Hermes Agent SuperGrok + Which AI Models To Run On It, With Real GoldieBench Scores For Grok 4.6, Grok 4.7 And Hermes MoA

Hermes Agent SuperGrok gives you xAI's chat, image and video models inside Hermes through one login, so the real question is which of those models to run.

GoldieBench has scored three xAI entries, and the spread between them is wide: 8.09, 7.15 and 5.95 out of 10.

Below are the scores I read on the live GoldieBench pages on 31 October 2026, what each one means for a Hermes setup, and which model I'd pick per job.

If you'd rather have my team make that call with you, book a free SEO strategy session.

What The SuperGrok Login Puts In Your Model Picker

The Hermes xAI Grok OAuth guide lists six chat models on the xai-oauth provider.

They are grok-4.6, grok-build-0.1, grok-4.3 and three grok-4.20-0309 variants.

grok-4.6 is the default and is pinned to the top of the picker.

The list is pulled live from the models.dev cache, so newer xAI releases show up once the cache refreshes.

The provider shipped in Hermes v0.14.0 on 16 May 2026.

The GoldieBench Scores For xAI Models

Every score below is an average of Julian's 0 to 10 verdicts across the same one-shot tasks the rest of the field runs.

The GoldieBench leaderboard covers 33 models and 50 tasks, with 1,310 live demos.

Hermes MoA sits in the table because it's the Hermes Mixture of Agents setup, not a single model.

How GoldieBench Builds These Scores

The leaderboard ranks each model by Julian's own 0 to 10 scores, averaged across every task it ran.

Medals are per-task ranks, so gold means the best result on that one task.

Models without scored verdicts stay unranked.

The Grok 4.6 page says its game builds were played through their full gameplay arc before scoring.

The Grok 4.7 page says each build was played on a Metal GPU before a vision judge scored the frame.

So these are played results, not screenshots.

Grok 4.6: The Hermes Default Scores 5.95

This is the model you get if you accept the defaults.

Its GoldieBench page shows 5.95 out of 10 across 20 game builds, with one gold medal.

The range is extreme.

Arcade scored 8.7 for gold, while Skyrim and Voxelcraft both scored 1.5.

The page names flight models as the weak spot: Flightsim scored 5.0 and Dogfight 4.8.

The strength that matters most for an agent is self-repair.

Handed the exact runtime error, it repaired 4 of 4 failed games on the first retry, per the page.

That's exactly what an agent loop in Hermes does, so the default suits error-and-retry work better than its average suggests.

Grok 4.7: 7.15 And 14 Of 20 Head-To-Heads

The Grok 4.7 page shows 7.15 out of 10 on the same twenty game briefs where Grok 4.6 averaged 5.95.

It won 14 of the 20 head-to-heads.

Its best one-shots were Neoncity at 8.7, Racing at 8.7 and Arcade at 8.6.

Two of its 20 builds still died on load or never moved.

The Hermes OAuth guide's model table doesn't list Grok 4.7.

Check your picker after the models.dev cache refreshes, and only switch if it appears there.

If it does, it's the obvious upgrade for game and app builds on these scores.

Both pages list a 500,000-token context window, so the switch costs you no context.

The Grok Entry: 8.09 Across 47 Tasks

GoldieBench also has an entry simply called Grok, with a 256,000-token context window and a 2026-04 release date.

It averages 8.09 out of 10, with 43 of its 47 tasks scored and five gold medals.

Its top score is Twilightvale at 9.5.

The page lists real-time X timeline access as its unique strength.

In Hermes, that same signal reaches the agent through the x_search tool, which switches itself on once xAI credentials exist.

The page doesn't map this entry to a Hermes model ID, so I won't either.

Why The Intelligence Index Numbers Don't Line Up

This trips people up, so read it carefully.

The Grok 4.6 page quotes an Artificial Analysis Intelligence Index of 61.

The Grok 4.7 page quotes 46, but on version v4.3.2 of that index.

Different index versions aren't comparable, so 46 is not a downgrade from 61.

Both pages also say these reference numbers are external, not GoldieBench medal scores.

For Hermes decisions, trust the same-prompt GoldieBench averages first.

What The Coding Numbers Say

The pages quote xAI's own coding benchmarks.

Grok 4.6 shows DeepSWE v1.1 at 65.9% and CursorBench v3.2 at 69.9%.

Grok 4.7 shows DeepSWE v1.1 at 71.0% on high effort and Terminal-Bench 4.0 at 38.0%.

The Grok 4.7 page calls Terminal-Bench the weak spot.

That matters in Hermes, because terminal work is a big share of what an agent does.

Pairing Grok With Hermes MoA

Hermes MoA scores 8.17 out of 10 across 47 tasks and ranks third overall.

It sends one prompt to a panel of models, then an aggregator writes the final answer.

The default panel is Claude Opus 4.8 and GPT-5.5, aggregated by Opus 4.8, all through OpenRouter.

The page says each build takes about 110 to 140 seconds, against a single call for a solo model.

So the pairing is simple: Grok on the SuperGrok login for fast daily work, MoA for the few high-stakes prompts.

Which Model I'd Pick Per Job

Rows without a GoldieBench score say so, because I only quote numbers I read on the site.

What The Scores Can't Tell You

A benchmark can't see your login.

Some standard SuperGrok accounts log in fine, then get an HTTP 403 because xAI gates OAuth access by tier.

That's issue #26847 on GitHub.

When that happens, the providers page route is an API key in ~/.hermes/.env plus the plain xAI provider.

The models and their scores stay the same on either route.

Want these models running inside a full system? Join the AI Profit Boardroom and pick your Grok model from day one.

Also On Our Network

• Agent Operatorsthe five commands that connect SuperGrok to Hermes, plus the 403 fix
• AI Income Deskwhether a SuperGrok-powered Hermes agent earns its place in a small business
• AI Tool Verdicta seven-criteria scorecard rating Hermes on SuperGrok at 7.5 out of 10
• agentos.guidefive prompts that make your agent connect SuperGrok to Hermes for you
• aiprofitboardroom.comthe complete Hermes Agent SuperGrok guide on the AI Profit Boardroom blog

Pick your model from the table above, and if you want my team to check the choice against your real work, book a free strategy session before you commit your Hermes Agent SuperGrok setup.