AI model comparison — text, image and video
The leading AI models from the US, Europe and China — text, image and video generation — compared by quality (market benchmarks), cost and open-source status, always separating remote (API) and local (on your hardware) models.
Data as of 2026-07-25 · automated research (Artificial Analysis, LMArena, official pricing) — verify before deciding.
📊 How quality is measured — three indices
We show quality three complementary ways. Here is how each index is built before you read the charts:
① Zendoric Quality (0-100) = equal-thirds average of SWE-bench-Pro (33% · real software development, checked against the maker) + LMArena (33% · human preference, normalised Elo) + Terminal-Bench (33% · agentic terminal capability). If a model lacks one of the three, its weight is shared among those present (at least two required).
② AA Index (Artificial Analysis Intelligence Index, 0-100) = a broader composite index (reasoning, science, code, maths). It gives a second reading: depending on how you measure, the maker ranking changes.
③ Cybersecurity (0-100) = capability on expert cyber tasks (hard «unguided pass@1» protocol: vuln-research and realistic exploitation). We use a non-saturated metric (the top is around 71, not 100, leaving headroom), not the Cybench «pass@k» the frontier already saturates. Sources: UK AISI, NIST-CAISI, CVE-Bench. We frame it as capability and risk, not an offensive ranking; where there is no direct eval it is estimated «est.».
📈 Zendoric Quality over time (frontier makers)
Quality index (0-100) of the top makers (their best model), last 24 months. Dashed line = quality estimated from the AA Index (labs without SWE-bench-Pro). Updated daily.
📈 AA Index over time (frontier makers)
AA Index (Artificial Analysis Intelligence Index, 0-100) of the top makers (their best model), last 24 months. It is a broader composite index (reasoning, science, code, maths) than ours. The historical series is reconstructed by anchoring each maker's trajectory to its current AA. Updated daily.
🛡️ Cybersecurity over time (frontier makers)
Cybersecurity index (0-100) of each maker's best model, last 24 months. Metric: EXPERT cyber tasks under a hard «unguided pass@1» protocol (no hints, one attempt; vuln-research and realistic exploitation). We pick it because it is NOT saturated — the top is around 71, not 100, so it discriminates and shows headroom (we drop Cybench «pass@k», where the frontier already scores ~100%). Sources: UK AISI (GPT-5.5 71.4% vs Anthropic preview 68.6%), NIST-CAISI, CVE-Bench. High confidence only for OpenAI/Anthropic (measured by AISI); the rest imputed by proximity → the whole series is marked «est.». We frame it as capability and RISK to govern, not an offensive ranking. Updated daily.
💰 Zendoric Quality vs cost
Flagship models of the top makers by quality (a maker may have several, e.g. Anthropic: Opus 5, Opus 4.8 and Fable 5). HIGHER = more quality; LEFT = cheaper (log axis). Hollow dot = quality estimated (AA Index). Colour by maker.
💰 AA Index vs cost
Same format as the quality/cost chart, but the vertical axis is the AA Index. HIGHER = more capability; LEFT = cheaper (log axis). Hollow dot = estimated AA (Terminal-Bench/SWE-Pro). Colour by maker.
🏁 Efficient frontier — quality vs cost
Every cloud model with data. Y = AA Index; X = output cost, log and REVERSED: further RIGHT is cheaper — the top-right corner is ideal. The green line joins the efficient frontier: models with no alternative that is both better and cheaper; picking off the line is only justified by non-price factors (ecosystem, context, open source). Hollow dot = estimated AA. Hover a dot for details.
