SuperCLUE's Qwen3.8-Max(max) (evaluation published 2026-08-06), LMArena's qwen3.8-max (added to its Text Arena on 2026-08-02 per its changelog), LiveBench's Qwen 3.8 Max (quoted at the same figures since 2026-08-10, before the snapshot existed) and Vals's Qwen 3.8 Max (released 2026-08-03 per its model page). Artificial Analysis marked its Qwen3.8 Max row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 13 | Muse Spark 1.3 | Meta | International | 2026-09-02 | no | 57.1 | 11 sources, 5 families, 4 operators; 36.2% budget coverage. Sensitivity rank range 12–21 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1494 ±6 | 48 | — | 157 (155 - 160) | — | — | 1531 ±20 | 1539 ±11 | 219 | 59 | 33 | 43.3 | Released 2026-09-02 per Meta AI Research's announcement (Introducing Muse Spark 1.3). Epoch, which carried no row for it as of 2026-09-05, lists it as Muse Spark 1.3 (dated 2026-09-02) in the 2026-09-22 read. LMArena added muse-spark-1.3-max to its text board on 2026-09-13 per its leaderboard changelog; the row stands on 12,343 votes in its 2026-10-02 dataset publish. Previously served here as muse-spark-1.3-contributor, Meta's Contributor tier, whose terms let Meta train future models on the prompts and completions sent to it. That tier's reasoning_effort stops at xhigh: the LMArena and Artificial Analysis rows quoted here are the max-effort setting, which it does not offer; the LiveBench rows are xHigh, and Vals and Epoch name no effort setting. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. ChinaAPI suspended this Contributor serving variant on 2026-10-24 after its sole supplier group was disabled; this historical evaluation row remains for comparison. | ||||||||||||
| 14 | GPT-5.5 | OpenAI | International | 2026-04-23 | no | 56.4 | 14 sources, 8 families, 5 operators; 45.8% budget coverage. Sensitivity rank range 11–19 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1481 ±4 | — | — | 159 (157 - 162) | — | — | 1517 ±10 | 1519 ±6 | — | — | — | — | The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 15 | GPT-6 Sol | OpenAI | International | 2026-09-22 | yes | 56.3 | 14 sources, 8 families, 5 operators; 45.8% budget coverage. Sensitivity rank range 8–23 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1457 ±7 | 48 | — | — | — | — | 1477 ±23 | 1520 ±13 | 104 | 49 | 44 | — | Released 2026-09-22 per OpenAI's announcement (Introducing GPT-6 Sol and Luna) and the API changelog entry of the same day. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 16 | DeepSeek V4.1 Flash 1017 | DeepSeek | China | 2026-10-17 | yes | 55.7 | 10 sources, 7 families, 5 operators; 26.8% budget coverage. Sensitivity rank range 10–16 across 56 recomputation scenarios (not a confidence interval).Source observations
| — | — | — | — | 61.26 ±1.05 | — | — | — | — | — | — | — | DeepSeek's dated 2026-10-17 checkpoint, published as DeepSeek-V4.1-Flash-1017 (https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017). The official API documentation states that deepseek-flash now serves this build (https://api-docs.deepseek.com/). The serving ID is present in the public ChinaAPI pricing catalogue read on 2026-10-23. September V4.1 Flash results remain on their historical row; none are inherited by this checkpoint. Official weights use the MIT license. Open weights · License: MIT · Verified: 2026-10-23 | ||||||||||||
| 17 | Kimi K3 | Moonshot AI | China | 2026-07-16 | yes | 53.4 | 15 sources, 9 families, 7 operators; 46.4% budget coverage. Sensitivity rank range 15–20 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1488 ±5 | 44 | 70.68 | 158 (155 - 160) | 52.95 ±1.00 | — | 1540 ±13 | 1541 ±8 | 37 | 51 | 13 | 38.3 | LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 18 | Qwen3.8-Max-0902 | Alibaba | China | 2026-09-02 | yes | 53.1 | 7 sources, 6 families, 3 operators; 18.6% budget coverage. Sensitivity rank range 16–20 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 5 of them.Source observations
| — | 45 | 72.62 | 155 (153 - 157) | — | — | — | — | 39 | 58 | 39 | 20.0 | Alibaba's 2026-09-02 snapshot of Qwen3.8-Max (alias qwen3.8-max-2026-09-02), listed on Model Studio on 2026-09-02 as an upgraded snapshot of qwen3.8-max; Epoch dates it 2026-09-01. Alibaba Cloud's update notices of 2026-09-02 (www.aliyun.com/notice/118616 and the international Model Studio notice Update Notice for Qwen3.8-Max Models) state that from 10:00 Beijing time on 2026-09-05, subject to the actual change time, the qwen3.8-max model name automatically moved to this snapshot with billing unchanged. Alibaba published no completion notice; Artificial Analysis has since marked the earlier build deprecated. This row therefore links to qwen3.8-max, the name on sale here; the same snapshot is also sold under its pinned name qwen3.8-max-0902. Artificial Analysis and Epoch list it as Qwen3.8 Max (0902), SuperCLUE as Qwen3.8-Max-0902(max) (evaluation published 2026-09-10). LMArena added qwen3.8-max-0902 only to its Code Arena (changelog, 2026-09-01), which this page does not cite, and neither LiveBench nor Vals carries a 0902 row; the 2026-08-02 build's figures on the qwen3.8-max row are not reused here. | ||||||||||||
