GoldieBench Blog · 7 min read
Hermes vs Perplexity Computer: Who Picks The Model In Each, Which AI Models To Run In Hermes, And What The GoldieBench Scores Mean
In Hermes you pick the model, so the bench helps directly. In Perplexity Computer the effort setting picks it. Here is what the scores mean on each side.

Hermes vs Perplexity Computer splits the model question in two: in Hermes you choose the model yourself, so GoldieBench scores can steer the pick directly, while Perplexity Computer chooses models for you behind an effort setting.
That makes the bench far more useful on one side than the other.
Every score on this page is one I read on a live goldiebench.com page on 6 November 2026.
GoldieBench doesn't score Hermes or Perplexity Computer as products; it scores the models inside them.
For help pairing models with your own setup, get your Backlink Gameplan, or try Polar Rank free for 3 days.
Who picks the model in each tool
Hermes runs on whatever provider you set with hermes model, and the Hermes providers page lists Nous Portal, Anthropic, OpenAI, OpenRouter, xAI, Google Gemini, Kimi, MiniMax, DeepSeek, local Ollama, LM Studio and custom endpoints.
You can switch per turn with /model, and cron jobs can pin their own model.
Perplexity Computer works the other way round: on web, iOS and Android you pick Light, Standard, High or Ultra, and the effort mode page says Computer then selects the model and reasoning level for you.
A Custom mode remains for choosing a specific model, and the Mac and Windows apps show a regular model picker instead.
Perplexity also says the model behind each effort level changes as models are added or retired.
What a GoldieBench score actually measures
The methodology page gives every model the same fixed prompt, once, with no help, no iteration and no best-of-N.
The output has to be a single self-contained HTML file that runs in a browser.
Julian scores each build from 0 to 10, and the leaderboard ranks models by their average across scored tasks.
So the bench rewards building something complete in one pass, which maps well onto long agent jobs and less well onto short research answers.
The scores, read on goldiebench.com
| Model | Avg score | Tasks | Hermes provider for that vendor | Page read |
|---|---|---|---|---|
| Fusion | 8.59/10 | 47 | OpenRouter (GoldieBench lists it as an OpenRouter product) | /models/fusion |
| Claude Opus 5 | 8.27/10 | 50 | Anthropic or OpenRouter | /models/opus5 |
| Hermes MoA | 8.17/10 | 47 | A Hermes mixture-of-agents panel | /models/moa |
| GPT-5.6 Sol | 8.16/10 | 50 | OpenAI or OpenRouter | /models/gpt56 |
| Claude Fable 5 | 8.10/10 | 47 | Anthropic or OpenRouter | /models/fable-5 |
| Grok | 8.09/10 | 47 | xAI provider | /models/grok |
| Kimi K3 | 7.89/10 | 50 | Kimi / Moonshot provider | /models/kimik3 |
| Gemini 3.6 Flash | 7.08/10 | 50 | Google Gemini provider | /models/gemini36 |
The provider column comes from the Hermes providers page, except Fusion's, which comes from its GoldieBench page.
Model lists change, so confirm each one shows up in hermes model before you rely on it.
Picking a Hermes model by job
- Long builds and multi-step jobs: Claude Opus 5 has the highest single-model average in the table at 8.27/10 over 50 tasks.
- A second strong option on a different vendor: GPT-5.6 Sol sits at 8.16/10 over 50 tasks, which makes it a sensible fallback provider.
- Long context: Kimi K3's page describes a 1M context tuned for long-horizon agent work, at 7.89/10 over 50 tasks.
- Quick, frequent cron jobs: Gemini 3.6 Flash scores 7.08/10 over 50 tasks, so test it on simple summaries before trusting it with builds.
These are starting points from the bench, not proof that a model will handle your inbox or your files well.
Pinning models for scheduled work
Scheduled jobs are where the two tools differ most on models.
In Hermes, the cron docs resolve a job's model in order: a per-job pin first, then a fleet default set with cron.model, then your main chat model.
That lets you keep a top-scoring model for chat and builds while a lighter model runs the hourly checks.