🏆 Zendoric Quality (SW dev + arena + agentic)
| Model | Zendoric Quality | SWE-bench-Pro | DeepSWE | LMArena | Terminal-Bench | LiveCodeBench | GPQA | ARC-AGI-2 |
|---|---|---|---|---|---|---|---|---|
| 🇨🇳 Kimi K3Moonshot AI · China | 90.7 | — | 67.5 | 1486 | 88.3 | 68.0 | 93.5 | — |
| 🇺🇸 Claude Fable 5Anthropic · USA | 90.1 | 80.3 | 69.7 | 1515 | — | 91.7 | 92.6 | — |
| 🇺🇸 Claude Mythos 5Anthropic · USA | 88.1 | 80.0 | — | 1507 | 84.3 | 91.7 | 94.1 | — |
| 🇺🇸 GPT-5.6 SolOpenAI · USA | 78.9 | 63.0 | 72.7 | 1470 | 88.8 | — | 87 | — |
| 🇺🇸 Grok 4.5xAI · USA | 77.3 | 64.7 | 53 | 1468 | 83.3 | — | 93.1 | — |
| 🇨🇳 GLM-5.2Zhipu AI · China | 76.9 | 62.1 | 46.2 | 1475 | 81.0 | 80.2 | 78 | 7 |
| 🇺🇸 Claude Opus 4.8Anthropic · USA | 76.5 | 69.2 | 59.0 | 1455 | 82.7 | 88.8 | 84 | 14 |
| 🇺🇸 GPT-5.5OpenAI · USA | 76.3 | 58.6 | 67.0 | 1475 | 82.7 | — | 85 | 16 |
| 🇺🇸 Claude Sonnet 5Anthropic · USA | 74.7 | 63.2 | 53.8 | 1461 | 80.4 | — | 83 | 12 |
| 🇺🇸 Grok 4xAI · USA | 74.2 | — | — | 1430 | 83.3 | 79.4 | 84 | 16 |
| 🇨🇳 Qwen3.7-MaxAlibaba · China | 72.6 | 60.6 | — | 1475 | 69.7 | 91.6 | 81 | 7 |
| 🇨🇳 Kimi K2.6Moonshot AI · China | 68.4 | 58.6 | — | 1460 | 66.7 | 89.6 | 78 | 9 |
| 🇨🇳 DeepSeek V4-ProDeepSeek · China | 66.1 | 55.4 | — | 1450 | 67.9 | 93.5 | 82 | 9 |
| 🇺🇸 Gemini 3 ProGoogle · USA | 65.8 | 43.3 | — | 1501 | 54.2 | 91.7 | 84 | 15 |
| 🇺🇸 Claude Sonnet 4.6Anthropic · USA | 63.4 | 58.1 | 29.9 | 1430 | 67.0 | — | 80 | 9 |
| 🇺🇸 MAI-1-previewMicrosoft · USA | 49.4 | 52.8 | — | — | 46.0 | 87.7 | 84.2 | — |
| 🇺🇸 Claude Opus 5Anthropic · USA | — | 79.2 | — | — | — | — | — | — |
| 🇺🇸 Llama 4 MaverickMeta · USA | — | — | — | 1288 | — | 43.4 | 70 | 5 |
| 🇪🇺 Mistral Large 3Mistral AI · Europa | — | — | — | 1415 | — | 74 | 72 | 6 |
| 🇪🇺 Magistral SmallMistral AI · Europa | — | — | — | — | — | 70.88 | 70.07 | 4 |
| 🆕 🇺🇸 GPT-5.6 TerraOpenAI · USA | — | — | — | — | 87.4 | — | — | — |
| 🆕 🇺🇸 GPT-5.6 LunaOpenAI · USA | — | — | — | — | 84.7 | — | — | — |
| 🆕 🇺🇸 Gemini 3.6 FlashGoogle · USA | — | — | 49 | 1485 | — | — | — | — |
| 🆕 🇨🇳 Qwen3.8-Max-PreviewAlibaba (Qwen) · China | — | — | — | — | — | — | — | — |
Quality = equal-thirds average of SWE-bench-Pro (SW development) + LMArena (human preference) + Terminal-Bench (agentic capability), the three with reliable sources (Zendoric Quality); if one is missing its weight is shared among those present (at least two; otherwise «—»). LiveCodeBench and GPQA are shown for reference (indicative, may be incomplete) but are NOT in the index; ARC-AGI-2 (arcprize.org) tracks AGI progress: models score VERY low → still far from AGI. %, except LMArena (Elo).