| 19 | GLM-5.3 | Zhipu AI (Z.ai) | China | 2026-08-14 | yes | 52.6 | 15 sources, 9 families, 7 operators; 46.4% budget coverage. Sensitivity rank range 16–19 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1478 ±6 | 45 | 71.29 | 156 (154 - 158) | 59.73 ±1.11 | — | 1531 ±17 | 1521 ±9 | 53 | 57 | 42 | 48.3 | Released 2026-08-14; served here via Zhipu's official channels since 2026-08-18. Independent boards picked it up between the 2026-08-19 and 2026-08-26 snapshots, so the vendor's own DeepSWE figure it carried before has been replaced by their readings; the LMArena row stands on 17,857 votes in its 2026-10-02 dataset publish. Epoch, which carried no row for it as of 2026-09-05, lists it as GLM-5.3 (dated 2026-08-14) in the 2026-09-22 read, and SuperCLUE lists GLM-5.3(max) with an evaluation published 2026-09-10. Vals lists it as GLM 5.3, released 2026-08-18 per its model page, the date of Z.ai's release notes. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 20 | Muse Spark 1.2 | Meta | International | 2026-08-05 | no | 49.5 | 10 sources, 6 families, 5 operators; 37.8% budget coverage. Sensitivity rank range 18–24 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1494 ±9 | 40 | — | 155 (153 - 157) | — | — | 1528 ±31 | 1531 ±17 | 172 | 49 | 7 | 31.1 | Epoch added a Muse Spark 1.2 row (dated 2026-08-05) between the 2026-08-31 and 2026-09-05 snapshots. Artificial Analysis marked its Muse Spark 1.2 (xhigh) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. Previously served here as muse-spark-1.2-contributor, Meta's Contributor tier, whose terms let Meta train future models on the prompts and completions sent to it; the xhigh setting the LMArena, LiveBench and Artificial Analysis rows quote is available on that tier. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. ChinaAPI suspended this Contributor serving variant on 2026-10-24 after its sole supplier group was disabled; this historical evaluation row remains for comparison. | ||||||||||||
| 21 | Gemini 3.6 Flash | International | 2026-07-21 | yes | 49.3 | 13 sources, 8 families, 5 operators; 43.4% budget coverage. Sensitivity rank range 19–23 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1483 ±4 | 34 | — | 154 (153 - 156) | — | — | 1530 ±12 | 1519 ±7 | 180 | 38 | 7 | 14.4 | Artificial Analysis marked its Gemini 3.6 Flash row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | |||||||||||||
| 22 | GPT-5.6 Terra | OpenAI | International | 2026-07-09 | yes | 48.2 | 15 sources, 9 families, 6 operators; 49.4% budget coverage. Sensitivity rank range 12–27 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1465 ±4 | 42 | — | 160 (157 - 163) | — | — | 1516 ±12 | 1515 ±7 | 82 | 47 | 35 | 31.7 | OpenAI announced a limited partner preview of the GPT-5.6 series (Sol, Terra and Luna) on 2026-06-26 and made the models generally available on 2026-07-09, the post the preview page links as the launch; release_date follows the launch. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 23 | Gemini 3.1 Pro | International | 2026-02-19 | no | 47.9 | 15 sources, 9 families, 6 operators; 49.4% budget coverage. Sensitivity rank range 20–32 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1487 ±3 | 30 | — | 155 (153 - 157) | — | — | 1529 ±7 | 1522 ±5 | 115 | 14 | 4 | 15.6 | LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | |||||||||||||
| 24 | Grok 4.5 | xAI | International | 2026-07-08 | no | 47.5 | 13 sources, 9 families, 6 operators; 45.8% budget coverage. Sensitivity rank range 20–27 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1466 ±4 | 39 | — | 154 (152 - 156) | — | — | 1504 ±11 | 1514 ±7 | 55 | 44 | 11 | 15.0 | LMArena labels this vendor SpaceXAI. Artificial Analysis marked its Grok 4.5 (high) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. Released 2026-07-08 per xAI's announcement as published that day and its API release notes; the live announcement page now shows Jul 16, the date it was edited when the model reached the EU. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 25 | Grok 4.6 | xAI | International | 2026-08-12 | no | 47.3 | 12 sources, 7 families, 4 operators; 39.8% budget coverage. Sensitivity rank range 15–28 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1454 ±5 | 44 | — | 157 (155 - 159) | — | — | 1503 ±14 | 1507 ±8 | 58 | 55 | 21 | 12.2 | Released 2026-08-12. LMArena labels this vendor SpaceXAI. Its row stands on 23,680 votes in LMArena's 2026-10-02 dataset publish. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 26 | DeepSeek V4.1 Flash | DeepSeek | China | 2026-09-10 | no | 46.8 | 15 sources, 8 families, 6 operators; 48.8% budget coverage. Sensitivity rank range 23–29 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1474 ±7 | 39 | 71.81 | 155 (149 - 158) | — | — | 1523 ±23 | 1528 ±12 | 220 | 55 | 27 | 22.8 | Released 2026-09-10 per DeepSeek's official announcement and served by the canonical deepseek-flash API model. The vendor retired V4 Flash and V4 Flash Vision Exp that day and temporarily routes their old identifiers to V4.1 Flash; their historical leaderboard rows remain separate and none of their figures are reused here. Artificial Analysis, LiveBench and Vals all published distinct V4.1 Flash rows by 2026-09-12; Epoch (DeepSeek V4.1 Flash, dated 2026-09-09) and SuperCLUE (DeepSeek-V4.1-Flash(max), evaluation published 2026-09-11) carry one as of the 2026-09-22 read. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. Historical 2026-09-10 checkpoint: the official deepseek-flash API now serves DeepSeek-V4.1-Flash-1017. This row keeps its original anchor and its own observations; the currently served build is listed separately. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 27 | MiMo-V2.6-Pro | Xiaomi | China | 2026-09-22 | yes | 46.5 | 7 sources, 3 families, 4 operators; 29.0% budget coverage. Sensitivity rank range 21–28 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 3 of them.Source observations
