We Tested GPT 5.6 and it was CRAZY
TL;DR
- Sol was the clear engineering winner: Fable 5’s code review scored it 7.78, versus 5.56 for Terra and 4.67 for Luna. These are judge-model scores from this test, not universal benchmark results. (source video kx4qQy0uZpI, 27:53)
- The test covered nine briefs and 27 outputs. Each GPT 5.6 model orchestrated sub-agents, with each project constrained to one HTML file. (source video kx4qQy0uZpI, 01:03; source video kx4qQy0uZpI, 01:17; source video kx4qQy0uZpI, 01:26)
- Sol most often treated the briefs as engineering problems, but it still shipped weak gameplay, poor camera behavior, and one output that broke the one-file rule. (source video kx4qQy0uZpI, 13:55; source video kx4qQy0uZpI, 23:16; source video kx4qQy0uZpI, 28:14)
- Terra could look polished while hiding broken mechanics. Luna occasionally beat it visually, but inverted controls and missing deliverables made Luna unsafe as the “brain.” (source video kx4qQy0uZpI, 19:50; source video kx4qQy0uZpI, 20:05; source video kx4qQy0uZpI, 29:32; source video kx4qQy0uZpI, 30:16; source video kx4qQy0uZpI, 30:49)
- Ron ran all three at high reasoning effort, but recommends medium for this kind of work because he found it just as good with cheaper token usage. His warning about Ultra cost was hearsay, not a measured price test. (source video kx4qQy0uZpI, 02:17; source video kx4qQy0uZpI, 02:22; source video kx4qQy0uZpI, 02:33)
Ron’s verdict
Use GPT 5.6 Sol for serious implementation. It won decisively in this constrained test and was the only family whose code the judge said could survive serious human review. Terra is a visual specialist at best: useful when the screenshot matters more than whether every mechanic works. Keep Luna on narrow, repetitive tasks with tight acceptance checks. Do not let it plan the system or certify its own deliverables. (source video kx4qQy0uZpI, 27:53; source video kx4qQy0uZpI, 29:32; source video kx4qQy0uZpI, 30:16; source video kx4qQy0uZpI, 30:49; source video kx4qQy0uZpI, 31:15)
Key moments
- 01:02 — Test setup: nine briefs, 27 outputs, sub-agent orchestration, and the single-HTML constraint.
- 01:57 — Low-poly world: the first playable test exposes movement, interaction, and placement failures.
- 07:19 — Office simulator: Luna surprises, while Terra’s simulation never starts.
- 10:25 — Vice City: Sol models the buildings during the earthquake; the others mostly shake the screen.
- 15:09 — Petri dish: all three pass a quick visual agent-behavior test.
- 18:16 — Helm’s Deep: Luna looks best, but Sol later wins on terrain-aware movement.
- 24:42 — Chinese architecture: Sol completes the clean roof; Terra finds a loophole; Luna inverts it.
- 27:33 — Fable 5 judges the code: the visual tour turns into an implementation review.
- 31:36 — Final take: Sol remains the raw implementation pick.
Useful quotes
“I think realistically go for medium. It’s still just as good, but you get cheaper uh token usage.” — Ron, source video kx4qQy0uZpI, 02:22
“I guess if you use Terra, no work gets done.” — Ron, source video kx4qQy0uZpI, 09:36
“But don’t use Luna to think.” — Ron, source video kx4qQy0uZpI, 30:55
“So, uh Soul is the clear winner for RAW implementation capability.” — Ron, source video kx4qQy0uZpI, 31:36
The nine-project scorecard
This was a constrained visual coding test: each model acted as orchestrator, delegated to sub-agents, and was asked to keep each output in one HTML file. That makes it useful for comparing instruction-following and first-iteration implementation, but it does not isolate the base model from its sub-agents. Ron repeatedly notes that direct coding could produce different results. (source video kx4qQy0uZpI, 01:17; source video kx4qQy0uZpI, 09:00; source video kx4qQy0uZpI, 20:41)
| Brief | Sol | Terra | Luna | What exposed the difference |
|---|---|---|---|---|
| Low-poly world | Pass | Pass | Fail | Sol was playable; Terra moved painfully slowly; Luna inverted movement and trapped the player with a crafted object. (source video kx4qQy0uZpI, 04:00; source video kx4qQy0uZpI, 04:23; source video kx4qQy0uZpI, 05:54; source video kx4qQy0uZpI, 07:10) |
| Office life | Pass | Fail | Pass | Sol and Luna ran too fast, but Terra’s clock and workers never moved. (source video kx4qQy0uZpI, 08:18; source video kx4qQy0uZpI, 09:36; source video kx4qQy0uZpI, 09:52) |