💵 Economics (USD / 1M tokens)
| Model | Input | Cache | Output |
|---|---|---|---|
| 🇨🇳 Kimi K3Moonshot AI · China | $3.0 | $0.3 | $15.0 |
| 🇺🇸 Claude Fable 5Anthropic · USA | $10.0 | $1.0 | $50.0 |
| 🇺🇸 Claude Mythos 5Anthropic · USA | $10.0 | $1.0 | $50.0 |
| 🇺🇸 GPT-5.6 SolOpenAI · USA | $5.0 | $0.5 | $30.0 |
| 🇺🇸 Grok 4.5xAI · USA | $2.0 | $0.3 | $6.0 |
| 🇨🇳 GLM-5.2Zhipu AI · China | $0.6 | $0.26 | $2.2 |
| 🇺🇸 Claude Opus 4.8Anthropic · USA | $5.0 | $0.5 | $25.0 |
| 🇺🇸 GPT-5.5OpenAI · USA | $5.0 | $0.5 | $30.0 |
| 🇺🇸 Claude Sonnet 5Anthropic · USA | until Aug 31, 2026 $2.0 from Sep 1, 2026 $3.0 | until Aug 31, 2026 $0.2 from Sep 1, 2026 $0.3 | until Aug 31, 2026 $10.0 from Sep 1, 2026 $15.0 |
| 🇺🇸 Grok 4xAI · USA | $3.0 | $0.75 | $15.0 |
| 🇨🇳 Qwen3.7-MaxAlibaba · China | $1.2 | $0.25 | $6.0 |
| 🇨🇳 Kimi K2.6Moonshot AI · China | $0.6 | $0.16 | $2.5 |
| 🇨🇳 DeepSeek V4-ProDeepSeek · China | $0.28 | $0.03 | $0.87 |
| 🇺🇸 Gemini 3 ProGoogle · USA | $1.25 | $0.31 | $10.0 |
| 🇺🇸 Claude Sonnet 4.6Anthropic · USA | $3.0 | $0.3 | $15.0 |
| 🇺🇸 MAI-1-previewMicrosoft · USA | — | — | — |
| 🇺🇸 Claude Opus 5Anthropic · USA | $5.0 | $0.5 | $25.0 |
| 🇺🇸 Llama 4 MaverickMeta · USA | $0.2 | — | $0.6 |
| 🇪🇺 Mistral Large 3Mistral AI · Europa | $2.0 | $0.2 | $6.0 |
| 🇪🇺 Magistral SmallMistral AI · Europa | $0.5 | $0.05 | $1.5 |
| 🆕 🇺🇸 GPT-5.6 TerraOpenAI · USA | $2.5 | $0.25 | $15.0 |
| 🆕 🇺🇸 GPT-5.6 LunaOpenAI · USA | $1.0 | $0.1 | $6.0 |
| 🆕 🇺🇸 Gemini 3.6 FlashGoogle · USA | $1.5 | $0.15 | $7.5 |
| 🆕 🇨🇳 Qwen3.8-Max-PreviewAlibaba (Qwen) · China | — | — | — |
Claude Sonnet 5: scheduled price increase (same model) — reduced pricing until Aug 31, 2026 and standard pricing from Sep 1, 2026.
🔓 Open source & type
| Model | Open source | License | Type |
|---|---|---|---|
| 🇨🇳 Kimi K3Moonshot AI · China | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Claude Fable 5Anthropic · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Claude Mythos 5Anthropic · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 GPT-5.6 SolOpenAI · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Grok 4.5xAI · USA | No | Proprietary | Proprietary (API only) |
| 🇨🇳 GLM-5.2Zhipu AI · China | Yes | MIT | Open-weight |
| 🇺🇸 Claude Opus 4.8Anthropic · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 GPT-5.5OpenAI · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Claude Sonnet 5Anthropic · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Grok 4xAI · USA | No | Proprietary | Proprietary (API only) |
| 🇨🇳 Qwen3.7-MaxAlibaba · China | No | Proprietary | Proprietary (API only) |
| 🇨🇳 Kimi K2.6Moonshot AI · China | Yes | Modified MIT | Open-weight |
| 🇨🇳 DeepSeek V4-ProDeepSeek · China | Yes | MIT | Open-weight |
| 🇺🇸 Gemini 3 ProGoogle · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Claude Sonnet 4.6Anthropic · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 MAI-1-previewMicrosoft · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Claude Opus 5Anthropic · USA | No | Proprietary | Proprietary (API only) |
| 🇺🇸 Llama 4 MaverickMeta · USA | Yes | Llama 4 Community | Open-weight |
| 🇪🇺 Mistral Large 3Mistral AI · Europa | Yes | Apache-2.0 | Open-weight |
| 🇪🇺 Magistral SmallMistral AI · Europa | Yes | Apache-2.0 | Open-weight |
| 🆕 🇺🇸 GPT-5.6 TerraOpenAI · USA | No | Proprietary | Proprietary (API only) |
| 🆕 🇺🇸 GPT-5.6 LunaOpenAI · USA | No | Proprietary | Proprietary (API only) |
| 🆕 🇺🇸 Gemini 3.6 FlashGoogle · USA | No | Proprietary | Proprietary (API only) |
| 🆕 🇨🇳 Qwen3.8-Max-PreviewAlibaba (Qwen) · China | No | Proprietary | Proprietary (API only) |
🖥️ Open source you can self-host
Small/medium models you can run on your own machine (laptop/PC/Mac). Quality = Artificial Analysis Intelligence Index (0-100; output quality), the measure with best coverage of small open models (LMArena does not list sub-32B). Memory estimated at 4-bit (Q4) and 8-bit (Q8) quantization; on Apple Silicon it is UNIFIED memory (RAM=VRAM).