| 1480 ±9 | 46 | — | — | 57.81 ±1.24 | — | 1511 ±33 | 1540 ±18 | 55 | 59 | 35 | — | Released and open-sourced 2026-09-22 per Xiaomi's release record (MiMo-V2.6 Series Release). | ||||||||||||
| 28 | GLM-5.3-Flash | Zhipu AI (Z.ai) | China | 2026-08-26 | yes | 44.2 | 13 sources, 7 families, 6 operators; 42.4% budget coverage. Sensitivity rank range 24–31 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1473 ±5 | 42 | 68.10 | 152 (149 - 154) | 50.86 ±1.25 | — | 1529 ±15 | 1523 ±8 | 65 | 57 | 33 | 51.1 | Released 2026-08-26 per Z.ai's own release notes; the first natively multimodal model in the GLM-5 line and its low-cost tier, served here on Zhipu's official channels. LMArena, LiveBench and Artificial Analysis all picked it up within days of release; the LMArena row is a ranked one, standing on 22,971 votes in its 2026-10-02 dataset publish. Epoch, SuperCLUE and Vals carried no row for it as of 2026-08-31. In the 2026-09-22 read Epoch lists it as GLM-5.3-Flash dated 2026-08-20, six days before Z.ai's release notes; the benchmark runs behind that figure are on the glm-5.3-flash model itself (Epoch's model versions glm-5.3-flash_max and glm-5.3-flash_high), so only the date differs. SuperCLUE lists GLM-5.3-Flash(max) with an evaluation published 2026-09-10, and Vals lists it as GLM 5.3 Flash, released 2026-08-26 per its model page. GLM-5.3's figures are not reused here. The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 29 | DeepSeek V4 Pro 0813 | DeepSeek | China | 2026-08-13 | yes | 43.2 | 12 sources, 7 families, 5 operators; 36.8% budget coverage. Sensitivity rank range 22–31 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1464 ±7 | 36 | 70.10 | 155 (154 - 158) | 52.97 ±1.04 | — | 1493 ±22 | 1505 ±11 | 68 | 47 | 14 | 17.8 | DeepSeek's change log dates this build's GA release to 2026-08-13 and states that the deepseek-v4-pro model name now serves it. Epoch added a DeepSeek V4 Pro 0813 row between the 2026-08-26 and 2026-08-31 snapshots, so the undated DeepSeek-V4-Pro figure is no longer borrowed for it. LMArena's 0813 row stands on 9,780 votes in its 2026-10-02 dataset publish. SuperCLUE announced on 2026-08-26 that it had added the official DeepSeek-V4-Pro; in the July 2026 edition's data file read on 2026-09-22 that row is labelled DeepSeek-V4-Pro-0813(max), with an evaluation published 2026-08-18, and the 64.40 row the board had labelled DeepSeek-V4-Pro(max) 预览版, now plain DeepSeek-V4-Pro(max), is the preview build, quoted on the deepseek-v4-pro row rather than here. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 30 | Grok 4.7 | xAI | International | 2026-09-21 | yes | 43.1 | 15 sources, 8 families, 6 operators; 48.8% budget coverage. Sensitivity rank range 19–33 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1442 ±8 | 46 | — | — | 58.56 ±1.01 | — | 1507 ±25 | 1488 ±16 | 39 | 60 | 26 | 15.0 | Released 2026-09-21 per the vendor's announcement, which now publishes under the SpaceXAI name; the row keeps the xAI label of the earlier Grok rows. Artificial Analysis lists Grok 4.7 (xhigh) and Grok 4.7 (high), both printing 46; the xhigh row is quoted because it is the highest effort setting and the higher score before rounding (46.4 against 46.3). LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 31 | GPT-5.6 Luna | OpenAI | International | 2026-07-09 | yes | 40.6 | 15 sources, 9 families, 6 operators; 49.4% budget coverage. Sensitivity rank range 26–34 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1453 ±4 | 37 | — | 156 (154 - 159) | — | — | 1476 ±12 | 1501 ±7 | 144 | 47 | 12 | 19.4 | The low-cost tier of the GPT-5.6 line. OpenAI announced a limited partner preview of the GPT-5.6 series (Sol, Terra and Luna) on 2026-06-26 and made the models generally available on 2026-07-09, the post the preview page links as the launch; release_date follows the launch. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 32 | DeepSeek V4 Flash 0731 | DeepSeek | China | 2026-07-31 | no | 38.7 | 8 sources, 6 families, 3 operators; 21.4% budget coverage. Sensitivity rank range 29–36 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 5 of them.Source observations
| — | 34 | 65.60 | 155 (152 - 157) | — | — | — | — | 221 | 46 | 12 | 13.3 | Historical 2026-07-31 V4 Flash build. DeepSeek retired V4 Flash on 2026-09-10 and its compatibility ID now routes to V4.1 Flash; this historical score row is not the currently served build and is not marked on sale. Artificial Analysis marked its DeepSeek V4 Flash 0731 (max) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 33 | GLM-5.2 | Zhipu AI (Z.ai) | China | 2026-06-16 | yes | 38.1 | 13 sources, 8 families, 5 operators; 39.8% budget coverage. Sensitivity rank range 25–34 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1476 ±4 | 34 | 63.27 | 152 (150 - 154) | — | — | 1520 ±11 | 1512 ±6 | 68 | 43 | 1 | 7.2 | Artificial Analysis marked its GLM-5.2 (max) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 34 | Kimi K2.6 | Moonshot AI | China | 2026-04-20 | yes | 35.1 | 12 sources, 6 families, 3 operators; 38.2% budget coverage. Sensitivity rank range 30–35 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 3 of them.Source observations
