OCTOPAI ENGINEERING ARTS · INTERNAL · CHASING THE COST QUESTION
If a failure costs what a success costs
Failed and passed SQLite attempts used a median 48 versus 47 core-seconds. The only honest denominator is accepted jobs.
A4 · INTERNAL · MEASURED · MEDIAN CORE-SECONDS
A5 · ZERO 429s ≠ NO MODEL WAITING
A6 · WHOLE-PILOT COST, NEVER PER JOB
SOURCE · PILOT #outcomes #cpuonly · FAILED vs PASSED ATTEMPTS
OCTOPAI ENGINEERING ARTS · INTERNAL · ONE MODEL, EVERY FIGURE
REFERENCE SELF-CHECKPASS
DRIFT → RENDERREFUSED
SOURCESoutputs.csv ONLY
COST RECEIPT · ONE ACCEPTED JOB · MODEL · PRICE ASSUMPTIONS ON SCREEN
FRONTIER-PRICED · C11
$0.0359
OWNED MARGINAL · C12
$0.0036
OWNED ALL-IN · C12
$0.0057
TOOLS, RETRIES AND VERIFICATION DO NOT CHANGE BETWEEN THE COLUMNS — THE TOKEN BILL IS NOT THE AGENT BILL
END-TO-END BLUEPRINT SCENARIO · MODEL · 500 SEATS · 4 AGENTS · 17 H/DAY · 3% ESCALATION
ALL-FRONTIER · C3
$419,451/mo
HYBRID ON-PREM · C4
$33,111/mo
92.1% BELOW ALL-FRONTIER · 4.4% BELOW ALL-CLOUD · 20:1 INPUT:OUTPUT · 90% CACHED AT 10% OF LIST · GPT-6 TERRA $1/$5
BREAK-EVEN, PAYBACK AND THE FOUR LEVERS · MODEL
PAYBACK · C7 / C8
1.6 / 33.1 mo
ESCALATION · C14
$24,920→$123,212
CACHE · C15
$1.17M→$344,498
$0.12/kWh · PUE 1.4 · 8 × 282 W MEASURED + 1.5 kW HOST · OPS $2,000/SERVER + $20/HALO BOX · HALO $3,999 × 1 PER 16 USERS · DO MI325X $3.80/GPU-HR · 36-MONTH AMORTISATION
Not in any film: what it cost us to make this.
One model, and everything else reads from it
C3–C15 · MODEL · ASSUMPTIONS ON SCREEN
C11/C12 QUANTITIES · A4 MEASURED
NO PRODUCTION SPEND SHOWN — BY RULE
SOURCE · shared/cost-model/outputs.csv · shared/COST-MODEL.md §2–§4