| Model | Quality (AA Index) | GPQA | Params | RAM Q4 | RAM Q8 | GPU | CPU / Mac | License |
|---|---|---|---|---|---|---|---|---|
| Qwen3.5-27BAlibaba | 42 | 85.5 | 27B | 16 GB | 31 GB | ≥16 GB | Limitado (mejor GPU/Mac ≥32 GB) | Apache-2.0 |
| Gemma 4 31BGoogle | 39 | 86 | 31B | 18 GB | 35 GB | ≥24 GB | Limitado (mejor GPU/Mac ≥32 GB) | Gemma |
| Qwen3.5-35B-A3BAlibaba | 37 | 84.2 | 35B | 21 GB | 40 GB | ≥24 GB | Limitado (mejor GPU/Mac ≥32 GB) | Apache-2.0 |
| Gemma 4 26B A4BGoogle | 31 | 82.3 | 26B | 15 GB | 29 GB | ≥16 GB | Limitado (mejor GPU/Mac ≥32 GB) | Gemma |
| Nemotron-Cascade-2-30B-A3BNVIDIA | 28 | 76.1 | 30B | 19 GB | 36 GB | ≥24 GB | Limitado (mejor GPU/Mac ≥32 GB) | NVIDIA Open Model |
| gpt-oss-20bOpenAI | 24 | 71.5 | 20B | 13 GB | 25 GB | ≥16 GB | Limitado (mejor GPU/Mac ≥32 GB) | Apache-2.0 |
| Gemma 4 12BGoogle | 22 | 78.8 | 12B | 8 GB | 15 GB | ≥8 GB | Sí (CPU lento · Mac 16 GB) | Gemma |
| Gemma 4 E4BGoogle | 19 | 58.6 | 4B | 4 GB | 6 GB | ≥8 GB | Sí (CPU/Mac, fluido) | Gemma |
| Gemma 4 E2BGoogle | 15 | 43.4 | 2B | 3 GB | 4 GB | ≥8 GB | Sí (CPU/Mac, fluido) | Gemma |
🗄️ Large open source (server / multi-GPU)
Powerful open models that need a server or multiple GPUs. Quality = LMArena Elo (human preference over output, source lmarena.ai), which does cover large models. For MoE, memory counts total parameters (all experts are loaded). Memory estimated at 4-bit (Q4) and 8-bit (Q8) quantization; on Apple Silicon it is UNIFIED memory (RAM=VRAM).
| Model | Quality (LMArena) | GPQA | Params | RAM Q4 | RAM Q8 | GPU | CPU / Mac | License |
|---|---|---|---|---|---|---|---|---|
| DeepSeek-V4-ProDeepSeek | 1465 | 90.1 | 1600B | 882 GB | 1762 GB | 12× 80 GB (servidor) | No (servidor GPU) | MIT |
| GLM-5.2Zhipu AI | 1465 | 91.2 | 744B | 411 GB | 820 GB | 6× 80 GB (servidor) | No (servidor GPU) | MIT |
| Kimi K2.6Moonshot AI | 1460 | 90.5 | 1100B | 552 GB | 1102 GB | 7× 80 GB (servidor) | No (servidor GPU) | Modified MIT |
| Qwen3.5-397B-A17BAlibaba | 1450 | 88.4 | 397B | 220 GB | 438 GB | 3× 80 GB (servidor) | No (servidor GPU) | Apache-2.0 |
| Llama 4 MaverickMeta | 1420 | 69.8 | 400B | 223 GB | 444 GB | 3× 80 GB (servidor) | No (servidor GPU) | Llama 4 Community |
| Mistral Large 3Mistral AI | 1416 | 68.0 | 675B | 373 GB | 744 GB | 5× 80 GB (servidor) | No (servidor GPU) | Apache-2.0 |
| gpt-oss-120bOpenAI | 1352 | 80.1 | 117B | 66 GB | 130 GB | ≥80 GB | No (servidor GPU) | Apache-2.0 |
🎨 Image generation
The leading IMAGE generation models, split between remote services (pay per image via API) and open models you can run on your own hardware.
