GoldieBench Blog · 6 min read
Grok Bot For Research + Which AI Models To Pair With It: Real GoldieBench Scores And What They Mean For Research Briefs
GoldieBench scores the models around your research Bot, not the one inside it; here is how to read the overall, Pages and reference columns for research work.

Grok Bot for research picks its own model for every request, so the benchmark question is which models you run alongside the Bot, not which one is inside it.
This page reads the live GoldieBench pages on 7 November 2026 and explains what the scores can and can't tell you about research work.
The Grok Bot facts come from xAI's FAQ and files and results page; every score is quoted from the GoldieBench page it appears on.
Want help choosing models for your own research stack? Get your Backlink Gameplan with my team, or try Polar Rank free for 3 days.
You can't choose Grok Bot's model
xAI's FAQ is direct: Grok Bot chooses the model per task, offers no model picker, and doesn't show which model handled a request.
The choice can change from one request to the next.
So no number on GoldieBench is a Grok Bot score, and the Grok row below shouldn't be read as one.
What the board can help with is the rest of your setup: the model in your Agent OS that writes the research question, and the one that turns the Bot's findings into a page or brief.
What GoldieBench tests, and why research isn't on it
The methodology page sets the rules: one fixed prompt per task, one attempt, and the result must be a single self-contained HTML file.
Scores are Julian's own 0 to 10, and the ranking is the average across every task a model was scored on.
The tasks are games, simulations, visuals and web pages.
The Grok deep dive spells out the gap: the bench doesn't measure multi-turn agent work, long-context recall or API latency.
Research is multi-turn agent work almost by definition, so treat these scores as a guide to who builds the clearest report, not who finds the best sources.
The overall board on 7 November 2026
| Rank | Model | Avg / 10 | Scored |
|---|---|---|---|
| 1 | Fusion | 8.59 | 47/47 |
| 2 | Claude Opus 5 | 8.27 | 50/50 |
| 3 | Hermes MoA | 8.17 | 47/47 |
| 4 | GPT-5.6 Sol | 8.16 | 50/50 |
| 5 | Claude Fable 5 | 8.10 | 47/47 |
| 6 | Qwen 3.8 | 8.10 | 45/45 |
| 7 | Grok | 8.09 | 43/47 |
Every row comes from the GoldieBench leaderboard as it read on 7 November.
The top seven sit within half a point of each other, so the order matters less than the category you care about.
LongCat-2.0 also shows 8.12 in the site's header ticker, but it isn't in the ranked rows above, so it's left out of this table.
Fusion is described on the models page as a multi-model panel, which is worth knowing if you prefer one vendor in your stack.
Pages is the nearest category to a research report
The Pages category has three tasks, and single-file pages are the closest thing on the bench to a research brief or a tracker your team opens.
| Model | Pages avg / 10 (3 tasks) |
|---|---|
| Fusion | 8.50 |
| Claude Opus 5 | 8.47 |
| GPT-5.6 Sol | 8.47 |
| Qwen 3.8 | 8.40 |
| Grok | 8.33 |
Those are the top five Pages averages on the leaderboard's Pages tab, read 7 November.
The board's Best at column tags GPT-5.6 Sol with Pages.
Three tasks is a small sample, so a gap of a few hundredths between rows says very little.
The one research-flavoured number on the board
The leaderboard has a Reference benchmark column next to the GoldieBench scores.
For Fusion, it lists 69.0% on "DRACO deep-research (premium panel)".
That's an outside reference shown beside Fusion's row, not a GoldieBench score, and it says nothing about Grok Bot.
It also isn't a Grok Bot feature: xAI's docs describe no deep-research mode, and the Bot does its research through public X, its browser and your files.
If you want a research-leaning model in your Agent OS to plan questions or check a draft, it's one data point to weigh alongside the Pages column.
The Grok row is a model, not the Bot
The Grok model page describes it as "Snappy + real-time — the X-native model", with a 256,000-token context window.
It lists live X context as a strength and suggests the model for workflows that need it.
Don't carry that over to Grok Bot.
The Bot's X access is a documented feature of its own: it searches and reads public X without a connected X account, read-only, whatever model handles the request.
On the bench, Grok's Pages builds include Landing and Webos at 9.0 each and Aipbpromo at 7.0, per its scorecard.
Conflicts on the bench to check before you quote
The methodology page still says models with zero scored tasks are unranked, and it names Grok among them.
The Grok model page has a trade-off line saying none of its demos has a curated verdict yet.
The live leaderboard, though, ranks Grok seventh at 8.09 with 43 of 47 tasks scored.
The Grok deep dive, marked "Data refreshed 2026-10-27", gives a board rank of eighth of 26.
Go by the dated leaderboard row, and recheck it before you quote a rank.
A pairing for a consultancy's weekly market brief
- In your Agent OS, a top Pages model such as Claude Opus 5 or GPT-5.6 Sol writes the week's research question and the brief's layout.
- Grok Bot takes the question, reads rivals' sites in its browser and searches public X for the client's niche.
- It returns a Google Doc with a dated source link after each claim and a list of what it couldn't confirm, as asked.
- Back in the Agent OS, the same model turns the doc into a one-page brief.
- A consultant opens the key links and signs it off before the client sees it.
A pairing for a shop watching its niche on X
- The shop's Agent OS model drafts a weekly X query: product terms, the shop's name and two rival names.
- A Grok Bot routine runs it every Monday, groups the posts by theme and links each one.
- Where Pages is available, the Bot builds a page with a table and filters; Pages isn't on Enterprise teams yet.
- The owner reads it and chooses what to stock or say; the Bot posts nothing.
What the scores can't tell you about research
No GoldieBench task checks whether a model picks reliable sources or quotes them accurately.
For that, use the checks xAI recommends: ask for direct source links, timestamps with time zones, and an explicit note of anything unverified.
Keep the outside-content risk in mind as well: xAI's security page says text from web pages is marked untrusted, which lowers the risk without removing it.
For a wider look at research agents, see our Grok Bot vs Perplexity Computer comparison.
The model trainings and the Agent OS are inside the AI Profit Boardroom, and my team is a Backlink Gameplan, or try Polar Rank free for 3 days.
Pick your brief-writing model from the dated board, keep a person on the final check, and join the AI Profit Boardroom to run Grok Bot for research.
FAQ
Which model does Grok Bot use for research?
xAI does not say. Its FAQ says Grok Bot picks the model for each task, has no model picker and does not show which model handled a request.
Does GoldieBench measure research quality?
No. It scores one-shot single-file HTML builds. The Grok deep dive says it does not measure multi-turn agent work, long-context recall or API latency.
Which GoldieBench category matters most for research briefs?
Pages, because single-file pages are closest to a report. On 7 November the top Pages averages were Fusion 8.50, Claude Opus 5 8.47 and GPT-5.6 Sol 8.47.
Is the Grok row a Grok Bot score?
No. It scores the Grok model on the bench. Grok Bot chooses its own models, and its X reading is a documented Bot feature.
- 📺 Video notes + links to the tools
- 🆓 Get a FREE AI course + community + 1,000 AI agents
- 🎥 Learn how I make these videos