| 1461 ±4 | — | — | 151 (149 - 153) | — | — | 1516 ±13 | 1516 ±7 | — | — | — | — | The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 35 | GLM-5.1 | Zhipu AI (Z.ai) | China | 2026-04-07 | yes | 34.6 | 10 sources, 7 families, 4 operators; 33.8% budget coverage. Sensitivity rank range 31–38 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1465 ±4 | — | — | 150 (148 - 152) | — | — | 1514 ±10 | 1513 ±6 | — | — | — | — | The Vals Index V2 (2026-08-13) no longer lists this model; the V1.2 figure it once carried was removed with that revision. | ||||||||||||
| 36 | GLM-5 | Zhipu AI (Z.ai) | China | 2026-02-12 | yes | 33.7 | 6 sources, 3 families, 3 operators; 20.4% budget coverage. Sensitivity rank range 32–41 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 5 of them.Source observations
| 1458 ±4 | — | — | 146 (144 - 148) | — | — | 1516 ±14 | 1498 ±7 | — | — | — | — | Released 2026-02-12 per Z.ai's blog (GLM-5: From Vibe Coding to Agentic Engineering) and both of its release notes; Epoch records 2026-02-11, the US-time day of the launch post. | ||||||||||||
| 37 | DeepSeek V4 Pro | DeepSeek | China | 2026-04-24 | no | 33.0 | 10 sources, 6 families, 3 operators; 29.8% budget coverage. Sensitivity rank range 32–38 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 4 of them.Source observations
| 1458 ±4 | 30 | 64.40 | 149 (147 - 151) | — | — | 1489 ±11 | 1500 ±6 | 86 | 32 | 15 | 21.1 | The pre-0813 build; the deepseek-v4-pro API moved to DeepSeek-V4-Pro-0813 on 2026-08-12. LiveBench dropped this row in favour of the dated builds between the 2026-08-19 and 2026-08-26 snapshots, so its 71.6 is no longer quoted. SuperCLUE's DeepSeek-V4-Pro(max) row (64.40), which its July 2026 edition labelled 预览版, is quoted here from CA-971 on, overturning CA-605. CA-605 left it out because LMArena lists deepseek-v4-pro and deepseek-v4-pro-high-preview separately, but LMArena's own data gives those two rows the same checkpoint key (deepseek-v4-pro-ch1-text and deepseek-v4-pro-ch1-thinking-text) and links both to DeepSeek's 2026-04-24 announcement: they are one build in two modes, which DeepSeek's change log calls V4-Pro-Preview. Before the 2026-08-13 GA that build was the only one the deepseek-v4-pro API served, so an API evaluation published in the July edition can only be this row. SuperCLUE's data file has since dropped the 预览版 label and dates the row 2026.9.10, keeping the July figure. Artificial Analysis marked its DeepSeek V4 Pro (max) row deprecated between the 2026-09-12 and 2026-09-21 reads; the leaderboard hides such rows under its default Status: Current filter, so this row's leaderboard figures are read with the Status filter set to All. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 38 | GPT-6 Luna | OpenAI | International | 2026-09-22 | yes | 32.4 | 13 sources, 7 families, 5 operators; 42.8% budget coverage. Sensitivity rank range 25–39 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1443 ±7 | 37 | — | — | 52.32 ±1.12 | — | 1484 ±21 | 1512 ±12 | 157 | 43 | 13 | — | Released 2026-09-22 per OpenAI's announcement (Introducing GPT-6 Sol and Luna) and the API changelog entry of the same day. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 39 | Qwen3.8-27B | Alibaba | China | 2026-08-14 | yes | 32.1 | 7 sources, 3 families, 3 operators; 28.4% budget coverage. Sensitivity rank range 34–40 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 4 of them.Source observations
| 1438 ±5 | 34 | — | 149 (147 - 152) | — | — | 1500 ±17 | 1502 ±9 | 41 | 45 | 6 | 21.7 | Open-weight Qwen3.8 model, released 2026-08-14 per the Qwen team's release notes (Hugging Face and ModelScope); on sale here since 2026-09-04. Epoch, which carried no row for it as of 2026-09-05, lists it as Qwen 3.8 27B (dated 2026-08-14) in the 2026-09-22 read. Artificial Analysis lists several effort settings; the first-listed xhigh row is quoted rather than a lower-effort sibling. The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| 40 | MiniMax-M3 | MiniMax | China | 2026-06-01 | yes | 31.3 | 10 sources, 5 families, 5 operators; 34.0% budget coverage. Sensitivity rank range 34–40 across 56 recomputation scenarios (not a confidence interval).Source observations
| 1440 ±4 | 29 | 56.90 | 147 (143 - 150) | — | — | 1473 ±11 | 1494 ±6 | 106 | 37 | 2 | 17.2 | LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. | ||||||||||||
| 41 | Kimi K2.7 Code | Moonshot AI | China | 2026-06-12 | yes | 29.9 | 8 sources, 7 families, 4 operators; 22.6% budget coverage. Sensitivity rank range 35–41 across 56 recomputation scenarios (not a confidence interval). Not eligible for a rank in 1 of them.Source observations
| — | 26 | — | 150 (148 - 152) | — | — | — | — | 61 | 26 | 1 | 16.1 | LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| — | MiMo-V2.6-Flash | Xiaomi | China | 2026-09-22 | yes | — | Unranked: insufficient families, operators. 5 sources, 2 families, 2 operators; 21.4% budget coverage.Source observations
| 1452 ±8 | — | — | — | 54.83 ±1.11 | — | 1502 ±27 | 1521 ±15 | — | — | — | — | Released and open-sourced 2026-09-22 per Xiaomi's release record (MiMo-V2.6 Series Release). The 2026-10-01 Vals Index v2.1 read independently covers this model, replacing the previously quoted vendor self-reported DeepSWE figure. | ||||||||||||
