China AI AtlasAlpha
Tech Buzz ChinaandAI ProemChina Research Collective
Key Takeaways
Model matchup

How China's models stack up

China's flagship models land within a few points of the best proprietary systems on Artificial Analysis Intelligence Index (via OpenRouter) — at a fraction of the price. Compare any two head-to-head, or scan the field below.

Last updated August 16, 2026 · Pricing August 16, 2026 · Benchmarks 2026-08-14

Head-to-head

The fastest way to see how a Chinese model compares to a Western one. Pick two and read the one-line verdict.

Pick any two models to compare. The radar is the shape at a glance — each spoke is scaled so the outer edge = best in class, so a bigger shape is stronger overall. The table has the exact numbers; bold = the winner on that row.

DeepSeek V4 Pro leads on price, value & open weights. GPT-5.6 Sol leads on intelligence, context & modality.

IntelligenceValue /$AffordabilityContextModalityOpen weights
DeepSeek V4 ProGPT-5.6 Sol
Intelligence (AA)43.760.9
Blended $/1M$1.46$11.25
Value (intel/$)29.95.4
Context1049K1050K
Modalitytextfile, image, text
Open weightsYesNo

Price vs. ability

Bottom-right is the sweet spot: smarter and cheaper.

$0.1$0.25$0.5$1$2$5$10$20010203040506070More capable → (Artificial Analysis Intelligence Index)Blended $ / 1M (log) ↑cheaper + smarter ↘Claude Fable 5 — intelligence 62.1, $20/1M blendedGPT-5.6 Sol — intelligence 60.9, $11.25/1M blendedClaude Opus 5 — intelligence 63.1, $10/1M blendedKimi K3 — intelligence 59.7, $6/1M blendedGemini 3.1 Pro — intelligence 47.7, $4.5/1M blendedMistral Medium 3.5 — intelligence 30.4, $3/1M blendedGPT-5.6 Terra — intelligence 56.6, $2.25/1M blendedGrok 4.3 — intelligence 37.9, $1.562/1M blendedDeepSeek V4 Pro — intelligence 43.7, $1.46/1M blendedKimi K2.6 — intelligence 45.1, $0.976/1M blendedGemini 3.7 Flash — intelligence 56, $0.75/1M blendedQwen3.6 Plus — intelligence 40.5, $0.731/1M blendedMiMo V2.5 Pro — intelligence 42.9, $0.544/1M blendedMiniMax M3 — intelligence 45.4, $0.525/1M blendedGLM-5.2 — intelligence 52.6, $0.473/1M blendedStep 3.7 Flash — intelligence 30.9, $0.438/1M blendedLlama 4 Maverick — intelligence 14.5, $0.35/1M blendedGPT-5.6 Luna — intelligence 52.3, $0.225/1M blendedDeepSeek V4 Flash — intelligence 40.3, $0.076/1M blendedClaude Fable 5GPT-5.6 SolClaude Opus 5Kimi K3Gemini 3.1 ProMistral Medium 3.5GPT-5.6 TerraGrok 4.3DeepSeek V4 ProKimi K2.6Gemini 3.7 FlashQwen3.6 PlusMiMo V2.5 ProMiniMax M3GLM-5.2Step 3.7 FlashLlama 4 MaverickGPT-5.6 LunaDeepSeek V4 Flash
Chinese labs US & Europe

All models

Click any column to sort — highest intelligence, lowest cost, best value (intelligence per dollar), longest context. DeepSeek V4 Flash leads on value.

20 models
Model▼Intelligence▼$/1M▼Value▼Context▼ModalityOpen
Claude Opus 5US
63.1
$10
6.3
1M
TIAVF
·
Claude Fable 5US
62.1
$20
3.1
1M
TIAVF
·
GPT-5.6 SolUS
60.9
$11.25
5.4
1.1M
TIAVF
·
Kimi K3CN
59.7
$6
10
1M
TIAVF
●
GPT-5.6 TerraUS
56.6
$2.25
25.2
1.1M
TIAVF
·
Gemini 3.7 FlashUS
56
$0.75
74.7
1M
TIAVF
·
GLM-5.2CN
52.6
$0.473
111.2
1M
TIAVF
●
GPT-5.6 LunaUS
52.3
$0.225
232.4
1.1M
TIAVF
·
Gemini 3.1 ProUS
47.7
$4.5
10.6
1M
TIAVF
·
MiniMax M3CN
45.4
$0.525
86.5
1M
TIAVF
●
Kimi K2.6CN
45.1
$0.976
46.2
262K
TIAVF
●
DeepSeek V4 ProCN
43.7
$1.46
29.9
1M
TIAVF
●
MiMo V2.5 ProCN
42.9
$0.544
78.9
1.1M
TIAVF
●
Qwen3.6 PlusCN
40.5
$0.731
55.4
1M
TIAVF
●
DeepSeek V4 FlashCN
40.3
$0.076
530.3
1M
TIAVF
●
Grok 4.3US
37.9
$1.562
24.3
1M
TIAVF
·
Step 3.7 FlashCN
30.9
$0.438
70.5
262K
TIAVF
●
Mistral Medium 3.5EU
30.4
$3
10.1
262K
TIAVF
·
Llama 4 MaverickUS
14.5
$0.35
41.4
1M
TIAVF
●
Hunyuan A13BCN—
$0.247
—131K
TIAVF
●
Modality — T text · I image · A audio · V video · F files● open weightsValue = intelligence per dollarClick a column header to sort.
Methodology

Blended $/1M tokens at a 3:1 input:output weighting. Intelligence = Artificial Analysis Index (blends coding, math, reasoning, agentic benchmarks), as republished by OpenRouter. Pricing is OpenRouter listed rates; native API pricing may differ. Intelligence as of 2026-08-14; pricing via OpenRouter, 2026-08-16. Benchmark configs vary by model, so treat the ability axis as directional. We refresh as new readings land.