Hermes Agent Cron Job + Which AI Models To Pin On Each Schedule, With Real GoldieBench Scores For Opus 5, GPT-5.6 Sol, Kimi K3 And Gemini 3.6 Flash

A Hermes agent cron job runs on whichever model it resolves to at the moment it fires, so the model you pin decides how good every scheduled run is.

GoldieBench gives a useful signal here: across the six models I'd shortlist, averages run from 8.27 down to 7.01 out of 10.

Below are the scores I read on the live GoldieBench pages on 1 November 2026, what they mean for scheduled work, and which model I'd pin to which kind of job.

If you'd rather have my team make that call with you, book a free SEO strategy session.

How Hermes Picks The Model When A Job Fires

The Hermes cron docs set a fixed order.

1. The job's own pin, set with --pin or --provider and --model.
2. cron.model, a default for every unpinned job.
3. Your main chat model, as it stands at fire time.

So an unpinned job follows your chat model around.

Switch chat models for an afternoon, and the nightly job switches too.

That makes model choice a per-job decision, which is exactly where benchmark scores help.

The GoldieBench Scores For A Cron Shortlist

Ranks are from the GoldieBench leaderboard, which covers 33 models, 50 tasks and 1,310 live demos.

Fusion tops the board at 8.59, but it's an OpenRouter entry, so I've left it off a list of models you'd pin to a single job.

What These Scores Do And Don't Tell You

GoldieBench scores are Julian's own 0 to 10 verdicts, averaged across every task a model ran.

Every model gets the same fixed prompts: one shot, one HTML file out.

That makes the scores a strong read on first-try quality, the thing that matters most when nobody is watching the run.

It doesn't measure a morning brief or a log check directly.

Treat the gap between 8.27 and 7.01 as a quality signal, not a promise about your exact job.

Heavy Weekly Jobs: Pin The Top Scorer At High Effort

A weekly strategy review runs once, reads a lot and has to get it right first time.

That's where I'd pin the highest single-model score on the list: Claude Opus 5 at 8.27 across 50 tasks, with 13 gold medals on its page.

GPT-5.6 Sol is close behind at 8.16, and its page says it was benched at medium reasoning effort.

Pair either with a per-job effort pin.

hermes cron edit weekly-review --provider <provider> --model <model> --reasoning-effort high

The docs say an effort level a model doesn't support is clamped, so pinning xhigh on a model that caps at high just runs at high.

Long Unattended Runs: Kimi K3

Some scheduled jobs run long, like an overnight audit across many pages.

Kimi K3 scores 7.89 across 50 tasks with 8 gold medals.

Its GoldieBench page says it's wired into the Agent OS as the kimi-k3 Hermes profile and used for long unattended agent runs.

The same page lists its trade-off plainly: it's slow on hard tasks.

For a job that runs while you sleep, slow is fine.

Just check the doctor report for late dispatches if it starts bumping into its next slot.

Frequent Light Checks: Gemini 3.6 Flash, Or No Model At All

An hourly "anything new?" job doesn't need the top of the board.

Gemini 3.6 Flash scores 7.08 across 50 tasks, and its page describes it as faster and using fewer tokens.

Pin it with a low effort level and the hourly job stays light.

Better still, ask whether the job needs a model at all.

No-agent mode runs a script and delivers its output with zero model calls, and a pre-check script can skip the agent on any tick where nothing changed.

The best benchmark score for a watchdog is no score, because no model ran.

Where Hermes MoA Fits

Hermes MoA scores 8.17 across 47 tasks and ranks third.

Its page describes a panel of frontier models, merged by an aggregator, run from the Mixture tab in the Agent OS.

That's a system, not a single model you pin with --model, and the cron docs don't describe pinning a mixture.

So I'd use MoA when you're running a big piece of work by hand, and pin a single model for the scheduled version.

Re-Check The Pick After Every Model Change

New models land on GoldieBench often, and the leaderboard order moves with them.

When a new entry overtakes your pinned model, test it before you switch the job.

Create a paused copy of the job on the new model, run it once by hand with hermes cron run, and compare the two outputs side by side.

Only then edit the live job's pin.

Want help reading the results? Bring both outputs to a coaching call in the AI Profit Boardroom, where four calls run every week.

Keep the old pin noted somewhere, so you can roll back with one edit if the new model's output is worse.

Pinned Jobs Don't Fall Back

This matters as much as the score.

The provider recovery section says a pinned job never walks your fallback_providers chain.

If the pinned route fails, the run fails and the alert says so.

An unpinned job does walk the chain.

So for a brief that must arrive, leave it unpinned and set cron.model; for a report that must stay on one model, pin it and accept the occasional failed run.

A Model Plan For Five Common Jobs

Whatever you pin, set a missed-run rule on each job too.

The docs give a catch-up window of half the job's period, between 2 minutes and 2 hours, and --grace-minutes 0 skips missed slots entirely.

If a run with your chosen model goes wrong, hermes cron cancel or the Desktop Stop button ends it without touching the schedule.

Also On Our Network

• Agent Operatorsthe seven-step operator setup, from gateway install to stopping a stuck run
• AI Income Deskthree small-business jobs to schedule, and the two safety rules to set first
• AI Tool Verdictan eight-row scorecard of Hermes scheduling and the final 8.3 verdict
• agentos.guidefive prompts that make your agent schedule, stop and repair its own jobs
• aiprofitboardroom.comthe complete Hermes Agent Cron Job guide on the AI Profit Boardroom blog

Pin one model per job from the table above, and if you want my team to check the plan against your real work, book a free strategy session before you schedule your next Hermes agent cron job.