The models we test

The current field for each area, with the newest version, list price and whether it works in mainland China without a VPN. Checked 29 Oct 2026; we refresh it before every monthly test round.

Video

ModelLatest versionWhy it's on the benchList priceChina
Seedance 2.5ByteDance23 Jun 2026#1 image-to-video on Artificial Analysis (1148). 30 s, 50 references, first/last frame, native audio. 4K is announced, not live.from $0.134/sYes
Kling 4.0Kuaishou15 Oct 20263–30 s, up to 4K, lip-sync, Omni Reference with 15 assets. Official API is still “coming soon”.WaveSpeed $0.157–0.235/sYes
Wan 3.0Alibaba24 Aug 2026#1 text-to-video on Artificial Analysis (1174). Closed weights.$0.05–0.20/sYes
Gemini Omni 1.1 FlashGoogle~27 Aug 2026#1 text-to-video on Arena; replaces Veo 3.1 Preview. Extends to 40 s.~$0.10/s at 720p?
MiniMax Hailuo H3 / H3 Max3 Aug 2026#3 on both Artificial Analysis video boards; up to 9 reference images.$0.05–0.08/s?
FLUX 3 Video / MaxBlack Forest LabsMax 20 Oct 2026#2 text-to-video on Arena; native audio.??
HeyGen Video 1late Sep 2026#4 text-to-video on Artificial Analysis; cheapest strong model right now.~$0.015–0.03/s?
PixVerse V630 Mar 2026Multi-shot from one prompt and on-screen Chinese text, both core to comic and short drama.??

Also on the list: Luma Ray3.2 (HDR, 16 keyframes), Vidu Q3 Pro, Grok Imagine Video 1.5, LTX-2.5 (open weights), Runway Gen-4.5.

Image

ModelLatest versionWhy it's on the benchList priceChina
GPT Image 2.5 Sunburst / FlareOpenAI8 Sep 2026#1 and #2 on both Artificial Analysis and Arena text-to-image. In our 10-poster test (31 Oct) it got all 58 Chinese characters right.~$0.21 at max; 0.5 Higgsfield credits at 1K medium?
Nano Banana 2.1Google6 Oct 2026Up to 14 reference images (4 characters, 10 objects), so it is the character-consistency benchmark.~$0.034?
Seedream 5.0 ProByteDance12 Aug 2026Strongest China-native option; long text stable.from $0.045Yes
Midjourney V8.130 Apr 2026Still the taste benchmark; weak on exact text.$10–120/mo?
FLUX 3 ImageBlack Forest Labsannounced 23 Jul 2026Native 4K and layout control with bounding boxes.??
Ideogram 4.0Jun 2026Won a designer typography test; open weights.~$0.06?
Recraft V4.114 May 2026Native SVG for logos and icons; text unreliable past ~5 words.$0.035?
Qwen-Image 3.0Alibaba~Jul 2026Readable 10-px text and infographic grids; invite-only API.?Likely

Also on the list: Krea 3 (15 Oct), MAI-Image-2.6, Grok Imagine Image 2.0, Adobe Firefly Image 5.

Writing, work and study (chat models)

ModelLatest versionWhy it's on the benchList priceChina
Claude Opus 5.5Anthropic22 Sep 2026#2 on the Artificial Analysis Intelligence Index (58); leads WebDev Arena; most-praised prose.$4 / $20No
GPT-6 familyOpenAIAstra 3 Sep; in ChatGPT for all 7 Oct; 6.1 Sol 29 Sep; 6.1 Terra 26 OctMost-used assistant; Sol is near Astra at under a quarter of the cost.Astra $10/$50, Sol $2/$10No
Gemini 4 ArgonGoogleGA 13 Oct 2026Best long-document recall and a low hallucination rate. Gemini 4 Carbon (#1 on Artificial Analysis, 59) is partner-only for now.$2/$10 offer, then $4/$20No
DeepSeek V4.1 Flash-101717 Oct 2026#1 on OpenRouter by tokens; MIT open weights; very cheap.~$0.15 / $0.60Yes
Kimi K3.1 ProMoonshot23 Oct 2026Best open-weights score on Artificial Analysis (46); strong long PDFs.$2 / $10Yes
Doubao 2.1 ProByteDance23 Jun 2026382M monthly users in China; good youthful and social copy.¥6 / ¥30Yes

Also on the list: Qwen3.8-Max, GLM-5.3, Grok 4.7, Meta Muse Wonder, Tencent Hy4 preview (元宝 for WeChat tone).

“?” means not verified yet. Prices are list prices from the maker or the named reseller and change often; see the price tracker.