gpt 5.6 sol vs gpt 5.6 terra vs gpt 5.6 luna vs gpt 5.5

the test – 4 prompts, single-file html, three.js from a cdn, fully procedural, no external assets. all run via @aimlapi

the takeaway up front: every openai frontier model here needed correcting before it would run. even sol, the one benched against fable, shipped broken code. the only thing that reliably fixed it was claude code on opus 4.8

prompts:

1. oceania beach house opening onto turquoise sea

2. wooden treehouse high inside a giant jungle tree

3. cozy finnish cabin at night, lit stove, snow outside

4. warm african earthen room, arched openings onto sand dunes

overall results (summed across all 4 scenes):

cost
#1 luna – $0.20
#2 terra – $0.55
#3 sol – $1.72
#4 gpt 5.5 – $1.84

tokens
#1 luna – 24,697
#2 terra – 28,001
#3 sol – 43,821
#4 gpt 5.5 – 47,147

lines of code
#1 luna – 2,041
#2 terra – 2,356
#3 gpt 5.5 – 3,273
#4 sol – 3,978

broken files / bugs shipped
#1 sol – 1
#2 gpt 5.5 – 2
#3 terra – 3
#4 luna – 4

the mistakes – all analyzed and fixed by claude code on opus 4.8 (high):

• the recurring one, across almost every model: three.js loaded as an es module from the cdn with no import map. orbitcontrols internally imports bare "three", the browser can't resolve it, the whole script aborts – blank page. terra, luna and 5.5 all made this call

• sol's snowy cabin named a variable throw – a reserved keyword. hard syntax error, nothing runs

• luna's earthen room went further: it reached for the legacy global-script build (three.min.js + examples/js) that three.js deleted years ago, so both cdn tags 404 and three is undefined – plus a shader that referenced an undeclared uniform

none of these are logic bugs. they're setup and boilerplate the model got wrong and never checked

observations:

• gpt 5.6 sol and gpt 5.5 spend essentially the same money per task – $1.72 vs $1.84

• gpt models are fast, but the tradeoff is no double-check. they commit to the first plan, write the code, and hand back something that doesn't run. every file we got failed to load out of the box

• gpt 5.6 sol is a smooth 3d html builder when it works, but it's still nowhere near fable 5 pre-nerf, and at this point even @claudeai opus 4.8 outperforms it – opus is the model that caught and fixed all four models' code