Hermes Agent 3D Model + Which AI Models To Drive It: Real GoldieBench Scores For Meshy And Blender Work

A Hermes Agent 3D model comes out of Meshy, but the AI model you run as Hermes' brain decides how well the prompt is written, which tools get called and how cleanly Blender is driven afterwards.

So the question isn't only "can Hermes make 3D?" It's "which model should sit behind it?"

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

If you'd like my team to help plan the agent side of your content, book a free SEO strategy session.

Two brains in one 3D job: Meshy's and yours

The geometry itself comes from Meshy, through the official Meshy plugin, listed at v0.5.0 on the Hermes catalog page.

Your chosen model never draws a polygon. It reads your request, picks among 28 Meshy tools, writes the generation prompt and decides on format, detail level and next steps.

The polish stage is different. The Blender plugin connects Hermes to the Blender Lab MCP server, whose tools run Python inside Blender.

That stage leans hard on your model's coding and spatial reasoning.

For the plugin basics, our Hermes Meshy explainer covers setup; this page is about model choice.

What GoldieBench measures, and why it's relevant to 3D

The methodology page sets the rules: same fixed prompt for every model, one shot, one self-contained HTML file out.

Julian scores each result 0 to 10, and the leaderboard ranks by the average across scored tasks.

Several tasks are real-time 3D builds: a voxel sandbox, a procedural terrain explorer, a solar system and a path tracer.

None of them call Meshy, so read the scores as a signal of 3D scene reasoning and graphics coding, not of Meshy output quality.

The overall leaderboard, as read on 1 November

The overall average is a good default, but 3D work rewards looking at the task pages too.

The 3D task scores that matter most

Two spreads stand out. Kimi K3 scored 4.5 on Terrain but 8.6 on Pathtracer.

Gemini 3.6 Flash went the other way: 8.4 on Terrain and only 3.5 on Pathtracer.

So a single average hides real swings on 3D work.

Best brain for routing Meshy jobs

Routing is the lightest stage: pick the right generation tool, state the credit cost, wait for approval, poll status.

The Meshy skill already carries the rules, including choosing the output format first and never re-creating an uncertain task.

Any model near the top of the leaderboard handles that. Claude Opus 5 at 8.27 across 50 scored tasks is the steady pick.

If you're running lots of small jobs, Gemini 3.6 Flash at 7.08 overall is the lighter option, with its strong Voxelcraft and Terrain scores showing it understands 3D scenes.

Best brain for driving Blender

This is where model choice matters most, because the Blender Lab tools execute Python inside Blender.

Claude Fable 5 is the standout here: 9.0 on Voxelcraft and 8.7 on Pathtracer, the winning score on both pages.

Fusion is the most consistent across all four 3D tasks, never below 8.0 in the scores read.

Avoid models with a weak score on the kind of scene you're building: the Pathtracer page shows Gemini 3.6 Flash at 3.5 on that lighting-heavy task.

Whatever you pick, run it on a disposable scene, as the Blender README advises.

Where Hermes MoA fits

Hermes MoA is a panel of models merged by an aggregator, and the leaderboard tags it best at Visuals with an 8.17 average.

Its model page lists a default panel of Claude Opus 4.8 plus GPT-5.5, aggregated by Opus 4.8.

The same page notes the trade-off: roughly 110 to 140 seconds per single-file build versus one call for a solo model.

For a Meshy job that already takes minutes, that extra wait matters less than it would in chat.

I'd use it for writing the concept and generation prompts on hero assets, not for routine polling.

How to switch models for each stage

You don't need one model for the whole job.

Run the generation chat on your routing model, then open a fresh chat on your Blender model for the polish stage, passing it the saved file path.

That split also helps with spend: the 200-credit limit is per chat, so each stage gets its own budget and its own approvals.

Keep the generation prompt and the seed from Desktop's "Made from" panel, so the Blender chat knows exactly what it's working on.

And if a model stumbles on a task type, check its score on the matching GoldieBench page before blaming the plugin.

A weak Pathtracer score, for example, is a warning sign for lighting and render work in Blender.

Pairings I'd start with

Settings that matter no matter which model you pick

• Every paid Meshy call asks first, and each chat has a 200-credit spend limit.
• Unattended runs refuse paid calls by default, whatever the model.
• Meshy API assets normally expire after three days, so tell the agent to download.
• The plugin needs Node.js 18+ and Hermes 0.22.3 or newer.

A stronger model won't lift those limits; they're set by the plugin, not the brain. Check the Hermes releases and Hermes docs when approval behaviour changes.

My bottom line on model choice

Start on Claude Opus 5 for routing, switch to Claude Fable 5 when Blender gets involved, and test Gemini 3.6 Flash for volume.

Use the task pages, not only the average, because the 3D scores swing more than the leaderboard suggests.

Want the trainings for setting these pairings up? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe six-step operator path from one prompt to a finished 3D file
• AI Income Deskfour real business jobs a Hermes 3D model handles, plus the spend rules
• AI Tool Verdicta nine-stage scorecard for the whole Hermes prompt-to-print 3D workflow
• agentos.guidefive copy-paste prompts that run the whole 3D pipeline in your Agent OS

Start on Claude Opus 5, move Blender work to Claude Fable 5, and if you want help, book a free SEO strategy session or join the AI Profit Boardroom to run your next Hermes Agent 3D model.