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

GoldieBench Blog · 6 min read

Grok Bot vs Gemini Spark: Which AI Models Each One Runs, With Real GoldieBench Scores

One agent keeps its model to itself. The other lets you choose per task. Here is how to use the GoldieBench scorecards on the side where you hold the dial.

Dark navy benchmark leaderboard with gold bars, a sealed cloud computer on one side and an open model selector dial feeding a task panel on the other

On models, Grok Bot vs Gemini Spark is lopsided: Grok Bot picks its model for you and never shows it, while Gemini Spark, now called Tasks in Gemini, puts a model and thinking-level menu at the top of every task.

So GoldieBench can say nothing about one side and something useful about the other.

Below are the scores I read on live GoldieBench pages on 6 November 2026, what they measure, and how to use them when you set up a Gemini task.

The tool comparison is my reading of the xAI and Google docs, since neither company compares the two.

If you'd like my team to match models to your workflows, get your Backlink Gameplan, or try Polar Rank free for 3 days.

The model question, side by side

Model questionGrok BotTasks in Gemini
Can you pick the model?No customer-facing pickerYes, from the model menu, plus a thinking level
Does the vendor name the model?NoGoogle's posts named Flash models at different points
Scored on GoldieBench?Not as a productGemini 3.6 Flash has a full scorecard
Bring your own keyNot documentedNot documented
Where the agent runsA shared cloud computer in Cursor's cloudYour Chrome or Google's remote browser

Why GoldieBench can't score Grok Bot

The Grok Bot security page says the product manages model selection and offers no model picker to customers.

There's no single model to benchmark, so no GoldieBench number applies to Grok Bot as a product.

That isn't a knock on it; it's a design choice that moves the decision from you to xAI.

What you control instead is the brief: a clear job per Bot, a skill saved from a run that worked, and approval rules on anything risky.

Judge it on its workflow: the shared cloud computer where every Bot gets its own screen, its routines and its approval cards.

What Gemini tasks run on

Google's launch post said Spark ran on Gemini 3.5 with the Antigravity harness, and the Gemini 3.5 post said it used 3.5 Flash.

Google's India rollout post in July described it as driven by Gemini 3.6 Flash.

Today's tasks help page adds a model menu, so you choose the model and thinking level per task.

Because the default has moved with each Flash release, I won't name one model as "the" Tasks in Gemini model.

What a GoldieBench score measures

The methodology page gives every model one fixed prompt, one shot, with no retries.

Each answer must be a single self-contained HTML file that runs in a browser: games, visuals, simulations and pages, each scored 0 to 10.

That's a test of building, not of running errands across Gmail and Calendar, which is what Tasks in Gemini mostly does.

Use the scores when a task asks Gemini to make something, like a web page or a code file, and ignore them for inbox triage.

Gemini 3.6 Flash on GoldieBench

On its model page, Gemini 3.6 Flash shows a 7.08/10 average across 50 tasks, with medals of one gold, two silver and one bronze.

Its best scorecard entries were Matrix (Visual) at 8.7, then Voxelcraft, Landing Page, Fractal, Galaxy and Plasma at 8.6.

Its weakest were Pathtracer and Particleforge at 3.5 and Waves at 3.0.

The same page labels it a Flash-tier model up against flagships, and suggests it for high-volume work and fast prototypes you iterate on.

The Grok model scores, and why they aren't Grok Bot

The Grok model page shows an 8.09/10 average across 47 tasks, with five golds, including Twilightvale at 9.5 and Landing Page at 9.0.

The same site also calls Grok "currently unranked but on the bench" on its models page, so treat that average as provisional.

More important, that page covers xAI's Grok model, not the Grok Bot product, and xAI doesn't say which model Grok Bot runs.

Don't read the Grok score as a Grok Bot score.

How to set the model in a Gemini task

  1. Open gemini.google.com or the Gemini app on Mac and click Task below the text box.
  2. Click the model menu at the top and choose the model and thinking level for this task.
  3. Describe the output you want, for example a single HTML file or a spreadsheet with formulas.
  4. Review the file in the thread; Markdown and text open in Canvas, and other files come as links you can download.

Google lists the file types tasks can produce, from PDF and DOCX to HTML, JS and Python, on the same help page.

How the leaderboard frames the choice

The GoldieBench menu I read listed Fusion at 8.59, Claude Opus 5 at 8.27, Hermes MoA at 8.17, GPT-5.6 Sol at 8.16 and Claude Fable 5 at 8.10.

None of those can be selected inside Tasks in Gemini, which offers Google's own models.

So the useful comparison for a Gemini task is between Google models in its menu, and GoldieBench currently scores one of them, Gemini 3.6 Flash.

Small-business pairings, step by step

A bakery that needs a quick landing page

  1. In Gemini, click Task and describe the page: opening hours, menu sections and a booking link.
  2. Pick the model from the menu; if Gemini 3.6 Flash is listed, its 8.6 on GoldieBench's Landing Page task is a fair signal for this kind of build.
  3. Ask for a single .html file, review it in the thread and download it.

An estate agent's weekly listings digest

  1. In Grok Bot, create a Bot for listings research and ask for a weekday routine that browses the listing sites.
  2. Have it post a linked summary in its chat and never email clients without approval.
  3. Model choice isn't yours here, so judge the routine on accuracy and sources, not benchmark scores.

For the same model question against OpenAI's agent, see Grok Bot vs ChatGPT Agent on GoldieBench.

When a benchmark score shouldn't drive the choice

Most small-business agent work is errands: reading mail, checking calendars, filling forms and drafting replies.

GoldieBench doesn't measure any of that, so for those jobs compare the agents on approvals, triggers and accounts instead.

Tasks in Gemini always confirms before it sends messages, deletes data, buys anything or submits a form.

Grok Bot keeps emails written by a routine as drafts until you press Send, and lets you add Ask first rules, per the approvals page.

Those guardrails matter more for a salon's bookings inbox than a tenth of a point on a game demo.

Verdict by job

  • Building pages, files or small apps inside a task: Tasks in Gemini, where the model menu and GoldieBench scores give you a real dial.
  • Standing jobs, team work and event triggers: Grok Bot, judged on routines and approvals rather than on a model score.
  • Anyone in the UK or EEA: Grok Bot, since Google doesn't offer tasks there yet.

Read the scorecards before you pick a model, and for the trainings behind this, join the AI Profit Boardroom.

Open the Gemini 3.6 Flash scorecard, set your task's model on purpose, and join the AI Profit Boardroom or get your Backlink Gameplan to settle the model side of grok bot vs gemini spark, 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 choose the model in Tasks in Gemini?

Yes. Google's help page says you can change the Gemini model and thinking level from the model menu at the top of a task.

What does Gemini 3.6 Flash score on GoldieBench?

Its GoldieBench model page shows a 7.08/10 average across 50 tasks, with one gold, two silver and one bronze medal.

Does the GoldieBench Grok score apply to Grok Bot?

No. That page covers xAI's Grok model, and GoldieBench also lists Grok as currently unranked. Grok Bot doesn't name its model.

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