Hype Check · Matt Johnston | AI Engineer

Worth your time?

Name: Is "Grok 4.6 Just Dethroned Fable On My Bench" worth watching? Hype 19/100 | Clip'd
Item: Grok 4.6 Just Dethroned Fable On My Bench
Rating: 81
Author: Clip'd

Grok 4.6 Just Dethroned Fable On My Bench by Matt Johnston | AI Engineer

Watch on YouTube

Hype score · lower is better

19/100

Useful / Authentic

Teaches: solid hands-on benchmarks and demos with clear metrics and few sales pushes.

Substance78

Proof68

Pitch pressure10

Claim inflation20

The 15-second version

Post it, send it, settle the argument.

Download clip

Receipts

Why this score

0:28The host reads vendor claims that Grok 4.6 is 'essentially at the level of the flagships' (Fable, GPT 5.6), which sets a high expectation and makes independent proof important.
3:11During the Halo demo the model initially shows missing UI and no sound, a concrete failure that demonstrates the video runs live tests and reveals real, usable failure modes.
9:50The host highlights that Grok correctly flagged a fake JavaScript function instead of inventing one, a specific correctness check that increases the video's evidentiary strength.
13:39He reports exact cost and throughput numbers (Grok $6 per million output vs Fable $50, 117 tokens/sec) and uses those to rank Grok above Fable, giving measurable backing to the 'dethroned' claim.

Title deliveredTitle claims Grok dethroned Fable; the video shows benchmark scores tying/edging Fable and crowns Grok champion with cost and throughput evidence.

Skip the video

The short version

Kept by Clip'd

Matt runs Grock 4.6 blind through his multi-task benchmark suite and compares it to Fable, Opus, and others. He steps through interactive tests (Halo Arena, an XCOM-style tactical grid, a 3D printable model, a sales-page UI, and an interactive spreadsheet), reruns a flaky case, and finds Grock 4.6 scoring at the top of his weighted leaderboard while costing far less per token.

Keep the lesson

What it teaches

Evaluate models across diverse, weighted tasks rather than a single benchmark. Use a mix of interactive, visual, spatial, and code-reasoning tests so a high visual quality model that hallucinates or a logically strong model with weak UI doesn't mislead you. Avoid judging purely on one flashy demo.
Always rerun suspicious failures and check harness behavior. If a run shows obvious glitches (no audio, missing UI), rerun the test to separate a model bug from a harness or sandbox error. The common mistake is trusting a single run and drawing conclusions without replication.
Include cost and throughput in model selection. Raw accuracy matters but so do tokens-per-second and price-per-token because they change the practical ROI of deploying a model. The frequent error is optimizing only for accuracy while ignoring operational cost, which can make a cheaper model dominant in production.

Jump straight there

Key moments

0:58Product claim: Grock 4.6 is presented as "essentially at the level of the flagships," framing the whole test.
3:12Halo anomaly observed: First Halo run shows missing bullets, no audio, and odd camera angle, signals harness or generation instability.
6:37Tactical grid peak: "Absolutely crushed it, 100% on the XCOM", demonstrates Grock's ability to solve structured tactical reasoning accurately.
7:27High-fidelity 3D output: The printable Tardis STL is nearly perfect aside from a small ergonomic omission, showing strong spatial reasoning.
10:18Anti-hallucination moment: Grock correctly flags a fake JavaScript function in the sales-page reasoning trap, scoring the automatic test.
12:57Top-of-benchmark surprise: The harness reports a 99 overall score, placing Grock at or above Fable on the weighted leaderboard.

Lines worth keeping

Quotes

Grock 4.6 has just come out.0:00
Absolutely crushed it, 100% on the XCOM.6:37
Grock 4.6 is $6 per million token output versus Fable 5, which is $50.13:39
I think we crowned a new champion today on my benchmark15:43

Clip'd

Skip the video. Keep the lesson.

Score your next video free

The Hype Board