What changed: GPU rental price moves over the last seven days, the cheapest card to rent today in every VRAM tier, and everything new on the site. Follow it by RSS at /feed.xml.
Oct 7:25 more Hugging Face models measured: LTX-2.5, Wan 14B, HunyuanVideo, Hunyuan3D, GLM-4.5-Air. The AI models list now tests nearly everything popular on Hugging Face that fits one GPU (up to 288GB). New: LTX-2.3 and LTX-2.5, Wan 2.1 14B text and image to video, HunyuanVideo 1.5, Cosmos-Predict2, AnimateDiff-Lightning and CogVideoX-5B for video; HiDream-I1 Fast, SD3 Medium, SDXL-Lightning and SD 2.1 for images; Hunyuan3D 2.0 and 2mini, Stable Fast 3D and Shap-E for 3D; GLM-4.5-Air, Laguna-S, OLMo 3, Apertus and more for text. The RTX 5060 got its first measured LLM numbers too.
Oct 7:Image-to-video measured on the cards people own. SVD-XT, Wan 2.2 I2V 5B, LTX-Video I2V and CogVideoX-5B I2V now have real speeds on the RTX 3090, 4060 Ti, 4080, 4090, 5060 Ti, 5070 Ti, 5080 and 5090, most of which said "won't fit" before. 16GB cards run them with CPU offload, and the result says so; 12GB cards still don't fit them at these settings, now measured rather than guessed.
Oct 7:TRELLIS.2 on a gaming card: 57 3D assets an hour on an RTX 4090. Microsoft's TRELLIS.2 timed to a finished, textured GLB on an RTX 4090 (63 seconds an asset) and an RTX 3090 (120 seconds), next to the H100, L40S and A100. It peaked at about 14GB of VRAM.
Oct 7:Bad rental hosts caught and re-measured. Some earlier image and video numbers came from power-capped or slow rented machines. Those cards were re-run on two or three different hosts and the site now shows the median, with a badge giving the number of hosts and the spread: the RTX 4080 Super's Z-Image speed went from 0.75 to 2.30 images a minute, the RTX 5080's SDXL from 8.9 to 12.2, the RTX 4090's LTX-Video from 4.7 to 6.7 frames a second.
Oct 7:Rent or buy with your own price, and blind voting in the Shootout. Every GPU page now asks what you paid (or would pay) and works out how many months until owning beats renting at RunPod's rate, electricity included. In the Shootout you can vote between two outputs without seeing the model names, and the community pick builds up from those votes.
Oct 7:More models measured: Qwen3, Gemma 4, GPT-OSS, CogVideoX, Chroma. Qwen3-8B on 13 more cards, GPT-OSS-20B on 12, Gemma 4 12B on 10, CogVideoX-2B text-to-video on 13, plus DreamShaper XL Turbo and Chroma1-HD image models. AMD and Intel cards now list text generation only: image, video and 3D models are built for NVIDIA's CUDA, and we no longer show estimates for them on other cards.
Sep 30:Text-to-video shootout, and cleaner head-to-heads. A new Text to video tab in the Shootout: the same four prompts through Wan 2.1 1.3B, LTX-Video and Wan 2.2 5B, with speeds down to a 12GB RTX 3060. The home head-to-head now groups AI workloads by task (text, image, video, 3D) with a winner for each, and shows when a model won't fit a card instead of dropping the row.
Sep 30:Real AI numbers for the RTX 4080, 4070 Ti Super, 3080 Ti and more. The 22 newest LLMs measured on the RTX 5080, 5070, 4080, 4080 Super, 4070 Ti Super, 4070, 3090 Ti, 3080 Ti, 3080, 3070 and 3060 Ti, plus first real numbers for the Titan V, Quadro RTX 5000, RTX 4500 Ada and A100 40GB PCIe. Six new image models in the Shootout: SD Turbo, LCM DreamShaper and SSD-1B (they fit 4-8GB cards), DreamShaper XL Lightning, Kolors and Baidu's ERNIE-Image Turbo, each timed on the cards people own.
Sep 29:GPU rental prices: every card, both providers, one table. RunPod and Vast.ai prices side by side for every rentable GPU, 30 days of price history, and what a million tokens or a hundred images actually costs on each card.
Sep 29:Shootout: see what every image and video model makes. The same prompts through 23 image models and the same stills through 6 video models, each with its speed on your GPU. Tick up to four to compare side by side. New: LTX-Video and Stable Video Diffusion, the video models that fit a 24GB and a 16GB card, measured on the RTX 3090, 4090, 5090 and 4060 Ti.
Sep 29:The giant models, measured on one GPU. MiniMax M2.7, DeepSeek V4 Flash, Ornith 397B, Qwen3.5 122B and Nemotron 3 Super 120B at 4-bit, plus DeepSeek V3.2 and GLM 5.2 at 2-bit, each on a single GPU. Everything that needs more than one card stays an estimate.
Sep 29:Image models on the cards you own. SD 1.5, SDXL Turbo, Playground v2.5, Z-Image and FLUX.2 klein measured on the RTX 4060 Ti, 3090, 4090 and 5090, and 23 image models in the output gallery.
Sep 29:10 new LLMs measured, and see what every image and video model makes. Qwen3.8 27B, Qwen3.6 27B and 35B-A3B, Gemma 4 26B-A4B and 31B, gpt-oss-120b and four small models, measured on 12 GPUs including the RTX 3060, 4060 Ti, 3090, 4090 and 5090. Every image and video model now shows its output next to its speed: the same prompts through each one.
Sep 29:Speed and quality for every quant on consumer cards. How fast Q3 through Q8 actually run on the RTX 3060, 4060 Ti, 3090, 4090 and 5090, and how much quality each quant costs, measured rather than guessed.
Sep 28:Workflows are now builders: swap any model, set your batch. Every workflow stage is a slot. Swap the image, video or 3D model, switch optional stages on or off and set how many items you need; time and rental cost update on every GPU. Five new batch jobs: photos to 3D, animate images, transcribe and subtitle, voiceovers and dataset captioning.
Sep 28:VRAM for every quant, not just Q4. Every LLM now shows how much VRAM it needs from Q2_K to BF16, and with a VRAM filter or your GPU picked, the best quant that fits.
Sep 28:The home page starts with your GPU. Pick your card once and the front page shows its rank, how many of our AI models it runs, and the cheapest faster card to rent.
Sep 28:140 more models, straight from Hugging Face. The AI models database now also lists the most-downloaded LLM, image, video and 3D models we have not benchmarked yet, with an estimated VRAM floor. Search them the same way.
Sep 28:The AI models database is now searchable. Search 160+ models, filter by task, model size and the VRAM you have, or type your GPU to see everything it can run.
Sep 28:Workflows now show what each job costs to rent. Every pipeline is priced on every rentable card at today's RunPod and Vast.ai rates. Set how many jobs you need and sort by cost or speed.