Thread 4374bdabOct 9, 2026 · 00:47 to 04:04 UTC+71 Fable 5.1 orchestrator + 19 Opus 5.5 sessions
Total cost, API list price
$118.35
Opus 5.5 subagents$92.7978%
Fable 5.1 orchestrator$25.5622%
≈ $36 per hour over 3 h 17 min
253.8M
tokens processed
97.8% were cache reads
1,372
API requests
20 sessions · avg $0.086 each
$1,259
cost without caching
caching cut the bill about 10x
511K
peak context
main thread, last turn
Timeline
When each session ran. The 7 build agents ran side by side from 01:25 to 01:47. Hover a bar for details.
Fable 5.1 orchestratorOpus 5.5 sessionPaused
Cost per session
Grouped by phase, in time order. Uncached input is under $0.05 in total, so it is folded into output. Hover a bar for token counts.
Cache readCache writeOutput (incl. thinking)
SessionModel · effortCost
Where the money went
Share of total cost by token type.
43.5%34.5%22%
Token type
Tokens
Fable $
Opus $
Total $
Cache read
248.3M
9.14
42.35
51.48
Cache write, 1h
4.30M
10.08
30.36
40.44
Cache write, 5m
70.9K
0.00
0.35
0.35
Output
1.11M
6.31
19.72
26.02
Uncached input
5,954
0.03
0.01
0.04
Total
253.8M
25.56
92.79
118.35
Rates per MTok. Fable 5.1: $10 in, $50 out, $20 1h write, $0.25 read. Opus 5.5: $4 in, $20 out, $8 1h write, $5 5m write, $0.20 read. Thinking was 278K of the 1.11M output tokens, about $6.40.
Main thread, turn by turn
Cost of each Fable turn. Your messages cost $13.82 over 10 turns. Wake-ups after a delegated task finished cost $11.73 over 17 turns.
Your messageTask-done wake-up
00:4704:04 UTC+7
A wake-up that only starts the next task still costs $0.45 to $0.65. Each request re-reads the whole context, which grew from 140K to 511K tokens.
Session detail
Times in UTC+7. Output shows thinking tokens in grey. Peak ctx is the largest single request. Click a column to sort.
Key findings
What stands out, and what would change the bill next time.
01Caching cut the bill about 10xThe same 253.8M tokens at full input price would cost $1,259 instead of $118.35.
02Subagents did 78% of the spendThe Fable orchestrator wrote only 126K output tokens. Opus 5.5 did the building at 40% of Fable's per-token price.
03The parallel build was the biggest phase7 Opus agents ran side by side from 01:25 to 01:47 and cost $38.05, which is 32% of the total.
041-hour cache TTL cost about $11.40 extra on subagentsSubagents sent requests seconds apart, so the 5-minute TTL would have worked: 3.79M write tokens at $8 instead of $5 per MTok. The main thread did need the 1-hour cache, with 10 gaps longer than 5 minutes.
05Review round 1 was the most read-heavyIt re-read a ~390K context 111 times (29.4M cache reads) to write 55K output tokens. It cost $9.96, almost as much as the Foundation build.
06Wake-ups were 46% of the Fable cost17 task-done wake-ups cost $11.73. Fewer, larger delegated tasks would mean fewer wake-ups on a 500K context.