c16
c16 / wf5.m2 - Agent Control Plane — hi-fi unconstrained (wf5.m2) / Cost Center·control-plane-v1/cost-center·draft

Finance · Cost Center

What every AI agent costs — governed, at-risk, and re-routable

$684.50 spent so far this cycle (11% estimated). At the recent ~$115.57/day run-rate this month projects to $3.5k16% over the $3.0k budget. $290.50 (42%) flows through at-risk agents; $122.10 is ungoverned and unchargeable. Top driver: Claude Code at $214.50.

Total spend

$684.50incl. est.

the single trusted number · 11% estimated

Projected this month

$3.5k

16% over $3.0k budget

Cost per session

$11.22/session

$684.50 ÷ 61 sessions

From cache

54%

cache-read tokens · cost saved

Budget vs. actual — forecast

$3.0k monthly budget · proposed: BudgetForecastMeter

At ~$115.57/day you cross budget in ~20 days.

budget $3.0k
proj $3.5k
Spent so far $684.50
Forecast addition +$2.8k
Budget ceiling $3.0k

Spend you can't charge back or control

at risk

proposed: GovernanceSpendBar · governanceTaggedSpend()

$290.50 / $684.50

42% of spend flows through agents with posture < 80 or no policy binding.

Governed · healthy$394.00 · 58%
At-risk (posture < 80)$168.40 · 25%
Ungoverned$122.10 · 18%

$122.10 ungoverned — no policy binding, no chargeback owner.

Spend by agent

$684.50 total · proposed: SpendTreemap

tile area = share of spend · color = governance posture · click to inspect the agent

governed · healthy at-risk (posture < 80) ungoverned / critical

Spend over time

daily · last 7d

Run-rate has climbed from $88.20 to $142.60/day — the slope driving the $3.5k projection.

Model efficiency — $ per 1k tokens

runtime-selection lever · proposed: EfficiencyLollipop

re-route latency-tolerant, high-volume work to the cheaper-per-token runtime

gpt-5-codex est
$0.032/1k
claude-opus-4-8
$0.032/1k
gpt-4o
$0.032/1k
claude-sonnet-4-6
$0.031/1k
claude-haiku-4-5
$0.006/1k

gpt-5-codex costs $0.032/1k — the cheapest (claude-haiku-4-5 at $0.006/1k).

Dollars only tell half the story

$684.50 bought 61 sessions — but how many succeeded? Cost per successful task exposes spend that produced nothing. See the join in Effectiveness.

By agent

7 agents · $290.50 at-risk
Agent
Governance
Risk
Tokens
Responses
Cost
Claude CodegovernedStandard6.9M1.2k$214.50
support-copilotgovernedElevated5.6M3.1k$168.40
CursorgovernedStandard3.1M980$96.20
Codex CLI
est
ungovernedElevated2.4M540$78.00
data-pipeline-agentgovernedStandard9.2M5.4k$52.30
marketing-gptungovernedCritical1.1M860$44.10
Claude DesktopgovernedStandard800.0k180$31.00

Each agent's spend joins back to its governance posture — so dollars carry the risk lens. Rows link to the agent profile. $122.10 of this column is ungoverned (highlighted) and has no chargeback owner.

By model

5 models
Model
Tokens
Cost
$/1k
claude-opus-4-8 7.7M$245.50$0.032
gpt-4o 6.7M$212.50$0.032
claude-sonnet-4-6 3.1M$96.20$0.031
gpt-5-codex
est
2.4M$78.00$0.032
claude-haiku-4-5 9.2M$52.30$0.006

By user

6 identities
User
Tokens
Responses
Cost
mark@qpoint.io 7.7M1.4k$245.50
svc-support@qpoint.io 5.6M3.1k$168.40
jane@qpoint.io 3.1M980$96.20
dev-7f3@duck.com
est
2.4M540$78.00
svc-data@qpoint.io 9.2M5.4k$52.30
(proxy-only, no identity) 1.1M860$44.10
est

est rows are estimated from token counts — some providers (OpenAI / Codex) don't report a billed cost, so we derive it from usage. This is usage-based observation, not a billed invoice. For chargeback, attribute spend per cost center in Findings.

Do next