GoldieBench Blog · 7 min read
Grok Bot vs Copilot: Which AI Models Each One Runs, What the GoldieBench Scores Say, and How To Pair Them
Copilot hands you the model dial on paid plans; Grok Bot keeps it. Here is what that means for every benchmark score you read.

Grok Bot vs Copilot splits hardest on model choice: GitHub Copilot lets you pick models on most paid plans, while Grok Bot picks for you and never says which one ran.
That single difference decides how much a benchmark score can help you.
With Copilot, a score can guide what you select.
With Grok Bot, a score describes models you cannot steer toward.
Copilot on this page is GitHub Copilot, the coding product.
Below are the GoldieBench numbers that matter for this pairing, each quoted with the page I read it on, on 7 November 2026.
Want help choosing models for your own stack? Get your Backlink Gameplan, or try Polar Rank free for 3 days.
Model choice is the real split
Most head-to-heads stop at features.
For a benchmark site, the useful question is who controls the model.
On the Copilot side, the plans page ties model access to the plan you are on.
On the Grok Bot side, the FAQ answers "Can I pick the model?" with a flat no.
Everything else on this page follows from those two answers.
What Copilot lets you pick, plan by plan
| Copilot plan | Models, as GitHub words it |
|---|---|
| Copilot Free | Auto model selection only |
| Copilot Student | Auto model selection only |
| Copilot Pro | A selection of models |
| Copilot Pro+ | Access to premium models |
| Copilot Max | Priority access to premium models |
| Copilot Business | Access to premium models |
Those labels come straight from the plans table, with prices left out.
In the IDE, GitHub's IDE page adds bring-your-own-key and models from VS Code extensions.
Copilot automations also let you set the model and reasoning level per automation, according to the automations page.
So on a paid plan, a benchmark can feed straight into a setting you control.
Why Grok Bot gives you no picker
xAI's FAQ says Grok Bot chooses the model for each task, aiming for the best outcome on that task.
The choice can differ from one request to the next.
Grok Bot does not show which model handled a request, and usage views record it all as Grok Bot usage.
A Cursor team's model allowlist or blocklist does not apply to it either.
The security page adds that only providers on Cursor's sub-processor list can process your data.
For a Grok Bot user, then, a model score is background reading, not a setting.
The GoldieBench scores for models Copilot lists
| Model | GoldieBench average | Tasks | Medals | Copilot plans that list it | Page read |
|---|---|---|---|---|---|
| Claude Opus 5 | 8.27/10 | 50 | 13 gold, 7 silver, 5 bronze | Pro+, Max, Business, Enterprise | /models/opus5 |
| Claude Fable 5 | 8.10/10 | 47 | 4 gold, 2 silver, 1 bronze | Pro+, Max, Business, Enterprise | /models/fable-5 |
| Claude Opus 5.5 | 7.57/10 | 50 | 0 gold, 0 silver, 1 bronze | Pro+, Max, Business, Enterprise | /models/opus55 |
| Grok 4.7 | 7.15/10 | 20 | 0 gold, 2 silver, 3 bronze | Pro, Pro+, Max, Business, Enterprise | /models/grok47 |
The plan column comes from the model tables on GitHub's plans page; Free and Student stay on auto selection.
Every score is a one-shot result: the same fixed prompt for every model, one try, one self-contained HTML file, scored 0 to 10.
The methodology page rules out iteration and best-of-N picks, and the full field is on the models page.
What Claude Opus 5's 8.27 tells a Copilot user
Claude Opus 5 posts the highest average of the four, 8.27/10 across 50 tasks with 13 golds.
On Copilot it needs Pro+ or higher; plain Pro does not list it.
For one-shot build quality, it is the bench's strongest pick that Copilot offers on these plans.
Its newer sibling, Claude Opus 5.5, averages 7.57/10 over the same 50 tasks, so newer is not automatically higher on this bench.
