Hermes Agent For Ecommerce: Which AI Models To Pin To Each Shop Job, With Real GoldieBench Scores

Hermes Agent for ecommerce lets each shop job run on its own AI model, so the useful question is which model belongs on the product-page job, the promo job and the overnight report.

I read every score below on the live GoldieBench pages on 8 November 2026 and lined them up against the jobs an online shop hands to Hermes.

The examples are a skincare brand, a vintage clothing store, a bakery with a web shop and a phone-case seller.

Want my team to pick models for your own store? Book a free SEO strategy session.

What GoldieBench Measures, And What It Leaves Out

The methodology page gives every model one fixed prompt, one attempt, and no retries or picking the best of several runs.

Each result must be a single HTML file that works in a browser, and Julian scores it from 0 to 10.

Two tasks sit close to shop work: Landing, a modern marketing landing page, and Aipbpromo, a motion-graphics video advert built as a web page.

No task scores product descriptions, customer emails or payment summaries, so for those jobs treat the overall average as a loose signal only.

Where Hermes Lets A Shop Pin Each Model

• Per job: the cron docs show hermes cron edit <job_id> --provider <provider> --model <model> for one job, or --pin to lock the current main model onto it.
• All scheduled jobs: hermes config set cron.model <model> covers every unpinned job without changing your chat model.
• Your chat: hermes model switches the main model, and the Hermes README lists Nous Portal, OpenRouter, OpenAI and your own endpoint as options.
• One big prompt: the /moa command sends a single prompt through the default Mixture of Agents preset, then restores your model.

So the skincare brand's sale page and its 6am payments report never have to share one model.

The Frontier Shortlist

Averages, task counts and golds come from the all-models page.

Product And Sale Pages: The Landing Scores

The Landing task page lists 26 models tested, 21 scored, a task average of 8.11 and Fugu Ultra as the winner.

The vintage clothing store would hand Hermes the drop date, the collection notes and its product-copy skill, and ask for one HTML draft of a sale page.

That job gets pinned to one of the 9.0 models, and the owner checks every word against the real stock before anything goes live.

Gemini 3.6 Flash or Kimi K3 at 8.6 make sense when the owner wants several quick variations to compare side by side.

Promo Videos For A Product Drop: The Aipbpromo Scores

The Aipbpromo task page asks for an auto-playing advert with animated stats and a closing call to action.

It is the harder task: 26 tested, 20 scored, a task average of 7.48, and Qwen 3.8 as the winner.

Notice the order flips: Grok ties for top on Landing but scores 7.0 here, while Qwen 3.8 goes the other way.

The phone-case seller planning a new-colour launch would pin Qwen 3.8 or GPT-5.6 Sol to the promo job and keep a 9.0 Landing model on the product page.

Hermes MoA: A Panel For The Big Launch Page

The Hermes MoA page describes a panel of frontier models with a chair that merges the drafts; the default panel is Claude Opus 4.8 and GPT-5.5, aggregated by Opus 4.8.

It averages 8.17 across 47 tasks with 3 golds and 10 silvers, and its deep-dive page ranks it third of 26 ranked frontier models.

On shop-like tasks it is solid rather than top: 8.4 on Landing and 7.0 on Aipbpromo.

Use /moa for the one launch page a year that deserves several opinions, not for daily jobs.

Members of the AI Profit Boardroom share how they set up panels like this inside the Agent OS.

Customer Data On Your Own Machine: The Local Board

Some shops would rather keep order exports off third-party servers, and Hermes accepts your own model endpoint.

The local models page ranks seven strictly on-device models, with nothing leaving the machine.

Qwable 5 27B Coder leads at 7.14/10 over 41 tasks, and the next, Agents-A1, sits at 4.83/10.

That gap says local models suit read-and-summarise jobs better than ambitious page builds.

The bakery could run its morning order summary on a local model and send only the sale-page job to a frontier model.

Jobs The Bench Doesn't Score

• Product descriptions: no task tests short sales copy, so judge drafts against your own skill file.
• Customer reply drafts: Hermes has no approval step on outgoing messages, so whichever model drafts, a person sends.
• Payment summaries: the Stripe catalog entry supplies the data; the model only reads and summarises it.
• Rival price checks: the bundled price-monitor skill does the fetching, so a mid-table model is usually enough.

A Starting Model Map For A Small Shop

These are starting points from the scores, not a guarantee of how any model writes for your products.

Scores also move as new models arrive and old ones get re-benched, so check the live pages again before you change a pin.

The skincare brand, for instance, would revisit its sale-page pick whenever a new model takes gold on the Landing task.

Before You Pin Anything

Run the same product brief through two shortlisted models and keep the one whose draft needs fewer edits.

Then pin it to the job and leave the rest of the shop on your main model.

Our Grok Bot for ecommerce model guide covers the hosted-agent side of the same question.

Members of the AI Profit Boardroom compare their own model picks on the weekly coaching calls.

Also On Our Network

• Agent Operatorsthe eight-step shop desk build, from profile to morning report
• AI Income Desksix small-shop use cases and the risks an owner should weigh
• AI Tool Verdicta ten-row scorecard of shop jobs, with the store-plugin gap
• agentos.guidefive copy-paste shop prompts and the weekly routine around them
• aiprofitboardroom.comthe complete Hermes Agent for Ecommerce guide on the AI Profit Boardroom blog

Shortlist from the scores, test on your own products, and book a free SEO strategy session if you want my team's help choosing.

Pin one tested model to each job, keep every send with a person, and that is the model plan for Hermes Agent for ecommerce.