How to Manage AI Spend in the Agentic Era: A Founder's Playbook

Key takeaways

• OpenAI reports the price per million tokens fell 97% from GPT-4 to GPT-5.4, yet AI bills keep climbing because agents consume autonomously, not per seat.
• The metric OpenAI recommends is useful work per dollar: tasks completed, time saved, decisions improved, not credits burned.
• Cheapest per token is not cheapest per outcome: retries, corrections, and review time belong in the real cost.
• OpenAI's five steps (visibility, outcome ROI, governance, compounding workflows, capacity matching) map directly onto how founders should model AI COGS.
• If your customers' agents consume without human pacing, your costs scale with work done, not users, and your pricing has to follow the same curve.

Why do AI bills rise while token prices fall?

OpenAI's own numbers frame the paradox: token prices dropped 97% between GPT-4 and GPT-5.4, and GPT-5.6 reportedly completes coding tasks with 54% fewer output tokens and 57% less time per task on the Artificial Analysis Coding Agent Index. Cheaper tokens, more efficient models, and yet enterprise AI budgets keep growing.

The reason is the shift from chat to agents. A person typing prompts is a natural rate limiter. An agent running a long workflow is not. When consumption becomes autonomous, spend decouples from headcount, and a growing bill can mean waste, healthy experimentation, or a workflow quietly becoming business-critical. Without visibility you cannot tell which.

What is useful work per dollar?

It is OpenAI's proposed north star for AI budgeting: measure tasks completed, time saved, decisions improved, and workflows ready to scale, per dollar spent. The sharper version for priority workflows is cost per accepted outcome: everything it took to get a result you would actually ship (model usage, tools, retries, review) divided by accepted results.

This flips the usual procurement instinct. The lowest token price does not always produce the lowest total cost. A cheaper model may fail, retry, or create work that needs correction. A more capable model may cost more per token but reach an acceptable result faster with fewer attempts.

How do the five steps translate for a SaaS founder?

OpenAI wrote this for enterprise admins buying AI. If you are a founder selling AI features, the same five steps read differently:

1. 1Visibility into usage and spend. Their version: know who uses what. Your version: know your token consumption per customer, per feature, per tier. If you cannot attribute spend, you cannot price.
2. 2Evaluate efficiency by outcome ROI. Their version: evals that reflect real tasks. Your version: pick models on cost per accepted outcome for your top workloads, not the pricing page.
3. 3Govern advanced workflows before they scale. Their version: approvals and limits. Your version: rate limits, caps, and stopping conditions in your product before a power user discovers an infinite loop on your dime.
4. 4Fund workflows that compound. Their version: portfolio management. Your version: invest AI COGS in features where usage correlates with revenue, not just engagement.
5. 5Match capacity to proven demand. Their version: guaranteed capacity and batch tiers. Your version: move predictable workloads to batch or cached processing and keep premium latency for the moments users actually pay for.

What does this mean for your pricing?

Here is the observation OpenAI's post does not make: as a founder you sit on both sides of this equation. Your customer's useful work per dollar is your value metric. Your own cost per accepted outcome is your COGS. Your pricing is the spread between the two, and the agentic era squeezes it from both directions: buyers get better at measuring value, while autonomous consumption inflates your costs.

Efficiency gains like fewer output tokens per task land in that spread too. When a model update cuts your cost per task, you choose: keep it as margin, or pass it through as price. Companies that never modeled their unit economics do not even notice they had the choice.

The takeaway: budget AI like an enterprise, price it like a founder. If you want to see how agentic usage patterns, model choices, and tier limits play out in your margins, you can simulate the whole thing in Calcaas.