| — | MiMo-V2.6-Pro-Ultraspeed | Xiaomi | China | 2026-09-22 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | — | — | — | — | — | — | — | Xiaomi's high-speed serving mode of MiMo-V2.6-Pro, released 2026-09-22. The vendor states flagship V2.6-Pro performance at up to 20x the output speed but publishes no separate figures, and no board lists it separately, so MiMo-V2.6-Pro's scores are not reused here. | ||||||||||||
| — | Step 5 Preview | StepFun | China | 2026-09-20 | yes | — | Unranked: insufficient families, operators. 5 sources, 2 families, 2 operators; 21.4% budget coverage.Source observations
| — | 44 | — | — | — | — | — | — | 83 | 53 | 33 | 16.7 | StepFun's announcement page prints no date; it says the model is released today and cites Artificial Analysis results published as of 2026-09-20, and reports dated that day already quote the release, which fixes the release date. Official BF16, FP8 and NVFP4 weights are now available in the StepFun collection, verified 2026-10-23 (https://huggingface.co/collections/stepfun-ai/step-5-preview). The StepFun Model License 1.0 has additional attribution and commercial-license conditions; it is not an unmodified Apache-2.0 license. Weight publication does not create a new model release or additional consensus ranks. Vals lists it as Step 5 Preview, released 2026-09-28 per its model page; the release date above follows StepFun's announcement. The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. Open weights · License: StepFun Model License 1.0 · Verified: 2026-10-23 | ||||||||||||
| — | Hy4-preview | Tencent | China | 2026-08-28 | yes | — | Unranked: insufficient categories. 3 sources, 3 families, 3 operators; 9.0% budget coverage.Source observations
| — | — | — | — | 53.21 ±1.13 | — | — | — | — | — | — | — | Tencent announced and open-sourced Hy4 preview on 2026-08-28. The Vals Index lists it as Hy4 Preview (tencent/hy4-preview); Artificial Analysis, LMArena, LiveBench and Epoch carried no row for it as of 2026-09-21, and those figures are left blank. Hy3 scores and AutoEval results are not attributed to this row. | ||||||||||||
| — | Qwen3.8-Flash | Alibaba | China | 2026-08-26 | yes | — | Unranked: insufficient sources, families, operators, categories. 2 sources, 2 families, 2 operators; 6.0% budget coverage.Source observations
| — | 40 | 66.41 | — | — | — | — | — | 54 | 56 | 25 | — | The boards list this model under its open-weight release name, Qwen3.8-Flash-Next (released 2026-08-26 per the Qwen team). The vendor's announcement states that the production version, with a 1M default context and built-in tools, is served under the name Qwen3.8-Flash, which is the qwen3.8-flash API on sale here since 2026-09-04; the two rows are joined on that statement, not on the row name. Artificial Analysis now prints 40 without an estimated marker (read 2026-09-21). SuperCLUE lists it under the API name itself, Qwen3.8-Flash(max), evaluated through the API with an evaluation published 2026-09-10, so that figure needs no join. Neither Epoch, LMArena nor Vals carried a row as of 2026-09-05. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| — | DeepSeek V4 Flash Vision Exp | DeepSeek | China | 2026-08-21 | no | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | 35 | — | — | — | — | — | — | 231 | 52 | 12 | — | DeepSeek's experimental vision tier, announced in the vendor's change log on 2026-08-21 as model='deepseek-v4-flash-vision-exp'. LiveBench names its row with the vendor's -Exp suffix and Artificial Analysis drops it, but DeepSeek publishes only one vision tier, so both rows are this model. LMArena, Vals, SuperCLUE and Epoch carry no row for it, and the sibling flash figures are not reused. DeepSeek retired this tier on 2026-09-10 and its compatibility ID now routes to V4.1 Flash. It is absent from the public catalogue read on 2026-09-12 and is retained here as a historical evaluation row without an on-sale badge. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| — | Hunyuan Hy3 | Tencent | China | 2026-07-06 | yes | — | Unranked: insufficient families, operators, categories. 3 sources, 1 families, 1 operators; 12.4% budget coverage.Source observations
| 1456 ±7 | 25 | 62.13 | — | — | — | 1509 ±21 | 1500 ±11 | 87 | 27 | 1 | — | |||||||||||||
| — | LongCat-2.0 | Meituan | China | 2026-06-30 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | 19 | 54.22 | — | — | — | — | — | — | 18 | 0 | — | Artificial Analysis no longer printed an output-speed value for this row on 2026-09-12, so the old figure was removed rather than retained without a live source. | ||||||||||||
| — | Doubao-Seed-2.1-pro | ByteDance | China | 2026-06-28 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | 65.94 | — | — | Terminal Bench 2.1 71.0 | — | — | — | — | — | — | |||||||||||||
| — | Doubao-Seed-2.1-turbo | ByteDance | China | 2026-06-28 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | Terminal Bench 2.1 67.6 | — | — | — | — | — | — | No independent board in this snapshot covers it; the figure shown is the vendor's own. The vendor writes the name without the 260628 snapshot suffix we serve it under. | ||||||||||||
