Multi-board

All three LLM boards refreshed on 2026-11-05, and the same models do not land in the same order. This page puts the three scores side by side; cells with no data stay empty rather than being filled with zeros. Roboflow is a vision-detection board and is listed separately below.

Artificial Analysis · 2026-11-05

60.0

Gemini 4 Carbon (high)
Restricted release, not generally available

SuperCLUE · 2026-11-05

73.62

Gemini-4-Argon
Overseas reference, not ranked

OpenCompass SiNan · 2026-11-05

78.47

Kimi K3.1 Pro (max)
Genuine #1 overall

The three numbers cannot be added, averaged or converted. The six models measured by all three boards are the only sample on this page that supports a direct comparison.

Three boards, one table

Sorted by SiNan V3.0 score; models without a SiNan result sit at the bottom. Grey text is the rank, and "Ref." means the board excludes it from ranking — SuperCLUE's treatment of overseas models.

Model / vendor	Highlight	AA v5	SuperCLUE	SiNan V3.0
Kimi K3.1 Pro max Moonshot AI	The only open model topping both Chinese boards	46.5	61.46#1	78.47#1
Claude Opus 5.5 max Anthropic	Second place, 0.22 points back	59.1	—	78.25#2
Claude Sonnet 5.5 max Anthropic	Actually first on the Vals Index	54.5	—	75.56#3
GPT-6.1 Sol max OpenAI	Top four on all three boards	54.6	71.76Ref.	75.19#4
GPT-6 Astra max OpenAI	Fastest task completion	56.9	—	74.3#5
Gemini 4 Argon high Google	The strongest Gemini you can actually use	50.4	73.62Ref.	72.52#6
Doubao-Seed-2.1-Pro 0915 ByteDance	Ranks higher on SiNan than on SuperCLUE	—	42.06#6	71.8#7
Qwen3.8-Max 0902 Alibaba Qwen	On all three boards; best Chinese score on SiNan knowledge	34.0	43.22#5	71.74#8
Grok 4.7 xhigh xAI	Highest in SuperCLUE's reference group	—	49.66Ref.	71.48#9
GLM-5.3 max Z.ai	Third among Chinese open models, strong on terminal tasks	32.0	50.66#3	71.45#10
Gemini 3.8 Flash high Google	Last-gen Flash still on the board	—	—	71.01#11
Hy4 high Tencent Hunyuan	The newest entry this round	—	42.96#5	70.65#12
DeepSeek-V4.1-Flash-1017 DeepSeek	The gap that most needs explaining	32.0	58.95#2	70.4#13
GLM-5.3-Flash Z.ai	A small model matching the flagship	—	47.18#4	—
Gemini 4 Carbon high Google	#1 on AA, and you still cannot buy it	60.0	—	—
Claude Fable 5.1 max Anthropic	Listed on AA only	50.0	—	—
Muse Wonder full Meta	AA only	43.6	—	—

AA scores use each model's highest effort setting: on AA a model's different reasoning tiers each get their own row, so only the top tier is shown here. Ranks follow each board's latest public standings; SuperCLUE treats models within 1 point as tied, so ranks repeat.

Roboflow RF-DETR 1.11.2

RF-DETR-2XL is the first real-time model to break 60 AP on COCO. This is a computer-vision object detector and has nothing comparable with the LLMs above — it shares a page only because all four were checked in the same round.

Model	AP50:95	Latency ms	Params M	Input	License
RF-DETR-N	48.4	2.3	30.5	384	Apache 2.0
RF-DETR-S	53.0	3.5	32.1	512	Apache 2.0
RF-DETR-M	54.7	4.4	33.7	576	Apache 2.0
RF-DETR-L	56.5	6.8	33.9	704	Apache 2.0
RF-DETR-XL	58.6	11.5	126.4	700	PML 1.0
RF-DETR-2XL	60.1	17.2	126.9	880	PML 1.0
YOLO26-N	40.3	1.7	—	—	AGPL-3.0
YOLO26-M	52.5	4.4	—	—	AGPL-3.0
YOLO26-L	54.1	5.7	—	—	AGPL-3.0
YOLO26-X	56.9	9.6	—	—	AGPL-3.0
YOLO11-X	50.9	10.5	—	—	AGPL-3.0

COCO val2017 AP50:95, NVIDIA T4, TensorRT 10.4 FP16, batch 1. Greyed rows are baselines; latency is single-image inference at batch 1. XL and 2XL ship under PML 1.0 and need a separate licence for commercial self-hosting; N through L and every segmentation size are Apache 2.0.

On real-world datasets the gap widens

RF100-VL averages 100 real-world industry datasets and sits closer to actual projects than COCO.

Model	AP50:95	Vs. best
RF-DETR-2XL	63.2
RF-DETR-XL	62.9
RF-DETR-L	62.2
LW-DETR-L	61.5
YOLO26-L	59.3
YOLO11-L	56.5

RF100-VL averages 100 real-world industry datasets and reflects transfer better than COCO. RF-DETR-L leads YOLO26-L by only 2.4 points on COCO, but the gap widens to 2.9 on RF100-VL — the DINOv2 backbone pays off on small datasets.

Three traps when reading this table

Only six models are comparable

Every other model is missing one or two boards. Before concluding that a model is #1 on board A, check whether it even entered board B — many gaps come from absence, not from ability.

Different scales, no conversion

AA is a nine-test composite, SuperCLUE a five-dimension weighted score and SiNan a six-dimension average. The same model can span 30 points across the three — DeepSeek does — and that is not noise, it is three different exams.

#1 is not always shippable

AA's leader, Gemini 4 Carbon, is open only to security teams inside the Fairwind programme and has no model code in the API. Being #1 on a board and being purchasable are two different things.