GoldieBench Blog · 6 min read
Hermes Agent For Sales + Which AI Models To Run On Each Sales Job: Real GoldieBench Scores For Pitch Pages, Drafts And Proposals
Hermes lets every Bot, channel and scheduled job run its own model. Here is how the live GoldieBench scores map onto a sales desk.

Hermes Agent for sales lets every sales job run on its own AI model, so the real question is which model to pin to the briefs, the follow-up drafts and the prospect pages.
Below are the GoldieBench scores I read on the live pages on 7 November 2026, matched to the places Hermes lets you choose a model.
The examples are a web agency, a kitchen fitter and a recruitment firm, and every outreach job stays draft only, because Hermes has no built-in approval for outgoing messages.
Want my team to match models to your own sales workflow? Book a free SEO strategy session.
What GoldieBench Scores, And Where That Helps Sales
The GoldieBench methodology gives every model the same fixed prompt, one shot, and the result must be a single self-contained HTML file.
Julian scores each result from 0 to 10, and the leaderboard ranks models by their average across the tasks they were scored on.
That makes one task directly useful for sales: Landing, a one-shot modern marketing landing page.
A one-page pitch built for a single prospect is the same kind of build.
GoldieBench doesn't score email drafts, call briefs or CRM reviews, so for those jobs the averages are a rough guide to overall strength, nothing more.
Six Places Hermes Lets You Pick A Model
- The main model: set with
hermes model; anything unpinned uses it. - Each cron job: a per-job pin with
--modeland--provider, orcron.modelas a default for every unpinned job, per the scheduled tasks docs. - Each job's thinking level:
--reasoning-effort, fromnoneup toultra. - Each Bot: Bot Mode gives every Bot a model and provider pin, and Bots can run different models side by side.
- Each chat channel:
channel_overridesin the messaging gateway set a model and system prompt per channel. - One prompt at a time: the
/moacommand sends a single prompt through the default Mixture of Agents preset, then returns to your session model, per the slash command reference.
So a sales desk can run one model for briefs and another for pitch pages without either job knowing about the other.
The AI Profit Boardroom trainings walk through this per-job model setup inside the Agent OS.
The Shortlist: Overall Averages
From the GoldieBench models leaderboard, read on 7 November 2026:
| Model | Avg score | Tasks | Page read |
|---|---|---|---|
| Fusion | 8.59/10 | 47 | Fusion page |
| Claude Opus 5 | 8.27/10 | 50 | Opus 5 page |
| Hermes MoA | 8.17/10 | 47 | Hermes MoA page |
| GPT-5.6 Sol | 8.16/10 | 50 | GPT-5.6 Sol page |
| Claude Fable 5 | 8.10/10 | 47 | Leaderboard |
| Grok | 8.09/10 | 47 | Grok page |
| Claude Opus 5.5 | 7.57/10 | 50 | Opus 5.5 page |
The gap between first and seventh is about one point, so the task-level scores below say more than the averages.
Prospect Pitch Pages: The Landing Task Scores
The Landing task page shows 26 models tested, 21 scored, an 8.11/10 task average and Fugu Ultra as the winner.
| Model | Landing score | Page read |
|---|---|---|
| Fugu Ultra | 9.0/10 (winner) | Landing task page |
| Fusion | 9.0/10 | Landing task page |
| GLM-5.2 | 9.0/10 | Landing task page |
| Grok | 9.0/10 | Landing task page |
| GPT-5.6 Sol | 8.7/10 | Landing task page |
| Gemini 3.6 Flash | 8.6/10 | Landing task page |
| Hermes MoA | 8.4/10 | Hermes MoA page |
| Claude Opus 5 | 8.4/10 | Opus 5 page |
| Claude Opus 5.5 | 7.5/10 | Opus 5.5 page |
| Claude Fable 5 | 7.2/10 | Landing task page |
For a web agency that sends each serious prospect a short custom page, here's the job:
- A Bot called Pitch is pinned to one of the 9.0 models.
- A rep posts the prospect's name, site and the problem from the discovery call.
- Pitch builds a single HTML page that restates the problem and the proposed fix.
- The rep opens it, edits the copy, and decides whether it goes to the prospect.
