Trailing 30 days · 26.9M API calls ·
69.6B tokens · synthetic data
Cost / successful task
$0.0112
the business denominator
Task success
96.1%
guardrail — must hold
Human intervention
1.2%
guardrail — must hold
Cost / 1K tokens
$0.0042
driver, not outcome
Revenue / 1K tokens
$0.0450
Cost / inference
$0.0108
GM after AI cost
90.7%
trailing 30 days
GPU utilization
87%
weekend: 64%
DAU / MAU
27%
14,595 avg weekday DAU / 54,845 MAU
trailing 30 days
Committed drawdown
63%
of usage revenue
F1 — Cost per successful task, daily (router deploy on day 45)
Routing is an engineering-product decision;
finance's job is to put its economics and its quality guardrails on the
same page. The router cut cost per successful task
41% at
identical customer pricing — and the trade is only good because the
guardrails held: task success 95.8% →
96.1%, human intervention
1.2% → 1.2%. A cost
drop with degraded quality is a regression wearing a savings costume, and
this page's validation fails it by design. GM after AI cost moved
85.8% → 90.3% across the
pre/post windows (the 90.7% headline covers only the
trailing 30 days — a shorter, later window, not a different calculation).
F2 — GPU utilization, daily (weekends shaded)
The reserved cluster bills 24/7; weekend demand halves.
Idle weekend capacity ≈ $145,061/yr
— a batch-workload / training-backfill candidate, not a cost cut.
F3 — p95 latency vs GPU utilization (each dot one day)
Above ~80% utilization, p95 degrades sharply
(1.5s hot vs 1.0s cool). This is
the cost-latency frontier: running the cluster hotter is cheaper per token
and slower per user, and the tradeoff should be priced, not discovered.