Hermes Agent Blender + Which AI Models To Run It With, Matched To Real GoldieBench 3D Scores

Hermes Agent Blender puts a model in charge of Blender's Python, so the model you pick decides how well your agent writes scene code, lights a shot and checks its own work.

The official Blender plugin for Hermes, version 0.4.0 from 24 October 2026, hands that model full bpy access through Blender Lab's MCP server.

Below I match models to Blender jobs using only scores I read on live GoldieBench pages on 1 November 2026, each quoted with the page it came from.

If you want my team to help plan the agent side of your marketing, book a free SEO strategy session.

Why the model matters more with Blender than with most plugins

The video shows the older Blender MCP route; this page covers the newer official plugin, which installs Blender itself after one yes.

Either way, the main tool is execute_blender_code, which runs whatever Python the model writes inside Blender.

The plugin's workflow skill asks the model to read the scene first, change it in small steps and query the result each time.

So you need a model that writes correct 3D code on the first try and follows a strict loop without drifting.

What GoldieBench measures, and how it maps to Blender

Per the methodology page, every model gets the same fixed prompt, one shot, and returns a single HTML file that Julian scores from 0 to 10.

No GoldieBench task drives Blender.

The closest signal is the 3D tasks: procedural terrain, a solar system and a path tracer, all of which test whether a model can build and light a 3D scene in code.

Read those as a proxy for scene-building skill, not a Blender score.

The overall board, as read on goldiebench.com

Claude Opus 5 shows 50 tasks scored and Hermes MoA 47, per their model pages.

The 3D task scores that matter for Blender

Sources: /tasks/terrain (field average 7.43, winner Fugu Mini at 8.5), /tasks/solar (average 7.57, Fugu Ultra and Fusion at 9.0) and /tasks/pathtracer (average 7.49).

Claude Opus 5 and Hermes MoA show medals on these pages rather than a number in my read, so I've left them out of the table instead of guessing.

Model picks for each Blender job

Best all-rounder for scene work: Claude Opus 5

Claude Opus 5 sits at #2 with 8.27, and the leaderboard lists Sims as its best category.

Its model page lists a 1,000,000-token context window, which leaves room for long scene summaries and the Blender API docs the skill tells the agent to search.

Use it as the default when a Blender job mixes reading a scene, writing code and checking the result.

Best for lighting and renders: Claude Fable 5

Claude Fable 5 won the Path Tracer task at 8.7, the most physically demanding light test on the bench.

That's the closest GoldieBench gets to "does this model understand light bounce, glass and shadows".

Point it at jobs where the render is the deliverable, such as product stills.

When to pick Hermes MoA

Hermes MoA ranks #3 at 8.17 and is listed as best at Visuals.

Its page describes a panel of Opus 4.8 and GPT-5.5 whose drafts an Opus 4.8 aggregator merges, run from the Mixture tab in the Agent OS.

That suits one-off hero scenes where you want several drafts weighed, not fast loops of small edits.

Kimi K3 and Gemini 3.6 Flash: read the split

Kimi K3 scored 8.6 on Path Tracer but 4.5 on Terrain, so its 3D results swing by task.

Gemini 3.6 Flash flips that: 8.4 on Terrain and Solar, then 3.5 on Path Tracer.

Both list context windows of about a million tokens on their model pages, 1,048,576 for Kimi K3 and 1,000,000 for Gemini 3.6 Flash.

Use either for layout and scene-blocking jobs, and hand final lighting to Fable 5 or Opus 5.

The readback rule covers for model mistakes

Even a top model will sometimes write Blender code that runs but doesn't do what you asked.

The skill is built for that: it tells the model that a clean transport reply is not proof the scene changed, and to query the state after every edit.

There's one trap the model has to know about: readback code must set result to a dict with string keys, and any other type is reported as an error after the code has already run.

A model that ignores that line retries edits that already happened, which is how scenes end up with duplicate objects.

That's another reason to favour models that score well on instruction-heavy 3D builds rather than raw speed.

What to put in the model's instructions

• Read the scene with get_objects_summary before any change.
• Make one small change per call and read it back.
• Ask before editing a saved project you didn't request edits on.
• Render once, at the end, with render_viewport_to_path, and report the real file path.

These mirror the plugin's own skill, and stating them again keeps cheaper models on the rails.

A pairing plan for Hermes Agent Blender

1. Setup and troubleshooting: any strong model; the skill's steps do the heavy lifting.
2. Scene edits and readbacks: Claude Opus 5 as the default.
3. Final lighting and renders: Claude Fable 5.
4. Big one-off hero scenes: Hermes MoA.

Whatever you pick, the bridge runs any Python the model writes, and Blender Lab's warning is that it isn't a sandbox, so test on a disposable scene.

Context size is a minor factor here

Every model in the pairing plan lists a context window big enough for normal Blender work.

A scene summary and a few edits rarely need much room, so don't pick on context alone.

Pick on how cleanly the model writes 3D code, which is what the task scores above show.

Where Meshy fits in the model choice

The Meshy plugin makes the 3D asset, and Meshy's own models do that generation, so your Hermes model mostly routes the calls there.

The model choice above matters once the asset lands in Blender and needs staging, lighting and a render.

Our Hermes Meshy breakdown and Hermes 3D page cover the generation side.

Check the Hermes Agent releases and Hermes docs as the plugin moves, and find the full Agent OS model setups inside the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe six-step setup for the official Blender plugin, from install to first render
• AI Income Deskwhether the Blender plugin earns its place in a small business, with three use cases
• AI Tool Verdictthe scorecard verdict on the official Blender plugin, strengths and limits
• agentos.guidefive copy-paste prompts that run Blender from your Agent OS

Set Claude Opus 5 as your default, move renders to Claude Fable 5, and for help planning the rest, book a free SEO strategy session or join the AI Profit Boardroom to run the right models with Hermes Agent Blender.