Fable 5.2 cost $35 more and finished 75 minutes faster.
Fable 5.2 ($81.10 · 1h 56m):
→ water that actually moves, with chop and a wake behind the hull
→ a whale surfacing right next to the boat, spout and all
→ ice with depth, the arch glows from inside
Opus 5.5 ($45.97 · 3h 11m):
→ glass-flat water with a full mirror reflection of the ship and the ice
→ a whale breaching across the arch
→ ice that reads more like a painted wall up close
they made different bets: Fable spent its budget on motion, Opus on stillness
the price gap is 1.76x. the time gap is 75 minutes. the taste gap is the part u can't put in a table
the limits are real:
> one run each, no word on effort level or fast mode
> the prompt isn't public, so the ship name might have been in it
> a pretty scene isn't a benchmark
paste this into Claude Code ↓
"Build [scene] as a single html file with three.js.
1. Before coding, list the 5 details that make it believable
2. Pick one hero detail and tell me what you'll cut to afford it
3. Build it, then render 3 frames at 20/50/80% of the timeline
4. Grade each frame against the list, honestly
Show me the list and the cut before writing any code"
every model has a budget. watch where it spends it
1
Fable 5.2 on xhigh costs 63% more than high and scores exactly the same
@ArtificialAnlys ran every effort level. here's the ladder:
→ medium: $3.80, score 56
→ high: $9.08, score 62
→ xhigh: $14.83, score 62
→ max: $26.54, score 65
so max is 7x the price of medium for 9 extra points
and xhigh buys nothing at all
the part nobody's talking about:
→ Fable high: $9.08, score 62
→ Opus 5.5 max: $10.04, score 59
the "expensive" model on high beats the cheaper one cranked to the limit
where this breaks:
> it's AA's test set, not ur codebase
> it's API cost. ur 5h subscription limit is a different story
> one chart, one benchmark run
rule I'm keeping: pay for the jump from medium to high. stop there unless u can name what max fixes
paste this into Claude Code ↓
"I want to pick an effort level per task instead of one default.
1. Look at [my last 20 tasks in this repo] and sort them: boilerplate, debugging, architecture
2. For each group, suggest medium, high or max and say why in one line
3. Flag any task where u think high would have failed
4. Put it in a short table
Don't change any settings. Report only."
what's the last task that actually needed max?
10
a $165M movie left one thing out of its black hole on purpose. most AI demos leave it out by accident
one-shot black holes are everywhere this week. @TokenGremlin's is going around, he says he's being routed to GPT-6.2 Sol, which OpenAI hasn't officially released
so I had Claude Code build one that has to pass physics, not vibes. one html file, one shader, no three.js
what's actually in it:
→ every pixel is a light ray bent through curved spacetime, not a ring texture
→ the disk stops at 6M, where stable orbits run out
→ the side spinning toward u is brighter and bluer, the far side dim and red
→ a star wanders in, gets stretched into a stream and wraps the hole
that third one is what Interstellar skipped. the team behind Gargantua turned doppler beaming off so a lopsided disk wouldn't confuse audiences
the hump over the top and the ring underneath aren't modeled. they're the back of the disk, bent toward the camera. once the rays are honest u get them for free
where it went wrong:
> the first render was blown out white. the beaming was there, u just couldn't see it
> the stream ends cut off like scissors until they got faded
> the star's spiral is art direction, not a real tidal disruption sim
get the light right and the shape draws itself
paste this into Claude Code ↓
"Build a black hole in one HTML file with WebGL2 and a single fragment shader.
1. Trace every pixel as a light ray through Schwarzschild spacetime, G=M=c=1
2. Add a disk from r=6 to r=20 with doppler beaming and gravitational redshift
3. Put a procedural starfield behind it, lensed by the same rays
4. Expose a deterministic render(t) so I can export frames
Before writing any code, list which effects are physics and which are art"
Interstellar's clean ring or the lopsided real one, which would u ship?
1
14

