Scene artist and audio producer: Fable high vs Opus xhigh vs Opus high
s05 Safe Word's first preparation batch, each made three times from the same brief: the four new sounds, the 2D cast, and ep01's 3D sets. Made 2026-11-16.
Time, tokens and cost
| Task | Setting | Wall time | Turns | Output | Cache write | Cache read | Cost |
|---|---|---|---|---|---|---|---|
| 4 sounds | Fable 5.2 · high * | 3.5 min | 16 | 20k | 121k | 1.4M | $2.29 |
| 4 sounds | Opus 5.5 · xhigh | 11.3 min | 38 | 61k | 160k | 3.8M | $2.78 |
| 4 sounds | Opus 5.5 · high | 5.3 min | 24 | 29k | 100k | 1.6M | $1.40 |
| Cast | Fable 5.2 · high | 43.0 min | 90 | 206k | 507k | 29.8M | $19.28 |
| Cast | Opus 5.5 · xhigh | 76.3 min | 165 | 282k | 635k | 72.4M | $23.30 |
| Cast | Opus 5.5 · high | 43.7 min | 120 | 176k | 482k | 38.0M | $13.53 |
| 3D sets | Fable 5.2 · high | 36.6 min | 62 | 177k | 481k | 19.4M | $15.77 |
| 3D sets | Opus 5.5 · xhigh | 68.0 min | 161 | 274k | 607k | 66.8M | $21.88 |
| 3D sets | Opus 5.5 · high | 37.0 min | 96 | 146k | 394k | 26.5M | $10.19 |
| All three tasks (times added up; the three agents of a setting ran at the same time) | |||||||
| All three | |||||||
| Fable 5.2 · high | 83.1 min | $37.34 | |||||
| Opus 5.5 · xhigh | 155.6 min | $47.96 | |||||
| Opus 5.5 · high | 86.0 min | $25.12 | |||||
- How it was counted: from each agent's own transcript: wall time from its first to its last message; tokens from the usage the API returned on each reply, counted once per reply. Thinking tokens are part of output.
- Cost = tokens × API list price per million (Claude API skill, prices cached 2026-11-03). Fable 5.2: input $8, cache write $10, cache read $0.20, output $40. Opus 5.5: input $4, cache write $5, cache read $0.20, output $20. Every agent used the 5-minute cache (writes at 1.25× input); none wrote to the 1-hour cache. On a subscription plan this is the API-equivalent cost, not what you are billed. This session's own orchestration is not included.
- * Fable's sound row stops where its narration work began (02:47). Its agent did the sounds, ep01's narration and the mix in one run; the Opus agents did the sounds only. Fable made two small level changes to its sounds later, during the mix, so its true sounds-only cost is a little higher than shown.
How the comparison was run
- Same brief: each Opus agent got the producer's exact brief to the Fable agent, taken from the s05 producer session's transcript. Changes: the sound brief keeps only the four sounds (no narration or mix, so no OpenRouter credit); each art brief adds one last step, naming the lab frame that best shows each item below.
- Same starting point: each Opus run worked in its own git worktree at
e1ed6e2, the commit just before Fable started, so it could not see Fable's work. Fable's art is rendered from its own delivery commitee8fce2, before any later fix. Branches:cmp/opus-xhigh,cmp/opus-high. - Same set-up: per setting, three agents ran at once, as Fable's did: one sound agent and two artists sharing one tree, the cast artist using stand-ins for any set not yet committed. Same agent files and tools; only the model and effort changed.
- A bias in Opus's favour: the agents' preloaded skills came from the main repo, which now holds about 64 lines of lessons Fable wrote while doing this very task (e.g. silver hair made Mom read as 80; camera height 1.48 m). The briefs told the Opus agents to ignore those lines, but they read them.
- Frames matched by content: each artist built its own lab, so frame numbers differ. Each row below shows the frame its own artist named as best (I picked Fable's from its frame names). Many frames use stand-in sets or stand-in people, so judge the drawing itself, not the room around it.
- Sounds are not level-matched: each file is at its own level;
mix.pysets the real level from each sound'slevel_db. Nobody has listened to any of them yet.
Sounds
ring-in A phone ringing on a wooden table, two rings (ep01 s10; ep04, ep05).
1.25 s. Two rings of a fast two-note trill (D6 / F#6) over the vibration motor rattling on wood.