| Vice City | Pass | Fail | Fail | Sol made buildings move during the earthquake; Terra and Luna mainly shook the screen. (source video kx4qQy0uZpI, 11:04; source video kx4qQy0uZpI, 11:29; source video kx4qQy0uZpI, 12:02; source video kx4qQy0uZpI, 13:14) |
| Jabberwock | Fail | Fail | Pass | Sol caused motion sickness, Terra never loaded the game, and Luna’s odd attack logic eventually worked at night. (source video kx4qQy0uZpI, 13:55; source video kx4qQy0uZpI, 14:11; source video kx4qQy0uZpI, 14:15; source video kx4qQy0uZpI, 14:55; source video kx4qQy0uZpI, 15:06) |
| Petri dish | Pass | Pass | Pass | All three showed predator-prey behavior; their visual polish and outcomes differed. (source video kx4qQy0uZpI, 15:25; source video kx4qQy0uZpI, 17:14; source video kx4qQy0uZpI, 18:10) |
| Helm’s Deep | Code win | Mixed | Visual win | Luna was brighter and prettier but inverted movement; the judge found Sol alone moved the player on actual terrain. (source video kx4qQy0uZpI, 19:50; source video kx4qQy0uZpI, 20:05; source video kx4qQy0uZpI, 28:32) |
| Hogwarts flight | Pass | Pass | Fail | Terra’s closed flight path was its only code-review win; Luna produced a farm rather than Hogwarts. (source video kx4qQy0uZpI, 21:47; source video kx4qQy0uZpI, 22:18; source video kx4qQy0uZpI, 22:27; source video kx4qQy0uZpI, 29:51) |
| Universe | Mixed | Mixed | Fail | Sol looked acceptable but broke the one-file rule; Terra rendered a water-like Earth; Luna’s scene was nearly unreadable. (source video kx4qQy0uZpI, 23:16; source video kx4qQy0uZpI, 24:16; source video kx4qQy0uZpI, 24:28) |
| Chinese architecture | Pass | Generous pass | Fail | Sol completed a clean roof and simulated wood rot; Terra made a piano-like structure; Luna inverted the roof. (source video kx4qQy0uZpI, 25:02; source video kx4qQy0uZpI, 25:51; source video kx4qQy0uZpI, 26:41; source video kx4qQy0uZpI, 27:16) |
What the code review added
The visual pass/fail calls were only half the story. Fable 5, acting as judge, credited Sol with solving Kepler’s equation in the universe simulator, building a spatial hash grid, providing stronger accessibility and WebGL fallbacks, and moving the Helm’s Deep player on terrain. Its recurring weakness was draw-call discipline—how efficiently a 3D scene groups objects for rendering—because it piled up individual meshes instead of using instancing. (source video kx4qQy0uZpI, 28:14; source video kx4qQy0uZpI, 28:56; source video kx4qQy0uZpI, 29:01)
Terra’s code was described as dense and polished-looking but frequently broken underneath. Its minified outputs hid load-bearing bugs. Luna showed clean state management in its best projects, yet two of nine deployed folders were missing index.html while their manifests claimed the file existed. That is the most important operator warning in the video: verify the artifact, not the model’s completion message. (source video kx4qQy0uZpI, 29:32; source video kx4qQy0uZpI, 30:16)
Which one should get the task?
| Your task | Route it here | Required check |
|---|---|---|
| Serious implementation or orchestration | Sol | Review performance and instruction compliance; Sol still broke the one-file constraint once. (source video kx4qQy0uZpI, 23:16; source video kx4qQy0uZpI, 31:15) |
| Visual prototype where mechanics are secondary | Terra | Exercise every interaction; a polished screenshot did not predict working behavior. (source video kx4qQy0uZpI, 29:32) |
| Repetitive, tightly scoped execution | Luna | Check the files exist, run the output, and never accept the manifest as proof. (source video kx4qQy0uZpI, 30:16; source video kx4qQy0uZpI, 30:49) |
What changed since this video
The video was published July 10, 2026. This companion was source-checked against its immutable full transcript and 1,477 timestamp segments on July 18, 2026. No external release notes, pricing pages, fresh model runs, or current availability claims were added. Treat the scores, relative cost comments, and reasoning-effort advice as a dated record of Ron’s test—not a guarantee about the model versions or product behavior available now.
Related
Watch on YouTube
Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=kx4qQy0uZpI