| — | Qwen3.7-Plus | Alibaba | China | 2026-06-01 | yes | — | Unranked: insufficient operators. 6 sources, 4 families, 2 operators; 22.8% budget coverage.Source observations
| 1456 ±4 | 25 | — | 147 (146 - 149) | — | — | 1504 ±12 | 1501 ±6 | 65 | 13 | 1 | 9.4 | Released 2026-06-01 (10:00 China time) per the Qwen blog, the day Model Studio also lists it; Epoch records 2026-06-02, the China-time day of the launch post on X. The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. | ||||||||||||
| — | Step 3.7 Flash | StepFun | China | 2026-05-29 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | 19* | 48.16 | — | — | — | — | — | 193 | 17 | — | — | |||||||||||||
| — | Qwen3.7-Max | Alibaba | China | 2026-05-20 | yes | — | Unranked: insufficient operators, categories. 6 sources, 4 families, 1 operators; 13.8% budget coverage.Source observations
| 1475 ±10 | — | — | 154 (152 - 156) | — | — | 1527 ±35 | 1524 ±18 | — | — | — | — | LMArena lists the preview build (qwen3.7-max-preview); the released model has no separate row. Released 2026-05-20 (10:00 China time) per the Qwen blog; Epoch records 2026-05-19, and the blog itself renders the date in the reader's time zone, so US readers see the 19th. The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| — | Qwen3.6-Flash | Alibaba | China | 2026-04-26 | yes | — | Unranked: insufficient operators, categories. 5 sources, 4 families, 1 operators; 12.6% budget coverage.Source observations
| — | — | — | 143 (141 - 145) | — | — | — | — | — | — | — | — | |||||||||||||
| — | MiMo-V2.5 | Xiaomi | China | 2026-04-22 | no | — | Unranked: insufficient families, operators, categories. 4 sources, 1 families, 1 operators; 18.4% budget coverage.Source observations
| 1434 ±4 | 25* | — | — | — | — | 1479 ±12 | 1491 ±7 | 34 | 24 | 0 | — | The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. Removed from ChinaAPI on 2026-10-24 after Xiaomi retired this model on 2026-10-21; historical benchmark evidence is retained. | ||||||||||||
| — | MiMo-V2.5-Pro | Xiaomi | China | 2026-04-22 | no | — | Unranked: insufficient families, operators, categories. 5 sources, 2 families, 2 operators; 18.4% budget coverage.Source observations
| 1468 ±4 | 26 | 53.01 | — | — | — | 1515 ±10 | 1522 ±6 | 45 | 30 | 0 | 9.4 | The exact build is absent from the Vals Index v2.2 cohort read on 2026-10-23; its former v2.1 figure is no longer quoted. Removed from ChinaAPI on 2026-10-24 after Xiaomi retired this model on 2026-10-21; historical benchmark evidence is retained. | ||||||||||||
| — | Step 3.5 Flash 2603 | StepFun | China | 2026-04-02 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | 17* | — | — | — | — | — | — | 118 | — | — | — | StepFun's agent-optimised variant of step-3.5-flash, whose reasoning_effort takes only low and high. Released 2026-04-02 per StepFun's announcement on its X account ("Step 3.5 Flash 2603 just shipped!"). Artificial Analysis lists it as Step 3.5 Flash 2603 (released 2026-04-02 on its model page) and marks the row deprecated; the leaderboard hides such rows under its default Status: Current filter, so its figures are read with the Status filter set to All, and its Intelligence Index is an estimate (*). LMArena carries no 2603 row: its step-3.5-flash row is the base build and is not reused here. LiveBench, Epoch, Vals and SuperCLUE carry no row for it either. | ||||||||||||
| — | Qwen3.6-Plus | Alibaba | China | 2026-04-01 | yes | — | Unranked: insufficient operators. 8 sources, 5 families, 2 operators; 25.0% budget coverage.Source observations
| 1443 ±4 | — | — | 148 (145 - 149) | — | — | 1482 ±12 | 1496 ±6 | — | — | — | — | The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. LiveBench 1.0 scores were withdrawn on 2026-10-23 after the site moved them into its historical-release view and launched the incompatible LiveBench 2.0 categories. | ||||||||||||
| — | MiniMax-M2.7 | MiniMax | China | 2026-03-18 | yes | — | Unranked: insufficient families, operators, categories. 3 sources, 1 families, 1 operators; 12.4% budget coverage.Source observations
| 1415 ±3 | — | — | 146 (138 - 148) | — | — | 1447 ±9 | 1480 ±5 | — | — | — | — | The previously cited Vals Index V2 row is absent from the V2.1 board read on 2026-10-01, so its old V2 figure is no longer quoted. | ||||||||||||
| — | GLM-5-Turbo | Zhipu AI (Z.ai) | China | 2026-03-16 | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | — | — | — | — | — | — | — | No independent board in this snapshot carries a row for it. | ||||||||||||
| — | Step 3.5 Flash | StepFun | China | 2026-02-02 | yes | — | Unranked: insufficient sources, families, operators, categories. 2 sources, 1 families, 1 operators; 10.0% budget coverage.Source observations
| 1393 ±4 | 17* | — | — | — | — | 1432 ±10 | 1449 ±6 | 126 | — | — | — | StepFun's platform documentation lists step-3.5-flash as the base version and step-3.5-flash-2603 as an agent-optimised variant built from it; the 2603 build shipped on 2026-04-02 and has its own row here. LMArena added step-3.5-flash to its text board on 2026-02-10, linking the original open weights, seven weeks before 2603 shipped, so its three text-board rows are this build (60,207 votes on the overall board in its 2026-10-02 dataset publish). Artificial Analysis lists this build as Step 3.5 Flash (released 2026-02-02 on its model page) and marks the row deprecated; the leaderboard hides such rows under its default Status: Current filter, so its figures are read with the Status filter set to All, and its Intelligence Index is an estimate (*). The vendor's own SWE-bench Verified figure (74.4) that stood here while no independent board was quoted has been removed. LiveBench, Epoch, Vals and SuperCLUE carry no row for it. | ||||||||||||