Claude Fable 5 and the data note on Copilot
Claude Fable 5 averages 8.10/10 over 47 tasks, a whisker behind Opus 5.
Its page also lists an outside reference number, SWE-bench Verified at 95.0%, and labels it as not a GoldieBench medal score.
GitHub's plans page attaches a note to Claude Fable 5 and Fable 5.1: when either is used, Anthropic keeps prompts and outputs by default to run safety classifiers.
Approved enterprises can ask for zero data retention on those two models through the end of 2026, and an admin still has to switch each model on.
Check that note against your client contracts before you make Fable 5 the default.
Grok 4.7 in Copilot is not a Grok Bot score
Copilot lists Grok 4.7 on every paid plan, Pro included.
On GoldieBench, Grok 4.7 averages 7.15/10 over 20 scored tasks, and its deep-dive page ranks it #18 of 26 ranked frontier models.
That score tells you how Grok 4.7 behaves when you pick it inside Copilot.
It does not tell you anything about Grok Bot, because xAI does not say which model Grok Bot runs.
For a Grok Bot user, the bench matters more on the coding side: xAI's How Grok Bot Works page says Bots hand coding work to Cursor Cloud Agents.
What a one-shot score misses for agent work
- Loops. Agents plan, run tools, read results and try again; a one-shot build does none of that.
- Review gates. Copilot's pull request review and Grok Bot's Auto Review sit outside any model score.
- Tool access. Copilot's MCP tools run without asking once configured, per the MCP settings page; no benchmark captures that risk.
- Browser work. Grok Bot spends much of its time on a shared cloud computer clicking through sites, which the bench does not test.
A pairing plan for a small business
A web agency
Developers on Copilot Pro+ test Claude Opus 5 and Claude Fable 5 on a few real issues assigned to the cloud agent, reviewing every pull request.
The account manager runs a Bot that reads merged pull requests through its GitHub connector and drafts client updates, with no model setting to worry about.
A bakery with an online order page
If the order page is a hosted builder, Copilot has no code to work on, and model scores do not matter for the Bot.
A Bot reads new orders from email each morning, drafts confirmations and stops at an Ask first rule before sending.
A solo developer selling a small app
On Copilot Free, model choice is automatic, so the bench is background reading.
Moving to Pro or above opens model selection, and that is when the GoldieBench table becomes a shortlist.
For the model side of a related matchup, read Grok Bot vs Cursor on GoldieBench.
Verdict: pick by job, then by model
Choose Copilot for code on GitHub, then use the bench to pick a model your plan offers.
Choose Grok Bot for browser and app work, and accept that the model is xAI's call, not yours.
Run both if you have code and a back office, and spend your benchmark attention on the Copilot side, where it changes a setting.
Inside the AI Profit Boardroom you can compare model settings for both tools with other members.
Check your Copilot plan's model list, test the top scorer on one real task, and get your Backlink Gameplan if you want my team to help you weigh grok bot vs copilot, or try Polar Rank free for 3 days.
FAQ
Can I choose the model Grok Bot uses?
No. xAI's FAQ says Grok Bot picks the model for each task, does not show which one ran, and ignores Cursor team model allowlists.
Does Copilot Free let me pick a model?
GitHub's plans page lists Copilot Free as auto model selection only. Model choice opens up on Pro and above.
Which Copilot plans list Claude Opus 5 and Fable 5?
GitHub's plans page lists both on Pro+, Max, Business and Enterprise, not on Pro. Fable 5 carries a note that Anthropic keeps prompts and outputs by default for safety checks.
Is the Grok 4.7 score a Grok Bot score?
No. Grok 4.7 is a model you can pick inside Copilot, and its 7.15/10 describes that model on one-shot builds. xAI does not say which model Grok Bot runs.
- 📺 Video notes + links to the tools
- 🆓 Get a FREE AI course + community + 1,000 AI agents
- 🎥 Learn how I make these videos


