Grok 4.5 is HERE! (Real Tests and Review)
TL;DR
- Ron’s first impression is simple: Grok 4.5 was fast and cheap, but its output quality was “50/50” and needed more guidance. (source video XoPS2ULsKo4, 00:39)
- Four mini-projects plus Composer use showed $2.33 in the recorded session, although Ron also noted high token usage. Treat that as one dated run, not a current price quote. (source video XoPS2ULsKo4, 02:53; 02:56)
- The medieval castle earned a pass. The Chinese timber demo improved after Ron removed a feature, while the Helm’s Deep and Hogwarts builds failed his visual and functional checks. (source video XoPS2ULsKo4, 07:51; 09:04; 09:27; 10:52)
- The strongest result came from a real repository review: Grok 4.5 identified inconsistent time-window filtering in Ron’s X-news pipeline. Ron then planned to turn that finding into a proposal or
spec.mdfor his execution model. (source video XoPS2ULsKo4, 13:08; 13:36)
Ron’s verdict
Grok 4.5 is not the model Ron would trust to turn a feature-heavy creative brief into a polished one-shot result. His tested position is closer to “cheaper Opus 4.7” than “Fable 5 rival”: fast enough and inexpensive enough to explore with, but best when the task is narrow, the operator can review the output, and the work benefits from its X-data context. For this evidence set, use it as a supervised reviewer or scoped executor—not the only brain in the loop. (source video XoPS2ULsKo4, 03:31; 13:19; 13:48; 14:24)
Key moments
- 00:00 — The question and first impression: Ron frames the review around real coding quality, speed, cost, and how much guidance the model needs.
- 01:32 — Context and benchmark claims: the video records the launch-time context-window and benchmark positioning.
- 03:52 — Chinese architecture test: the same loaded prompt used with Fable 5 exposes weak roof and weathering detail.
- 08:03 — Medieval castle: a more convincing build earns a pass, with timeline-control and polish caveats.
- 09:06 — Helm’s Deep: rain is present, but the scene is too dark and does not resemble the target.
- 09:30 — Hogwarts: simplifying the prompt still does not produce the requested look or accessible interior.
- 10:54 — Real repository review: Grok reviews an X-news pipeline and finds a concrete filtering problem.
- 13:48 — Final verdict: Ron separates low cost and speed from raw creative capability.
Useful quotes
“it’s definitely fast, it’s definitely cheap, but the quality is kind of 50/50.” — Ron, source video XoPS2ULsKo4, 00:39
“That is not a roof.” — Ron, source video XoPS2ULsKo4, 05:36
“maybe loaded prompts are not very good for Grok 4.5.” — Ron, source video XoPS2ULsKo4, 07:58
“I’d say it’s like Opus 4.7, but faster and cheaper.” — Ron, source video XoPS2ULsKo4, 14:24
What the four creative tests actually showed
A loaded prompt asks for several difficult features or constraints at once. Ron’s first Chinese timber prompt combined a detailed 3D building, user photo upload, and a single self-contained HTML file. (source video XoPS2ULsKo4, 04:05) Its first result had multiple structures but lacked a finished roof and used color change instead of cracked, weathered texture. (source video XoPS2ULsKo4, 05:13; 05:39; 05:59) Ron explained that Fable 5 had rejected the incompatible upload requirement and built a demo, while Grok 4.5 attempted the whole request. (source video XoPS2ULsKo4, 06:24; 06:36)
Ron then removed the upload feature and asked only for a demo. (source video XoPS2ULsKo4, 06:45) The second attempt reached a finished roof but still lacked roof tiles; Ron said its scaffolding logic was correct, while the aging remained a color change. (source video XoPS2ULsKo4, 07:28; 07:32; 07:47) Ron called it a slight improvement over the first iteration. That supports a practical lesson: reduce the brief before assuming the model itself has failed. It does not turn the result into a pass. (source video XoPS2ULsKo4, 07:51)
| Test | What Ron observed | Result in this video |
|---|---|---|
| Chinese timber architecture | Multiple structures and better scaffolding on the second try, but weak roofing and color-only weathering instead of cracked-wood texture. (source video XoPS2ULsKo4, 05:39; 05:59; 07:32) | Improved after scope reduction; still below the requested finish. (source video XoPS2ULsKo4, 07:51) |
| Medieval castle | Correct-looking construction order, a siege x-ray, useful small-building detail, and no timeline control; Ron found it less attractive than Longcat. (source video XoPS2ULsKo4, 08:07; 08:24; 08:31; 08:36; 09:02) | Pass. (source video XoPS2ULsKo4, 09:04) |
| Helm’s Deep | It did not resemble the target, the scene was too dark, and rain detail was present. (source video XoPS2ULsKo4, 09:11; 09:14; 09:23) | Big fail. (source video XoPS2ULsKo4, 09:27) |
| Hogwarts | Ron had the model improve and plan a simpler prompt, but the scene was too bright and he could not reach the requested Great Hall interior. (source video XoPS2ULsKo4, 09:56; 10:05; 10:16) | Fail. (source video XoPS2ULsKo4, 10:52) |
These are four demonstrations, not a general benchmark suite. They are most useful as failure-mode evidence: Grok 4.5 could produce structure, but it needed the brief narrowed and still missed visual polish or requested interaction in three tests; the castle was the one pass. (source video XoPS2ULsKo4, 07:51; 09:04; 09:27; 10:52)
Where the real code review was better
Ron moved from creative demos to an early X-news project that collected and filtered social posts for a daily report. He had previously used Fable 5 for planning and MiniMax M3 for execution, then asked Grok 4.5 to learn the repository and judge the pipeline. (source video XoPS2ULsKo4, 11:07; 12:05; 12:36)
The review said the broad pipeline was sound, then isolated time-window consistency as the main gap in how relevant and irrelevant posts were filtered. Ron’s next move was not to give Grok full control: he proposed asking it for a spec.md, handing that proposal to his MiniMax agent, and iterating from there. (source video XoPS2ULsKo4, 12:58; 13:08; 13:36)
Decision aid: should you route this task to Grok 4.5?
Use this only as a decision aid for the tests recorded on July 9, 2026:
- Is this a one-shot creative build with several visual and functional demands? → Do not choose Grok 4.5 on this evidence alone; three of the four creative tests needed major correction or failed Ron’s check. (source video XoPS2ULsKo4, 07:51; 09:27; 10:52)
- Can you split the brief into one demo-sized objective? → Try the smaller task and inspect the output. Removing the upload requirement produced a slight improvement in the Chinese architecture attempt. (source video XoPS2ULsKo4, 06:45; 07:51)
- Is the job a review of an X-heavy news or aggregation pipeline? → This is the clearest fit in the video. Ask for a diagnosis, then turn the finding into a proposal or spec for a separate execution pass, as Ron planned. (source video XoPS2ULsKo4, 13:19; 13:36)
- Does cost matter, and can a human or stronger planning model supervise? → Grok 4.5 is a candidate for the cheap pass. The video’s $2.33 session is evidence for that run only, not a guaranteed budget. (source video XoPS2ULsKo4, 02:53; 14:12)
What changed since this video
The video was published July 9, 2026, and this companion was source-checked on July 18, 2026 against its saved transcript and timestamp segments. No outside release note, pricing page, benchmark result, or follow-up test was added. The recorded claims about a 500k context window, a possible one-million-token upgrade “by next week,” benchmark placement, and the $2.33 session therefore remain launch-time statements from the video, not assertions about the model available today. (source video XoPS2ULsKo4, 01:33; 01:41; 02:35; 02:53)
Related
Watch on YouTube
Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=XoPS2ULsKo4