| — | Agnes 2.0 Flash | Agnes AI | China | — | no | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | Claw-Eval Pass^3 60.9% | — | — | — | — | — | — | No independent board in this snapshot covers it; the figure shown is the vendor's own. Retired from sale 2026-09-03 (CA-654): the vendor deprecated it in favour of agnes-2.5-flash; kept as a ranked row with on_chinaapi false so the roster count and maintained.at still describe the 2026-08-31 board read. | ||||||||||||
| — | Agnes 2.5 Flash | Agnes AI | China | — | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | — | — | — | — | — | — | — | Artificial Analysis carries a different build of this family (Agnes 2.5 Pro Alpha, 40) and the vendor publishes its own figures only as a chart image, so there is no number here that can be quoted. | ||||||||||||
| — | Agnes 3.0 Flash | Agnes AI | China | — | yes | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | 36* | — | — | — | — | — | — | — | 53 | 7 | — | Artificial Analysis lists Agnes 3.0 Flash separately from the 2.5 family. Its Intelligence Index is currently estimated, so the trailing * is retained exactly as the board prints it. Artificial Analysis no longer printed an output-speed value for this row on 2026-09-21 (the column reads --), so the 238 tokens/s figure was removed rather than retained without a live source; the index, GDPval-AA and Terminal-Bench figures are from the 2026-09-21 board read. | ||||||||||||
| — | DeepSeek V4 Flash | DeepSeek | China | — | no | — | Unranked: insufficient families, operators, categories. 4 sources, 2 families, 2 operators; 16.0% budget coverage.Source observations
| 1436 ±4 | — | — | 146 (144 - 148) | — | — | 1471 ±11 | 1484 ±6 | — | — | — | — | The pre-0731 build. LMArena still lists this build and carries no 0731 row in its 2026-10-02 dataset publish; LiveBench dropped its row between the 2026-08-19 and 2026-08-26 snapshots, so its 65.5 is no longer quoted. Epoch lists this build as DeepSeek-V4-Flash, dated 2026-04-24, separately from its DeepSeek V4 Flash 0731 row. | ||||||||||||
| — | Doubao-Seed-Evolving | ByteDance | China | — | yes | — | Unranked: insufficient sources, families, operators, categories. 1 sources, 1 families, 1 operators; 3.6% budget coverage.Source observations
| — | — | — | — | — | — | — | — | — | — | — | — | Doubao publishes Seed Evolving as a rolling model under the stable doubao-seed-evolving API ID, so capabilities can change without a dated version suffix. No independent score could be verified for this exact model on the boards cited by this snapshot as of 2026-09-12; Seed 2.1 scores are not reused. | ||||||||||||
| — | GLM-5V-Turbo | Zhipu AI (Z.ai) | China | — | yes | — | Unranked: insufficient families, operators, categories. 4 sources, 1 families, 1 operators; 18.4% budget coverage.Source observations
| 1433 ±6 | — | — | — | — | — | 1474 ±21 | 1488 ±11 | — | — | — | — | Vision-language model. LMArena's text board now carries a glm-5v-turbo row (10,436 votes in its 2026-10-02 dataset publish); Epoch does not score it. | ||||||||||||
| — | Ling-3.1-flash | InclusionAI | China | — | no | — | Unranked: insufficient sources, families, operators, categories, directAnchors. 0 sources, 0 families, 0 operators; 0.0% budget coverage.Source observations | — | — | — | — | — | — | — | — | — | — | — | — | Official inclusionAI weights and MIT license verified 2026-10-23 at https://huggingface.co/inclusionAI/Ling-3.1-flash. A 560B-total/25B-active native hybrid reasoning model with up to 1M context. A substantiated vendor announcement date has not been established here, so the release-date cell is left empty; the weight-repository creation date is not substituted. No exact-model observations were found in the sources available to this consensus run. Artificial Analysis is unavailable to consensus without its API key; a third-party index or the vendor's self-reported scores cannot fill that gap. Not present in the public ChinaAPI pricing catalogue read on 2026-10-23. Open weights · License: MIT · Verified: 2026-10-23 | ||||||||||||
| — | Qwen3.8-2.4T-A95B | Alibaba | China | — | yes | — | Unranked: insufficient sources, families, operators, categories. 1 sources, 1 families, 1 operators; 3.6% budget coverage.Source observations
| — | 40 | — | — | — | — | — | — | 38 | 55 | 11 | 0.0 | The open-weight Qwen3.8-2.4T-A95B release has its own Artificial Analysis row. Qwen identifies Qwen3.8-Max as a production version based on this model with additional features; the two catalogue models remain separate here, and Max scores from other boards are not reused. Artificial Analysis prints MLCR-AA 0.0 for this release, with no breakdown, against 19.4 and 20.0 for the two Qwen3.8-Max builds. | ||||||||||||
Image generation models
Quality coverage: 13/17 models, 4 boards from 2 evaluation operators. Public preference evidence, not comprehensive capability certification or a ChinaAPI endpoint test. Tasks and audio pools are separate; scores across columns are not interchangeable. Overlapping intervals do not establish a significant lead.
