Hermes Agent For Content Creation + Which AI Models To Pin To Each Content Job, With Real GoldieBench Scores

Hermes Agent for content creation lets you run a different AI model for each content job, so the useful question is which model belongs on the drafting job, the page-building job and the visual job.

I read every score below on the live GoldieBench pages on 9 November 2026 and matched each one to a content job Hermes can run.

The examples are a wedding photographer, a bike shop, a language school and a podcast studio.

Want my team to match models to your own content plan? Book a free SEO strategy session.

What The Bench Measures, And What It Doesn't

The methodology page sets the rules: one fixed prompt per task, one shot, no iteration, and a single self-contained HTML file that has to run in a browser.

Julian scores each result from 0 to 10.

The tasks page lists 50 creative coding tasks across Game, Sim, Visual, Page and Other.

It says a model that ships the Visual tasks cleanly "has design taste — useful for content workflows".

No task scores blog posts, captions or newsletters, so for pure writing the overall average is a rough guide only.

And the bench grades chat models, not the image or video models Hermes calls through FAL.ai or its video backends.

Where Hermes Lets You Choose The Model

• Main model: hermes model sets the default for chat and every unpinned job.
• Per content job: the cron docs let you pin a model to one job with --provider and --model, and set its own --reasoning-effort.
• All scheduled jobs: cron.model gives unpinned jobs their own default, separate from chat.
• Research helpers: the delegation docs let subagents run on a different model under delegation.model.
• Media: the image model is picked in hermes tools, per the image generation page, and the video model is a user setting the agent can't change, per the tools reference.

So the long-form draft, the landing page and the caption run never have to share one model.

A pin is user-owned: the cron page notes the agent can't point a job at a different model on its own.

The Shortlist By Overall Average

All figures come from the models page, which describes Hermes MoA as a panel of frontier models merged by a chair.

Long-Form Drafts: The Podcast Studio's Show Notes

A small podcast studio turns each episode transcript into show notes, a blog post and a thread with the youtube-content skill from the skills catalog.

No bench task tests writing, so I'd lean on the overall average here: Claude Opus 5 at 8.27 and GPT-5.6 Sol at 8.16 lead the single-vendor models.

1. Pin Claude Opus 5 or GPT-5.6 Sol to the weekly show-notes job.
2. Set --reasoning-effort high on that job only.
3. Let the house-style skill and the humanizer pass handle tone.
4. Send the drafts to the studio's review chat, marked draft only.

Landing Pages: The Bike Shop's Spring Service Page

A bike shop wants a one-page spring service offer built as HTML with the bundled claude-design skill.

The Landing task page shows an 8.11/10 average with 21 of 26 attempts scored, and names Fugu Ultra as the winner.

Fusion, GLM-5.2, Grok and Fugu Ultra each score 9.0/10 there, GPT-5.6 Sol 8.7 and Gemini 3.6 Flash 8.6.

Claude Fable 5 scores 7.2 on that task, lower than its 8.10 overall average suggests.

1. Pin GLM-5.2 or Grok to the page job.
2. Feed it the service list, opening hours and two photos.
3. Review the HTML file before your web person uploads it.

Motion Promos: The Language School's Course Teaser

The Aipbpromo task asks for a motion-graphics video advert built as a web page, the closest bench task to an animated promo.

Its average is 7.48/10, with 20 of 26 attempts scored, and the winner is Qwen 3.8.

GPT-5.6 Sol takes silver at 8.3, Gemini 3.6 Flash scores 7.8, Claude Fable 5 and GLM-5.2 7.5, Grok 7.0 and Kimi K3 6.5.

For a language school's course teaser, I'd pin Qwen 3.8 or GPT-5.6 Sol to the job that writes the animated HTML.

Real video clips are a separate step: video_generate runs on whichever backend you set, which this bench doesn't score.

Visual Taste: The Wedding Photographer's Infographics

A wedding photographer wants planning-timeline infographics with the bundled baoyu-infographic skill.

Three Visual task pages show where design taste sits.

GPT-5.6 Sol scores 8.6 or higher on all three, which makes it the steady pick for visual layout jobs.

Claude Fable 5 is strong on Synthwave and Aurora but weaker on the Landing page.

Captions And Research Helpers: Keep Them Light

Short captions and research summaries don't need the top model.

Point subagents at a lighter model under delegation.model and keep the strong model for the final draft.

If you leave that key empty, the docs say subagents use the same model as the parent.

The bike shop from earlier could run its weekly caption batch this way: three subagents each draft captions for one product line, and the main agent picks and polishes the best.

The bench can't tell you which light model writes the best caption, so judge that from a week of real drafts in your review chat.

For live X research, the separate x_search tool needs xAI credentials and is read-only, so it works the same whatever chat model you pin.

Audio, Images And 3D Sit Outside The Bench

The official ElevenLabs Music and ElevenLabs Sound Effects plugins, the official Meshy 3D plugin and text_to_speech each run on their vendor's own models.

Your chat model only writes the prompt they receive.

So a stronger chat model helps by writing clearer music briefs and image prompts, not by changing the generator's output quality.

What The Scores Don't Change

• Hermes has no built-in hold before a post goes out; its approvals cover risky commands, so every content job still says "draft only".
• Publishing servers still need trust: untrusted so write calls ask first.
• A high score on a one-shot HTML task doesn't prove a model writes in your voice; the house-style skill does that work.

For campaign-side model picks, see Hermes Agent for marketing model scores.

Also On Our Network

• Agent Operatorsthe nine-step setup for a draft-only Hermes content studio, commands included
• AI Income Desksix small-business content jobs, from salon newsletters to 3D product previews
• AI Tool Verdicta nine-criteria scorecard rating Hermes as a content studio
• agentos.guidefive copy-paste prompts that turn one video into a week of drafts
• aiprofitboardroom.comthe complete Hermes Agent for Content Creation guide on the AI Profit Boardroom blog

Pin one model per job, check the drafts against these scores for a fortnight, and book a free SEO strategy session to aim the output at real searches, because the right model is only half of Hermes Agent for content creation.