Hermes Agent For Lead Generation: Which AI Models To Pair With Each Lead Job, With Real GoldieBench Scores

Hermes Agent for lead generation lets you pick a different model for each job, so the useful question isn't which model is best but which one fits finding, reading and scoring prospects.

GoldieBench can help with part of that answer, as long as you read the scores for what they are.

Every number below is one I read on a live goldiebench.com page, with the page linked.

Want my team to map models to your own lead flow? Book a free SEO strategy session.

What GoldieBench Measures, And What It Doesn't

The methodology page is clear: each task is one fixed prompt, one shot, and the result must be a single self-contained HTML file.

Scores are Julian's 0-10 judgement of those builds.

There's no lead-research task on the bench, no web-search test and no scoring-accuracy test.

So an average is a broad signal of how well a model reasons and follows a brief, not proof it will qualify a clinic or a café correctly.

Treat the bench as a shortlist, then test the shortlist on ten of your own prospects.

Where Hermes Lets You Choose The Model

Your main chat model is set with hermes model, and that's what research chats use.

The cron docs add two more levers: a per-job pin with --model and --provider, and a fleet default with cron.model.

Resolution order is per-job pin first, then cron.model, then the main model at the moment the job fires.

Each job can also pin its own reasoning effort, from none up to ultra, so a heavy weekly research run and a light daily scorer don't have to share one setting.

That means a lead desk can run a strong model for weekly research and a lighter one for the daily shortlist.

The Shortlist: Overall Averages From The Models Page

Those averages come from the models page as it read on 9 November 2026.

The gap between the top five is small: 8.27 down to 8.09.

For lead work, that gap matters less than context length, speed and where your data goes.

Hermes MoA, listed as a Mixture of Agents that merges a panel of models through a chair, sits second at 8.17, which is worth a look for the one-off job of writing your scoring rubric.

Long Context For Big Research Batches

Research batches pile up text fast: web_extract returns up to 15,000 characters a page by default and takes five URLs a call, per the tools reference.

The Claude Opus 5 page lists a 1,000,000-token context.

The Grok page lists 256,000 tokens.

Kimi K3's card on the models page describes a 1M context tuned for long-horizon agent work, and it averages 7.89.

For a batch of forty supplier websites in one session, the bigger window gives the agent more room before it has to drop earlier pages.

Private Lists: The Local Leaderboard

Prospect lists hold names and work emails, which count as personal data in many places.

Hermes can run open models entirely on your own machine, per the local models guide, with no network access once a model is downloaded.

GoldieBench keeps a separate local leaderboard for strictly on-device models:

The drop from 7.14 to 4.83 is steep, so on a local setup the first choice is clear on this bench.

Agents-A1 is described there as an agent-tuned open model built for long-horizon tool work, which suits tool-heavy research, but its build score is far lower.

Run your ten-prospect test before trusting any local model with a full week of scoring.

Opt-In Pages: The Landing Task Is The Closest Match

Some lead desks also need a page that collects enquiries, and that's the one place the bench maps almost directly.

The Landing task page asks for a modern marketing landing page in one shot; it shows a task average of 8.11 with 21 of 26 attempts scored.

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

Claude Opus 5's own page lists 8.4 on Landing Page.

Keep the form wired to a Hermes webhook route that checks HMAC signatures, so each enquiry is scored as it lands.

X Buying Signals Don't Depend On Your Chat Model

Hermes's x_search is read-only public X search that runs through xAI's built-in tool, per the tools reference.

It needs xAI credentials, but your chat model can be anything.

Grok's model card calls it the X-native model, which is a reason to try it for summarising signals, not a requirement for searching them.

Keep Every Model On A Short Leash

Whichever model you pick, the guardrails are the same, and the AI Profit Boardroom trainings cover them on video.

Hermes approvals cover dangerous commands and writes to MCP servers marked untrusted, per the security page; there's no built-in hold on outgoing email.

So prompts say draft only, and deliveries go to your own chat.

Respect each site's terms and the anti-spam and privacy rules where you and your prospects are, whatever the model scores.

Qualified leads then move to a sales setup; the model picks for that side are in Hermes Agent for sales and which models to pin.

A Worked Pairing: An Office Furniture Supplier

Take a small office furniture supplier hunting for companies moving into new premises.

The owner writes the scoring skill once in a chat on a top-average model, because the rubric is the piece every later run depends on.

The weekly research job, which reads dozens of commercial property and company pages, gets a long-context model pinned with --model.

The daily cron that rescores new rows gets a lighter pinned model and a low reasoning-effort setting.

Because client names sit in the sheet, the owner trials the top local model on the same ten rows before deciding whether privacy is worth the drop in score.

The enquiry page on the supplier's site comes from a model that did well on the Landing task, and the form feeds a webhook route.

The Pairing Cheat Sheet

If you want to see these pairings set up on video, the AI Profit Boardroom walks through each one.

Also On Our Network

• Agent Operatorsthe eight-step prospect desk build, from ICP skill to morning cron list
• AI Income Deskfour small-business lead-gen use cases and who should hold off
• AI Tool Verdicteleven top-of-funnel lead jobs scored out of ten, with reasons
• agentos.guidefive copy-paste Agent OS prompts for finding and scoring prospects
• aiprofitboardroom.comthe complete Hermes Agent for Lead Generation guide on the AI Profit Boardroom blog

Shortlist two models from the table, test them on ten real prospects, and book a free SEO strategy session if you want help pinning the right one for Hermes Agent for lead generation.