Ordered by LMArena Text-to-Image Arena; a model with no figure on that board follows the ones that have one, because absent from a board is not the same as last on it.
Third-party price and generation-time references
Not ChinaAPI prices or endpoint measurements. These references never affect quality coverage or ranking. Compare only matching tasks and configurations. Generation time is a trailing-three-day median of successful requests including download, not retry overhead.
Video generation models
Quality coverage: 15/18 models, 4 boards from 2 evaluation operators. Public preference evidence, not comprehensive capability certification or a ChinaAPI endpoint test. Tasks and audio pools are separate; scores across columns are not interchangeable. Overlapping intervals do not establish a significant lead.
Ordered by LMArena Text-to-Video Arena; a model with no figure on that board follows the ones that have one, because absent from a board is not the same as last on it.
Third-party price and generation-time references
Not ChinaAPI prices or endpoint measurements. These references never affect quality coverage or ranking. Compare only matching tasks and configurations. Generation time is a trailing-three-day median of successful requests including download, not retry overhead.
Real usage on ChinaAPI
Which models the calls ChinaAPI actually served went to, as positions and shares of this platform's own traffic. It is a measurement of one gateway, not of the Chinese AI market, and it is not comparable with the benchmark results above. No call count, token count or revenue figure is published here or anywhere else on this site: every number below is a proportion. Recomputed hourly. A model with too little traffic in a window to rank is left unnamed and counted under “other”, as is anything outside the published catalogue. The same board as JSON is at /api/model-rankings/usage.
Share of calls served, the last 7 days
| # | Model | Vendor | Share of calls | Trend vs the previous 7 days |
|---|---|---|---|---|
| 1 | hy3 | Tencent | 21.4% | up |
| 2 | agnes-3.0-flash | Agnes AI | 18.0% | up |
| 3 | agnes-image-2.5-flash | Agnes AI | 15.8% | up |
| 4 | deepseek-flash | DeepSeek | 7.2% | up |
| 5 | gpt-5.6-luna | OpenAI | 2.8% | up |
| 6 | gemini-3.1-flash-lite | 2.6% | up | |
| 7 | claude-sonnet-5-5 | Anthropic | 2.5% | up |
| 8 | claude-opus-5-5 | Anthropic | 2.2% | down |
| 9 | gemini-3.8-flash | 1.7% | up | |
| 10 | gpt-6-luna | OpenAI | 1.4% | down |
| 11 | stepaudio-3-chat-preview | StepFun | 1.3% | flat |
| 12 | agnes-video-2.5-flash | Agnes AI | 1.1% | up |
| 13 | gemini-4-argon | 0.8% | up | |
| 14 | claude-haiku-5-5 | Anthropic | 0.8% | down |
| 15 | claude-sonnet-5 | Anthropic | 0.8% | flat |
| 16 | gpt-6.1-sol | OpenAI | 0.7% | down |
| 17 | agnes-2.5-flash | Agnes AI | 0.5% | flat |
| 18 | glm-5.3-flash | Zhipu AI | 0.5% | down |
| 19 | qwen3.8-max | Alibaba | 0.5% | flat |
| 20 | claude-opus-4-8 | Anthropic | 0.4% | flat |
| All other models | 16.9% |
How those calls arrived: 6.5% over agent protocols (Claude Messages and the OpenAI Responses API), 71.8% over chat protocols (OpenAI chat completions and Gemini generateContent), 21.7% over everything else (audio, images, embeddings, rerank and asynchronous tasks).
Share of calls served, the last 30 days
| # | Model | Vendor | Share of calls | Trend vs the previous 30 days |
|---|---|---|---|---|
| 1 | hy3 | Tencent | 21.2% | up |
| 2 | agnes-3.0-flash | Agnes AI | 6.3% | up |
| 3 | gpt-6-luna | OpenAI | 4.4% | up |
| 4 | deepseek-flash | DeepSeek | 4.3% | down |
| 5 | claude-opus-5-5 | Anthropic | 3.7% | up |
| 6 | agnes-image-2.5-flash | Agnes AI | 3.5% | up |
| 7 | gpt-6-sol | OpenAI | 2.0% | up |
| 8 | glm-5.3-flash | Zhipu AI | 1.7% | down |
| 9 | gpt-6.1-sol | OpenAI | 1.6% | up |
| 10 | gemini-3.8-flash | 1.5% | up | |
| 11 | claude-sonnet-5 | Anthropic | 1.5% | flat |
| 12 | agnes-2.5-flash | Agnes AI | 1.5% | down |
| 13 | gpt-5.6-luna | OpenAI | 1.3% | flat |
| 14 | claude-opus-4-8 | Anthropic | 1.2% | flat |
| 15 | claude-sonnet-5-5 | Anthropic | 1.2% | up |
| 16 | deepseek-v4-pro | DeepSeek | 1.1% | down |
| 17 | gpt-5.6-sol | OpenAI | 1.0% | down |
| 18 | stepaudio-3-chat-preview | StepFun | 0.9% | up |
| 19 | gemini-3.1-flash-lite | 0.9% | flat | |
| 20 | claude-opus-5 | Anthropic | 0.9% | flat |
| All other models | 38.3% |
How those calls arrived: 11.9% over agent protocols (Claude Messages and the OpenAI Responses API), 81.7% over chat protocols (OpenAI chat completions and Gemini generateContent), 6.5% over everything else (audio, images, embeddings, rerank and asynchronous tasks).