How does AgentOS track cost?

Every task carries a cost budget as part of its execution contract. Spend is tracked per task at sub-penny precision and rolled up to the unit of work, the work order, the phase, the role, the model. Cost is a property of the task.

Why does it get cheaper over time?

Three loops: frontier spend trends down as memory absorbs explanations and local models handle a growing share of routine work; review collapses to minutes because work arrives with its own evidence packet; and the institutional record compounds, so tomorrow's agents inherit today's lessons without retraining.

What is model routing?

AgentOS routes tasks to models by task class and cost class.

Economics · The Cost of the Work

The savings ledger

Measured on our own estate, from the graph.

Recorded numbers from our production work graph. Every figure is computed from per-call cost records and the memory ledger, reproducible by query.

LEDGER · CACHE DISCIPLINE

$10,603 of work for $1,990.

94.5% of input tokens were served from cache at a tenth of the price. Computed per call from each call's own pinned rates: the identical work, uncached, would have cost 5.3 times more. Cache discipline returned about $8.6K.

$10,603 uncached→$1,990 actual

LEDGER · THE ALL-IN NUMBER

A full delivery arc for $1,990.

Planning, architecture, development, QA, adversarial review, and governance: 11,193 model calls, 3.20B tokens, and 36,515 governed transitions carrying the work from authoring to close. All-in, with every dollar attributed to the task that spent it.

36,515 governed steps→5.5¢ per step

LEDGER · RETRIEVAL LEVERAGE

5.7× fewer tokens per code question.

Semantic search over the AST index answers a code question in about 5.1K tokens where manual file exploration averaged 29.1K, measured across 536 recorded searches. Workers ask the graph.

29.1K exploring→5.1K retrieved

production ledger · a real work order, attributed to the penny

priced model calls11,193

tokens3.197B

actual spend$1,990.52

uncached equivalent$10,603

cache hit rate94.5%

Work phase	Primary model	Tokens	Cache reuse	Cost
QA	Codex GPT-5.5	201M	92%	~$227
Development	Claude Sonnet 4.6	450M	97%	~$210
Orchestration	Claude Opus 4.8	208M	97%	~$143
Review	Codex GPT-5.5	38M	89%	~$55
Architecture	Claude Opus 4.8	53M	96%	~$46
Gatekeeping	Codex GPT-5.5	22M	91%	~$36
Planning	Claude Opus 4.8	26M	97%	~$22

Every dollar traces to the phase, role, model, and action that spent it, where most platforms can only report a monthly total.

Recorded numbers from the production work graph · anonymized estate · reproduction queries in the evidence library

The compounding insight

Cheaper and smarter the longer it runs.

Three loops produce the ledger above, bending the cost curve down and the capability curve up at the same time. None of them require retraining.

LOOP · INFERENCE COST

Frontier spend trends down.

Every expensive explanation is captured once and reused forever. As the memory layer fills, local models absorb a growing share of routine work, and frontier models get reserved for high-leverage reasoning.

all frontier→frontier only when needed

LOOP · DEVELOPER LEVERAGE

Review collapses to minutes.

Work arrives with its own evidence packet. Reviewers verify the gates and spot-check the diff instead of re-reading every line, so throughput per engineer compounds.

days re-reading→minutes to verify

LOOP · INSTITUTIONAL INTELLIGENCE

The system gets smarter, no retraining.

Decisions, approved patterns, and failure modes accumulate. Tomorrow's agents inherit today's lessons, and that knowledge survives engineer turnover.

each agent blind→each agent grounded

Per unit of work

Most platforms report monthly AI spend. AgentOS reports the cost of the work.

Spend is a property of a task. Every dollar is attributed to the unit of work that spent it, all the way up the work graph.

Cost per work order Cost per role Cost per phase Cost per model Cost per action

Planning

Architecture

Development

QA

Governance

Routing by task class

The right model for the right task.

Cheap local models handle the high-volume, class-bounded work. Frontier models handle the high-leverage, decision-class reasoning. AgentOS routes by task class and cost class.

Local lane

Bulk, grounded, class-bounded.

Most of the work is routine: grounding lookups, embeddings, classification, summarization, formatting. Local models on your own hardware do this work at near-zero marginal cost.

Inference Fabric: 22× tokens per NVIDIA GPU
Bulk grounding · embedding · classification
Class-bounded execution

Frontier lane

Reserved for decision-class work.

When a task requires open reasoning, a frontier model is the right tool. AgentOS reserves the lane for that, so spend goes where leverage is.

Architecture, design, ambiguous QA
Adversarial review and adjudication
Frontier-model share tracked as a KPI

Cost discipline at the boundary

Budgets enforced where the work happens.

The contract that authorizes a task also bounds it. Budget is a required field, fail-closed when exhausted, and visible in real time.

Sub-penny precision

Spend tracked per task.

Token usage and model rates roll up per task in near-real time. No batch reconciliation, no surprise overruns at month-end.

Fail-closed at the gate

Budget exhausted means stop.

A task that runs out of budget halts at the boundary and surfaces. It does not silently borrow from the next task, the next phase, or the next month.

Frontier-model KPI

Cheap by default, expensive on purpose.

Frontier share is a tracked number. If it climbs, something is escaping the routing rules, and the signal fires the same week.

The cost of the work.

Get AgentOS and put AgentOS economics on your real work.

Get AgentOS Back to AgentOS