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

GoldieBench Blog · 6 min read

Grok Bot vs Perplexity Computer: Which AI Models Sit Behind Each, With Real GoldieBench Scores

One agent keeps its model hidden. The other names it and lets you swap it. Here is where the leaderboard helps, and where it can't.

A dark navy benchmark leaderboard with gold bars, a sealed cloud computer with several screens on one side and a dial feeding interchangeable model chips into a row of sandbox boxes on the other

Grok Bot vs Perplexity Computer is also a question about models: Grok Bot picks its model for you and never shows it, while Perplexity Computer orchestrates many models and lets you see or choose the one behind a task.

That means GoldieBench scores can't guide one side at all, and can guide the other when you choose a model yourself.

Every score below I read on live GoldieBench pages on 6 November 2026.

The product comparison is my reading of the xAI and Perplexity docs, since neither vendor compares the two.

For help matching models to your own workflows, get your Backlink Gameplan, or try Polar Rank free for 3 days.

The model question, side by side

Model questionGrok BotPerplexity Computer
Who picks the model?Grok Bot, every timeComputer by default, or you
Can you see the model?NoYes, the effort selector shows its current model
Ways to steer itNone documentedLight, Standard, High or Ultra effort, Custom mode, a picker in the desktop apps
Several models on one questionNot documentedModel Council, two to eight models
On-device modelsNot offeredHybrid compute on Mac, Portable Computer on supported NVIDIA hardware
Does GoldieBench help?NoYes, when you choose a benched model for a build

Why GoldieBench can't rank Grok Bot

xAI's Grok Bot security page says the product manages model selection and has no customer-facing model picker.

With no fixed model, there is nothing for a one-shot leaderboard to score.

The Grok page on GoldieBench covers xAI's Grok model, not the Grok Bot agent, and nothing in xAI's docs links the two.

Judge Grok Bot on its workflow instead: the shared cloud computer, where all your Bots work and each gets its own screen, plus its routines and approval cards.

How Perplexity Computer chooses models

Perplexity's launch post describes Computer as model-agnostic, with a core reasoning model that hands subtasks to other models suited to them.

It also says you can choose specific models for specific subtasks, and that the line-up changes as models advance.

The effort mode page adds four settings, Light, Standard, High and Ultra, on the web and in the iOS and Android apps.

Each setting picks the coordinating model and how deeply it reasons, and the selector names that model, though Perplexity warns the pairing will change.

Custom mode lets you pick the model and reasoning level yourself, and the Mac and Windows apps keep the regular model picker.

Model Council, added in August, runs two to eight models on one question across OpenAI, Anthropic, Gemini and open-source families, then shows where they agree.

What a GoldieBench score measures

The methodology page says every model gets the same fixed prompt, one shot, with no iteration and no best-of-N.

Each result must be one self-contained HTML file that runs in a browser, such as a game, a simulation or a page.

The ranking is the average 0 to 10 score across the tasks a model was scored on.

So the scores speak to build-style jobs, which matches Perplexity's own list of things Computer creates: apps, websites and reports.

They say little about research quality or citation accuracy.

The scores I read on 6 November

All of these come from the GoldieBench models page.

ModelMakerAverageTasks scored
FusionOpenRouter8.5947
Claude Opus 5Anthropic8.2750
Hermes MoAHermes8.1747
GPT-5.6 SolOpenAI8.1650
Claude Fable 5Anthropic8.1047
GrokxAI8.0947
Claude Opus 5.5Anthropic7.5750
Grok 4.7xAI7.1520
Gemini 3.6 FlashGoogle7.0850

One oddity to know: the methodology page still lists Grok as unranked, while the models page shows it at 8.09 over 47 tasks, so I treat the models page as current.

What the Grok score does and doesn't tell you

Grok scored 8.09 over 47 tasks, which puts xAI's model close to the top group on one-shot builds.

The Grok 4.7 page sits lower at 7.15, but over only 20 tasks, so the two averages aren't directly comparable.

Neither number tells you how Grok Bot performs, because xAI doesn't say which model it routes a task to.

What you can judge on Grok Bot is the work around the model: one persistent computer for all your Bots, saved sign-ins, routines on schedules or events, and approval cards before consequential actions.

Using the scores inside Perplexity Computer

When Perplexity's effort selector or picker names a model that also sits on GoldieBench, the table gives you a starting point for build jobs.

Perplexity doesn't promise that any one of these models is in your list, so check the selector first.

For research or briefings, leave the choice to the effort setting, since GoldieBench doesn't test research.

Perplexity's own guidance is that Light suits routine work with clear instructions and Ultra suits open-ended problems with several approaches to test.

JobWhere to start in PerplexityDoes GoldieBench help?
Build a page, app or interactive demoCustom mode with a benched modelYes
Sort routine data with clear rulesLight effortNo
Write a brief from mixed sourcesStandard effortNo
Weigh a high-stakes decisionModel Council or Ultra effortPartly, for the models you include

Perplexity also notes that an effort setting doesn't cap the steps, sources or tools a task uses.

On-device models: Perplexity's list vs the local leaderboard

Perplexity's Personal Computer page lists Gemma 4 E4B, Qwen3.6 35B-A3B and a Perplexity model on Apple silicon Macs, and PPLX 27B and Qwen 27B on supported PCs and DGX Spark.

None of those appear by those names on the GoldieBench local models board.

The closest relative is Qwable 5 27B Coder, a fine-tune built on a Qwen3.6-27B base, at 7.14 over 41 tasks, but it isn't the model Perplexity ships.

So for local work, judge Perplexity on privacy, since its on-device modes keep files local and ask before sending a step to the cloud.

Small-business examples

A property manager.

Use Grok Bot, where the model doesn't matter: a Bot named Lettings reads repair emails sent to its own address each morning, logs them in a Google Sheet and drafts replies for approval.

A print shop.

Use Perplexity Computer in Custom mode: pick a high-scoring benched model if it is listed, and ask it to build a one-page order form site from the shop's price list file, then review it before publishing.

A wedding photographer.

Use Perplexity Computer on the Standard effort setting: it researches the season's venues and writes a cited brief that arrives by email.

For the same model question on another pair, read Grok Bot vs Manus on models.

Verdict by job type

Pick Grok Bot when you'd rather not think about models and want standing Bots with memory, logins and routines.

Pick Perplexity Computer when model choice matters, and use GoldieBench scores to choose for build-style jobs.

Want the stack these tests run in? Join the AI Profit Boardroom.

Check the leaderboard before your next build, and join the AI Profit Boardroom or get your Backlink Gameplan to pick the right models for grok bot vs perplexity computer, or try Polar Rank free for 3 days.

FAQ

Which model does Grok Bot use?

xAI doesn't say. Grok Bot manages model selection itself and has no customer-facing model picker.

Can I pick the model in Perplexity Computer?

Yes. Effort mode picks a model for you and shows its name, Custom mode lets you choose the model and reasoning level, and the Mac and Windows apps keep a regular model picker.

Is the GoldieBench Grok score a Grok Bot score?

No. The Grok page on GoldieBench scores xAI's Grok model on one-shot HTML builds. Grok Bot is an agent product, and xAI doesn't name its model.

Are Perplexity's local models on GoldieBench?

Not by the names Perplexity lists. The local board's closest relative is Qwable 5 27B Coder, a fine-tune on a Qwen3.6-27B base, which scored 7.14 over 41 tasks.

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