⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 6 min read

Hermes Agent Microsoft Office + Which AI Models To Run For Word, PowerPoint And Excel Jobs, With Real GoldieBench Scores

The skill does the file work with Python helpers; the model writes the specs and fixes the misses. Here are eight GoldieBench scores and how to use them.

Illustrated hero of a benchmark leaderboard feeding into document, slide and spreadsheet cards on a dark navy background with gold accents

The model behind hermes agent microsoft office jobs decides how good the JSON specs are, how well the agent reads helper results, and how cleanly it fixes a slide that overflows.

The skill itself is fixed: the Hermes docs describe Python helpers that build and check Word, PowerPoint and Excel files offline.

What you can change is the model driving them, so below are the GoldieBench scores I read on the live models page and model pages on 14 November 2026, and how to read them for document work.

If you would rather my team matched a model to your document workload, book a free SEO strategy session.

What The Model Actually Does In An Office Job

The heavy lifting is done by scripts, not by the model.

Helpers such as docx_create.py, pptx_create.py and xlsx_create.py turn a JSON spec into a real file using python-docx, python-pptx and openpyxl.

The model has four jobs around them.

It writes the spec, picks the right helper and flags, reads the JSON each helper prints back, and decides what to fix when a check fails.

For decks there is a fifth: after pptx_render.py makes slide images, the docs have the agent check each one with vision_analyze.

So you are choosing a model for structured output, instruction following and self-correction, not for writing prose alone.

The GoldieBench Scores I Used

ModelAvg scoreScored tasksGold medalsContext (model page)
Fusion8.59/104721Varies by panel
Claude Opus 58.27/1050131,000,000 tokens
Hermes MoA8.17/10473Varies by panel
GPT-5.6 Sol8.16/105021,050,000 tokens
MiniMax M37.97/10472see model page
Kimi K37.89/105081,048,576 tokens
GLM-5.27.77/10475see model page
Gemini 3.6 Flash7.08/105011,000,000 tokens

Averages, task counts and gold medals come from the models page; the context figures come from each linked model page.

The models page compares 33 frontier models in total.

What A GoldieBench Score Means For Document Work

Be clear about what is measured.

The tasks page lists 50 one-shot creative coding tasks across five categories, each a fixed prompt that has to come back as a single HTML file.

The methodology page says every score is my own 0 to 10 number, one shot, with no retries and no best-of-N.

There is no Word, PowerPoint or Excel task on the bench.

So read the scores as a signal of how reliably a model follows a fixed brief and ships a working artefact first time, which is close to what writing a correct JSON spec asks of it.

The tasks page says Page tasks test product instinct, which is the nearest category to building a deck or report people will read.

Averages Versus Medals

The table shows two different signals, and they answer different questions.

Per the methodology page, the leaderboard rank is the average 0 to 10 score across every task a model was scored on.

Medals are per task: the highest score on a task takes gold, the next silver, the next bronze, and ties at the top share gold.

A high average with few golds, like GPT-5.6 Sol, can point to a steady model that rarely fails badly.

Many golds with a lower average, like Kimi K3, can point to a model that wins some briefs outright but varies more.

For batch document jobs, steadiness usually matters more than peak wins, because one broken file in fifty still has to be fixed by hand.

Picking A Model For Specs And Decks

For the profile that builds new documents from scratch, start near the top of the single-model table.

Claude Opus 5 averages 8.27 across 50 scored tasks with 13 golds, and GPT-5.6 Sol averages 8.16 across 50.

Both list context windows of about a million tokens on their model pages, which leaves room for a long brief, a brand template outline and the reference notes the skill loads.

A deck job is where a strong model pays off most, because the agent has to plan the slides, write the spec, read the render check and fix what looks wrong.

Picking A Model For Repeat Template Fills

Filling the same Word template for fifty clients is a different job.

The structure already exists, docx_template.py does the filling, and the model mostly passes data and reads results.

Here a mid-table model can be enough.

Kimi K3 averages 7.89 across 50 tasks with 8 golds, and its model page describes it as tuned for long-horizon agent work, which suits a long batch.

Gemini 3.6 Flash sits lower at 7.08 across 50, and the bench tags it as Google's launch-day Flash; test it on a small batch before you hand it the whole client list.

When A Panel Model Makes Sense

Two of the top three entries are panels rather than single models.

Fusion leads the page at 8.59 across 47 tasks and is listed as an OpenRouter multi-model panel.

Hermes MoA averages 8.17 across 47; its page describes a panel of frontier models whose drafts are merged by a chair, run from the Mixture tab in the Agent OS.

For a document that matters, such as a contract summary or a board deck, a panel is worth trying, because a second model reading the same spec can catch what the first missed.

For bulk work it is usually more than you need.

Where The Model Cannot Help

Some limits sit in the skill, and no model fixes them.

It does not open legacy .doc, .ppt or .xls files, so those need converting with LibreOffice first.

It has no pure-Python path to PDF, so PDF export needs soffice.

And nothing in the skill computes formulas or Word fields; Excel fills values on open, or xlsx_recalc.py does with LibreOffice installed.

A top-scoring model will tell you these steps are missing, but it cannot skip them.

Testing Your Pick On Real Files

Run the same three jobs through two models before you commit.

Use one Word letter from a template, one six-slide deck and one workbook built from a CSV.

Count how many helper calls each model needs before every check passes, and how often it reads a failing JSON result correctly.

Keep the model that gets there with the fewest retries, then pin it to the profile that does document work.

Since the v0.23.0 release merged the three format skills into one microsoft-office skill, one model choice now covers all three formats.

For other skills worth testing the same way, see the best Hermes Agent skills.

My Verdict On Models For Office Files

Put a top single model such as Opus 5 or GPT-5.6 Sol on the profile that builds new decks and reports.

Test a mid-table model on repeat template fills, where the helpers do most of the work.

Try a panel for the few documents where a second read is worth it.

And remember the bench measures one-shot HTML builds, so your own three-file test is the final word for your work.

If you want to see these models wired into a working setup, the AI Profit Boardroom has the trainings.

Also On Our Network

Pin a top model to your deck builder, test a mid-table model on template fills, and if you want help, book a free SEO strategy session or join the AI Profit Boardroom to tune your hermes agent microsoft office setup.

FAQ

Which AI model is best for Hermes Agent Microsoft Office jobs?

For new decks and reports, start with a top single model. On GoldieBench, Claude Opus 5 averages 8.27 and GPT-5.6 Sol 8.16, read on 14 November 2026.

Does GoldieBench test Word, Excel or PowerPoint tasks?

No. It scores 50 one-shot creative coding tasks that return a single HTML file. Use the scores as a signal of following a fixed brief, then test on your own files.

Can a cheaper or smaller model fill Word templates?

Often, yes. docx_template.py does the filling, so the model mainly passes data and reads results. Test a mid-table model on a small batch first.

Does the model matter if the skill uses Python helpers?

Yes. The model writes the JSON spec, picks the helper, reads the JSON result and decides what to fix, so weak instruction following shows up as more retries.

Can any model make the skill open old .xls files?

No. Legacy .doc, .ppt and .xls files must be converted with LibreOffice first, whatever model you run.

Is a panel model like Hermes MoA worth it for documents?

For a few high-stakes files, it can be. Hermes MoA averages 8.17 across 47 tasks, and a second model reading the same spec can catch mistakes.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week