Pitch never sends the link itself; the prompt says draft only, and the rep shares the page.
Follow-Up Drafts: Pick For Steadiness, Not The Landing Score
A kitchen fitter runs a morning job that drafts follow-ups for quotes that have gone quiet.
- The job reads the quote sheet with the bundled
google-workspaceskill. - It writes one draft per stale quote, with "draft only" in the prompt.
- Output is delivered to the owner's own chat, since the job's final answer goes to its delivery target.
GoldieBench has no score for this kind of writing.
The Hermes docs' own channel_overrides example puts a fast model in a daily channel and a frontier model in a specialist one, which is a sensible split for this job and the pitch work above.
Pin the follow-up job with --model so a change to your main model never changes the tone of the drafts overnight.
Call Briefs And High-Stakes Proposals: Where Hermes MoA Fits
The Hermes MoA page lists 8.17/10 across 47 tasks, with 3 gold and 10 silver medals.
Its stack notes describe a panel of Claude Opus 4.8 and GPT-5.5 whose drafts an Opus 4.8 aggregator merges into one answer.
The same page lists the trade-off: latency is the panel's slowest draft plus the aggregator pass, roughly 110 to 140 seconds per build.
That fits a recruitment firm preparing a once-a-quarter proposal to a large client, where one careful answer matters more than speed.
In Hermes, /moa sends just that one prompt through the default preset and then hands control back to your normal model.
Group Chats: Keep The Router Light
A sales room in Bot Mode might hold a Researcher, a Writer and the Pitch Bot.
When a message names nobody, a room can pick who answers automatically with one short tool-less model call, and the docs say to route that call to a cheap model through auxiliary.group_router.
Save the strong models for the Bots doing the work.
Keep Every Model On A Short Leash
A better model doesn't make outbound safer.
Per the Hermes security docs, approvals cover dangerous commands, and cron and webhook runs deny them by default whichever model runs the job.
For CRM tools, set the Attio or Close MCP server to trust: untrusted, so every write-capable call asks first, per the MCP config reference.
Messages themselves have no approval gate, so delivery targets stay with your team.
The Pairing Cheat Sheet
| Sales job | Where to pin | What the bench says |
|---|---|---|
| Prospect pitch page | Bot pin | Landing: Fugu Ultra, Fusion, GLM-5.2, Grok at 9.0 |
| High-stakes proposal | /moa | Hermes MoA 8.17 average, 8.4 on Landing |
| Morning follow-up drafts | Cron --model pin | Not measured; pick for steady tone |
| Call briefs in chat | channel_overrides | Not measured; use the averages as a rough guide |
| Group chat routing | auxiliary.group_router | Not measured; docs suggest a cheap model |
For the hosted alternative, our Grok Bot for sales page looks at the same jobs from Grok Bot's side.
The Hermes facts here come from the v0.23.0 release of 5 November 2026 and its docs.
The AI Profit Boardroom is where members share which model they pinned to which job.
Also On Our Network
- Agent Operatorsa seven-step operator build for a draft-only Hermes sales desk
- AI Income Deskfive owner-sized sales jobs and whether Hermes is worth setting up
- AI Tool Verdictten sales jobs scored out of 10, with a reason for each
- agentos.guidefive copy-paste prompts that keep every sales draft in your hands
- aiprofitboardroom.comthe complete Hermes Agent for Sales guide on the AI Profit Boardroom blog
Pin one model to your first sales job today, and book a free strategy session if you want help choosing models for Hermes Agent for sales.
FAQ
Does GoldieBench measure sales email writing?
No. It scores one-shot builds of single HTML files. The Landing task fits prospect pitch pages; for drafts and briefs the averages are only a rough guide.
Which models scored highest on the Landing task?
Fugu Ultra, Fusion, GLM-5.2 and Grok each scored 9.0 on the GoldieBench Landing task page read on 7 November 2026, with Fugu Ultra listed as the winner.
How do I keep a follow-up job on the same model?
Pin it with hermes cron create or edit using --model and --provider, or --pin to lock in the current main model.
When is /moa worth the wait for sales work?
For one important prompt, such as a large proposal. The Hermes MoA page lists roughly 110 to 140 seconds per build because a panel drafts and an aggregator merges.