Each job can also pin its own reasoning effort, from none up to ultra.
In Perplexity Computer, the enterprise help page says an automation carries its own model selection, and falls back to Computer's default model if that model is no longer available to your organization.
So on both sides, check what a scheduled job is really running on after any model change.
Hermes MoA: a panel instead of one model
Hermes MoA scores 8.17/10 over 47 tasks.
Its GoldieBench page describes a panel of frontier models merged by a chair model, built inside Julian's Agent OS rather than by a vendor.
That puts it above GPT-5.6 Sol and just behind Claude Opus 5 in the table above.
Perplexity Computer orchestrates several models too, but it does that for you, and GoldieBench has no score for Computer's own mix.
Local models on both sides
Both tools can keep inference on your own hardware, in different ways.
Hermes connects to local servers such as Ollama, LM Studio, vLLM and llama.cpp.
GoldieBench's local leaderboard puts Qwable 5 27B Coder first at 7.14/10 over 41 tasks, run through Apple MLX.
Perplexity's Personal Computer page lists its own local models by platform, such as PPLX 27B on supported RTX PCs, and Portable Computer runs the agent itself on NVIDIA hardware.
None of Perplexity's local models appear on GoldieBench, so I can't give you a bench score for them.
Using Perplexity's models inside Hermes
There is one documented link between the two sides.
The Hermes providers cookbook adds Perplexity's API as a custom provider, with sonar as the example default, for answers that search the live web and cite sources.
Sonar has no GoldieBench score either, and it isn't Perplexity Computer, so treat it as a research add-on rather than your main Hermes model.
A sensible pattern is a high-scoring model as default and a per-turn switch to sonar for questions that need fresh sources.
A small-business example: an estate agency's listing pages
A local estate agency wants each new listing written up as a clean one-page HTML sheet.
That task looks a lot like what GoldieBench measures, so the scores are directly useful here.
Step one, set Hermes to Claude Opus 5 or GPT-5.6 Sol and give it the property notes from the office folder.
Step two, Hermes builds the single-file page and posts a preview link to the agent on Telegram for review.
Step three, for the area guide, the agent asks Perplexity Computer for cited local information on schools and transport, then pastes the approved facts into the notes.
The bench informs the Hermes model choice, and Computer handles the sourcing it's built for.
Reading these scores honestly
GoldieBench tests one-shot creative coding builds, not email triage, research accuracy or tool-calling reliability.
Use the table to shortlist a Hermes model, then test it on one real job before you move your cron jobs over.
For Perplexity Computer, start on Light or Standard as its own docs suggest, and step up only when the output falls short.
I take the same approach in my Hermes vs n8n model guide.
The verdict on models
If model choice matters to you, Hermes wins, because every model in the table above is one you can actually select.
If you'd rather never think about models, Perplexity Computer wins, because the effort setting makes that decision for you.
Hermes itself is MIT-licensed on GitHub, and the v0.23.0 release added spend limits for paid tools, which helps once you run premium models all day.
The full model notes behind these picks are inside the AI Profit Boardroom.
Shortlist a model from the table, test it on one job, and get your Backlink Gameplan if you want help with hermes vs perplexity computer, or try Polar Rank free for 3 days.
FAQ
Can I choose the model in Perplexity Computer?
By default you choose an effort level and Computer selects the model. Its effort mode page says a Custom mode remains for picking a specific model, and the desktop apps show a regular model picker.
Which GoldieBench model should I start with in Hermes?
Claude Opus 5 has the top single-model average in the table at 8.27/10 over 50 tasks on its GoldieBench page. Test it on one real job first.
Does GoldieBench score Perplexity Computer?
No. It scores individual models and panels such as Hermes MoA, not Perplexity Computer's own model mix or its local models.
Can Hermes use Perplexity models?
Yes, through Perplexity's API as a custom provider, with sonar as the documented example. That is a separate product from Perplexity Computer.
- 📺 Video notes + links to the tools
- 🆓 Get a FREE AI course + community + 1,000 AI agents
- 🎥 Learn how I make these videos