☁️ Remote (API / cloud)
| Model | Quality (arena) | Price (per image) | Notes |
|---|---|---|---|
| GPT Image 2OpenAI · USA | 1338 AA Arena nº1 confirmado; 2º Reve 2.1 1300 (era 1305), gap ~38; arena.ai +242 (1512) confirmado | $0,211/img high (=$211/1K API OpenAI); fal.ai $0,005 low a $0,401 high 4K (fal.ai no verif. al detalle) | Nº1 t2i confirmado. Lanzado abr-2026. 2K, texto multilingüe y prompt adherence líder |
| MAI-Image-2.5Microsoft · USA | 1269 AA (era 1265), nº3 t2i confirmado; 2º Reve 2.1 1300 | Azure Foundry $47/M img out confirmado; ~$0,048/img estimado | Microsoft. Nº3 t2i (tras Reve 2.1); Flash más barata (~1207 Elo) |
| Nano Banana Pro (Gemini 3 Pro Image)Google · USA | 1222 AA t2i (era 1218), nº9 de 81 (CORREGIDO, era nº13); fuera del top-3 | $0,134/img 1-2K · $0,24/img 4K Gemini API (Batch 50% → 4K ~$0,12) confirmado | = Gemini 3 Pro Image. Salida hasta 4096×4096 |
| Nano Banana 2 Lite (Gemini 3.1 Flash Lite)Google · USA | 1262 AA t2i (era 1251), nº4 (CORREGIDO, era nº5) | $0,034/img (1K) Gemini API confirmado (Batch $0,017) | = Gemini 3.1 Flash Lite Image. Rápida/barata ~4s, solo 1K |
| GPT Image 1.5OpenAI · USA | 1260 AA (era 1256), nº6 t2i confirmado | $0,009 low · $0,034 medio · $0,133 high (1024 OpenAI API) — tiers no verif. al detalle | Generación anterior OpenAI, aún competitiva (nº6 AA) |
| Grok ImaginexAI · USA | 1203 AA (image-quality) nº12 (era 1201/nº14); variante base 1178 nº14 | $0,02/img confirmado (1K/2K, hasta 3 refs) | xAI. Precio bajo confirmado. i2v nº1 AA (1336) |
| Cosmos 3 SuperNVIDIA · USA | 1217 AA t2i (era 1214); Nº1 open weights confirmado en AA | Open weights (licencia OpenMDW-1.1, self-host gratis); API $0,04/img en fal y WaveSpeed | NVIDIA, omnimodal MoT ~64B (Qwen3-VL 32B). Nº1 open AA t2i confirmado |
💻 Local (open-weight, on your hardware)
| Model | License | Hardware | Notes |
|---|---|---|---|
| FLUX.2Black Forest Labs | open-weight + API | GPU dedicada (variantes dev/FP8) | Lanz. 25-nov-2025 (max 16-dic, klein 15-ene-2026), hasta 4MP; open + Pro/Flex/Max API |
| Qwen-Image 2.0Alibaba | Apache-2.0 | 7B — GPU 16GB+ | Lanzado 10-feb-2026, 7B + enc. 8B Qwen3-VL, 2K nativo, tipografía EN/ZH |
| Z-Image TurboAlibaba (Tongyi) | open-weight | ~1 s/imagen en H100; corre en Mac (MLX) | Alibaba Tongyi-MAI, 6B, sub-segundo en H100. El open más rápido |
| Stable Diffusion 3.5Stability AI | Stability Community | GPU 8-16GB | Upgrade gratis desde SD 3.0. Mayor ecosistema (Civitai, ComfyUI, LoRAs) |
| HiDream-O1-Image-1.5HiDream | open-weight | GPU dedicada | v1.5 CERRADO (hosted), hasta 2K. Top-4 AA confirmado |
Data as of 2026-07-29 · sources: Artificial Analysis Image/Video Arena · arena.ai · llm-stats.com · Pixazo · documentación de fabricantes · arena Elo is blind human preference over the output; indicative prices — verify before deciding.