1.15 s. Two rings, each four quick marimba-like notes over the buzz rattling on wood; a modern melodic ringtone.
1.43 s. Two rings of a trill flicking 1055 / 1320 Hz, 18 times a second, plus the motor rattling on wood.
call-end Three falling tones of a dropped, declined or busy call (ep01 s9, s12; ep03, ep06).
0.56 s. A5, F5, D5, soft sine tones, the last held a little longer.
0.55 s. Three soft beeps falling down a minor chord.
0.55 s. E6, C#6, A5, 0.11 s each; no room echo (an echo cancelled the first beep).
door-chime A doorbell app's motion alert, one soft double tone (ep03).
1.05 s. E6 then C6 0.27 s later, soft bell tones.
1.0 s. A soft ding-dong, the second note 0.22 s after the first.
1.0 s. A soft falling double tone, A5 then F5 at 0.22 s.
dish-drop A pie dish landing flat on a tiled floor (ep04).
0.75 s. Rim clack, the filling's slap, the tile's knock, then six rocking taps; nothing breaks.
0.45 s. Thud and splat under the rim's clack, a short settling chatter, a few crumbs. Remade once: the first was too dull for a phone speaker.
0.73 s. A dull thud, a short rim clack, then three settling taps.
Cast
Mom seated at the table, hands flat, eyes wide
Mom standing in the hall
Mom side-on under the lot lamp
Mom's back at the front door
Mom smiling at the table
Kate seated
Kate's contact photo
The man in the cap, from above
The man in the cap, far shape
The clerk behind the counter
Leo in the train seat
Leo standing, eyes wide
Leo in the store line
Leo's dark shape at the alley
ep01 3D sets
Kitchen wide
ep01 opening: Mom's place at the table, phone on it
Phone on the kitchen table, close
Train seat
Phones from straight above
Call screen LEO, blue wave
Amber wave screen
Calling LEO...
Leo can't take your call right now.
Bank app payment screen
Leo's bank alert
Leo's post with the sound wave
Every frame
Each lab in full, with each artist's own list of weak spots. Click a frame to open it full size.
Cast labs
Fable 5.2 · high: every frame of lab-safeword-cast (55)
Its own report on weak spots: Frames 0-48 on stand-in sets; 49-54 on its own real kitchen and train (they were committed mid-task).
Opus 5.5 · xhigh: every frame of lab-safeword-cast (39)
Its own report on weak spots: Mom's glasses read as a hair clip side-on; people small in the wides; the cap man facing the door is close to the turned-away view; the envelope can't be seen front-on. Frames 34-38 on its own real kitchen and train.
Opus 5.5 · high: every frame of lab-safeword-cast (26)
Its own report on weak spots: Every set is a stand-in; side-on hands on the table look like sausages; Kate's arms read a little as reaching; Mom's glasses small; the cap's brim reads as a band over the face; the queue hats barely show.
3D labs
Fable 5.2 · high: every frame of lab-safeword-3d (47)
Its own report on weak spots: People are stand-ins. (Its report listed no weak spots for the sets.)
Opus 5.5 · xhigh: every frame of lab-safeword-3d (60)
Its own report on weak spots: The bottom 40% of the wide is floor; the oak looks streaky close up; the pie kinds don't read from the wide; the phone on the tray is a thin sliver side-on. People are stand-ins.
Opus 5.5 · high: every frame of lab-safeword-3d (50)
Its own report on weak spots: Lots of empty floor in the wide and medium kitchen shots; the blue glow turns pinkish on the wood; sparse window lights on the train. People are stand-ins.































































































































































































































































































































































