Higgsfield has launched GPU Harness, a way for Claude to rent a cloud computer with an NVIDIA GPU from inside a normal chat. You connect it once, type /gpu, pick a card, and Claude installs the software, runs the job and brings the results back. Higgsfield says it works in ChatGPT too.
This kit gives you what it is, the setup, which GPU to pick, what it really costs, how to render a Blender scene, how to run a LoRA or open model, the prompts to copy, and the rules that keep the bill small.

Higgsfield's test: a mech Claude modelled in Blender, then rendered on one rented L40S.
1. What GPU Harness is
GPU Harness rents you a cloud computer with an NVIDIA graphics card, controlled by your AI agent. Higgsfield pitches it for three jobs:
Rendering in Blender, Houdini, Unreal Engine, DaVinci, Nuke and Higgsfield's own Craft apps
Running open-weight models for image, video and text
Training LoRAs, or importing one you already have from Hugging Face
You pay in Higgsfield credits, by the second, only while the machine is on. Higgsfield also says several agents can work on several machines at once, and that you can drive it from your phone.
Presets for surreal effects. The GPU page also lists ready-made Blender presets you can open straight in Claude:
Simulate destruction in Blender (a car hitting a post, for example)
Animate an object assembling or coming apart
Build 3D scenes in Blender
Create cartoon materials in Blender
2. Connect it to Claude
Open higgsfield.ai/mcp and click Connect to Claude.
Sign in to Higgsfield and allow the connection.
Make sure you have Higgsfield credits on your account. Renting a GPU spends them.
In a new Claude chat, type
/gpuand pick a card. The price per hour is shown before anything starts.
For a Blender job, Higgsfield's own tutorial also has Blender open on your computer and connected to Claude, so Claude can build the scene before it rents the GPU to render it.
3. Pick the right GPU
GPU | Memory | Good for |
|---|---|---|
L40S | 48 GB | Rendering and video. Start here for Blender. |
RTX PRO 6000 | 96 GB | Image, video and language models, LoRAs |
H100 | 80 GB | Model inference and fine-tuning |
H200 | 141 GB | Larger models and longer context |
B200 | 180 GB | Large-model inference and training |
B300 | 288 GB | The heaviest, memory-hungry jobs |
Rubin is listed too, but only through Higgsfield's sales team, not on demand.
4. What it really costs
Prices from Higgsfield's GPU page on 20 November 2026, in credits:
GPU | Per minute | Per hour |
|---|---|---|
L40S | 1.183 | about 71 |
RTX PRO 6000 | 1.633 | about 98 |
H100 | 2 | 120 |
H200 | 2.233 | about 134 |
B200 | 2.933 | about 176 |
B300 | 3.283 | about 197 |
Real examples from Higgsfield: a 62-frame walk loop at 1920x1080 rendered in 6 minutes 52 seconds on one L40S and cost 16.79 credits for the whole 14-minute rental. A 75-second film of 2,167 frames took 14 RTX PRO 6000 cards, 4 hours 11 minutes and 5,653 credits.
Higgsfield is running a Black Friday deal it describes as "up to 64% off" GPUs. It does not say what the discount is measured against, so compare the live price in /gpu before you rent.
5. Render a Blender scene
Build or open your scene in Blender, connected to Claude.
Ask Claude to render one test frame first and show it to you.
Approve it, then let Claude render the full sequence on the rented card.
Have Claude check every frame is there, assemble the video and stop the rental.
Ask for a report: GPU model, render time, credits spent.
The big version: a 75-second film written in code. In Higgsfield's Houdini tutorial, a black liquid crawls around a glass tank, bursts across the panes and pulls itself back together into the Higgsfield logo. Here is how it was made:
Claude wrote all of it: about 23,000 lines of scripts in one week. Houdini moves the liquid and Blender turns it into pictures. Nothing was built by hand.
The cheap parts ran locally: the simulation and the draft renders ran on a single Mac, so the rented GPUs only did the full-quality frames.
The render: 2,167 frames on 14 rented machines with one RTX PRO 6000 each, 4 hours 11 minutes, 5,653 credits (about 2.6 per frame).
Start small: Higgsfield says a 150-frame test shot cost about 640 credits.
You need Claude Code, Houdini (paid) and Blender (free) for this one. The full step-by-step requests are in Higgsfield's Houdini tutorial.
One user, Matt Shumer, reported on X that Fable 5.2 built a replica of Tokyo's Shibuya district in Blender in 12 hours, with NPCs, and that rendering it on 8 NVIDIA L40S cards through the harness took 2 hours, all from his iPhone. That is his report and we have not reproduced it.
6. Run a LoRA or open model
Give Claude a Hugging Face link to a model or a LoRA. It picks a machine, sets it up, loads the weights and runs your job. For language models, pick a card with 96 GB or more (RTX PRO 6000 and up). Higgsfield lists models like Wan 2.2, MiniMax H3 and GPT OSS 20B as fits for those cards.
Agencies are already using it this way: one LoRA per client and style, several machines running at once, one agent per client.
7. Prompts to copy
The first prompt is Higgsfield's own, from its Blender tutorial. We wrote the other two and have not tested them in the product yet.
Render a Blender file safely (Higgsfield's prompt)
Render the walk cycle of the attached Blender file on Higgsfield GPU Harness. Tell me the hourly price before you rent anything and wait for my yes.
- Add a camera that keeps the whole mech in frame, a ground plane and soft studio lighting.
- Render one test frame first and show it to me.
- After I approve it, render the whole loop at 1920x1080 with Cycles.
- Check that every frame is there, assemble the frames into an mp4 that repeats the loop three times, and stop the rental.
Report the GPU model, the render time and the credits spent.Run a LoRA from Hugging Face
Use Higgsfield GPU Harness to run this LoRA: [Hugging Face link]. Pick the cheapest GPU that has enough memory, tell me which one and its price per hour, and wait for my yes. Then generate [number] images of [subject] in this style, save them, show me the first four, and stop the rental as soon as the last image is done. Report the credits spent.Put a cap on the bill
Before you rent any GPU in this chat: tell me the card and the price per hour, estimate how long the job takes and what it will cost in total, and wait for my yes. If the job runs past [number] credits, pause and ask me before you continue. Always stop the rental when the job is done.8. Limits and safety
It is not free. Every minute costs credits, so always ask for the price first and make Claude stop the rental at the end.
Test small. One test frame or one image before a full run saves you from paying for a broken job.
Rubin is not on demand. The newest NVIDIA cards are offered through Higgsfield's sales team only.
The discount has no stated baseline. Treat "up to 64% off" as Higgsfield's marketing until you see the price in
/gpu.Speed claims vary. The page says "up to 20x faster" without saying faster than what.
Sources: Higgsfield GPU, the GPU price list, Higgsfield's Blender tutorial and its Houdini tutorial, all checked on 20/11/2026.
You're reading The AI Leverage, where we help you get sharper with AI every week. Tools, workflows and guides like this one, always free.
Know someone who renders or trains models? Send them this kit.
by bestapps.ai on instagram