🎬 Video generation
The leading VIDEO generation models, split between remote services (pay per generated second via API) and open models for your own hardware.
☁️ Remote (API / cloud)
| Model | Quality (arena) | Price (per second) | Notes |
|---|---|---|---|
| Kling 3.0 / TurboKuaishou · China | Con audio 1112 (Pro 1080p, ~#6 t2v). Gemini Omni lidera, no Kling. Sin audio 1244 NO VERIF; arena.ai 1991 NO VERIF | CORREGIDO: fal Kling 3.0 Turbo 1080p ~$0,14/s (no $0,08); ~$2,10 por 15s | Kuaishou feb-2026; Turbo/Omni 17-jun-2026. Hasta 15s + audio nativo |
| Veo 3.1Google · USA | CORREGIDO: NO top-3. Con audio 1096, ~#6-9 t2v. Lidera Gemini Omni Flash en AA | Lite ~$0,05/s · Fast $0,10/s (720p) · Std $0,40/s con audio (720p/1080p) — Vertex/Gemini API | Tres niveles + audio nativo (Vertex AI). Std $0,40/s audio incluido |
| Gemini Omni FlashGoogle · USA | 1244 con audio #1 t2v · 1325 sin audio #1 (orig 1326). Lidera t2v e i2v (i2v 1199 c/audio, 1375 s/audio) | ~$0,10/s (720p,24fps) preview Gemini API/AI Studio. Precio exacto NO VERIFICADO | Google, I/O may-2026. Multimodal, audio nativo, edición conversacional, 3-10s |
| Seedance 2.0 (Dreamina)ByteDance · China | 720p 1228 con audio #2 (orig 1227) · 1272 sin audio #4. Fast 1703 arena.ai NO VERIF | ~$0,10/s std (AA ~$9/min 1080p). EvoLink desde $0,045/s. Fast ~$0,081/s NO VERIF | ByteDance SEED 12-feb-2026. Multimodal, 15s + audio estéreo |
| Sora 2OpenAI · USA | — (ausente del leaderboard AA) | $0,10–0,70/s aún activo hasta cierre (orig '—') | EN RETIRADA: app cerrada 26-abr-2026; API se apaga 24-sep-2026 (410 después). Aviso 24-mar-2026 |
💻 Local (open-weight, on your hardware)
| Model | License | Hardware | Notes |
|---|---|---|---|
| Wan 2.7Alibaba | open-weight | 16-24GB VRAM (14B) | Alibaba, release 22-abr-2026. Suite T2V/I2V/ref-to-video con voz + edición por instrucciones, multi-imagen |
| LTX-2.3Lightricks | open-weight | desde 12GB VRAM (FP8: flujos de 32GB) | Lightricks, 22B DiT, audio+vídeo en una pasada; ~18x más rápido que Wan 2.2 (NO VERIF directo) |
| HunyuanVideoTencent | open-weight | 16-24GB VRAM | Tencent 13B, open dic-2024; sucesor HunyuanVideo-1.5 (8.3B, ~14GB VRAM, Apache 2.0) publicado 20-nov-2025. No en AA |
| Mochi 1Genmo | open-weight | fine-tune en 1×H100/A100 80GB | Genmo, 10B (AsymmDiT), oct-2024, Apache 2.0; 480p preview, alta adherencia al prompt. No en AA |
| SkyReels V1Skywork | open-weight | GPU dedicada | Skywork, feb-2025, primer open human-centric (fine-tune de HunyuanVideo); T2V+I2V. Ojo: SkyReels V4 (1110 AA) es otra versión |
Data as of 2026-07-29 · sources: Artificial Analysis Image/Video Arena · arena.ai · llm-stats.com · Pixazo · documentación de fabricantes · arena Elo is blind human preference over the output; indicative prices — verify before deciding.