Operations · June 2, 2026 · 9 min read
When AI Usage Costs More Than the Work It Supports
Cost comparison requires workload context, not hype. At sustained volume, poorly routed model usage can rival or exceed the cost of the workflow it was meant to support.
The claim needs a boundary
It would be careless to say that AI usage universally costs more than employee salaries. Many useful workflows are inexpensive. Many others are not. The real issue appears when recurring agent loops combine large context, repeated premium-model calls, long outputs, tool-driven follow-ups, and weak visibility into usage.
Provider pricing is public and variable by model. OpenAI, Anthropic, and Google publish different rates for different model tiers, cached inputs, outputs, and related processing options.1 The same workflow can therefore have very different operating costs depending on routing.
What drives sustained usage cost
- Large context windows sent repeatedly instead of reused deliberately.
- Premium-only routing for simple helper work.
- Long agent loops that generate additional model calls after every tool result.
- Retries and verification passes without request-level inspection.
- High output volume, especially when a workflow produces verbose intermediate responses.
- Missing separation between deterministic operations and model reasoning.
Worked example: one premium route for everything
Assume an organization runs 500,000 agent turns per month. Each turn sends an average of 20,000 uncached input tokens and receives an average of 2,000 output tokens. That is 10 billion input tokens and 1 billion output tokens per month.
As of June 2, 2026, OpenAI lists GPT-5.5 standard processing below its stated long-context threshold at US$5 per million input tokens and US$30 per million output tokens.2 Using that listed rate only as an illustration:
- 10 billion input tokens: 10,000 million-token units x US$5 = US$50,000.
- 1 billion output tokens: 1,000 million-token units x US$30 = US$30,000.
- Illustrative monthly model cost: US$80,000.
If the supported workflow represented a hypothetical US$65,000 monthly payroll allocation, this premium-only usage pattern would exceed that payroll allocation. Change the volume, token shape, cache behavior, route, provider, or model and the answer changes. That is precisely the point: architecture changes the bill.
Worked example: route the routine work
Now assume that review identifies 60 percent of the input and 50 percent of the output as routine helper work suitable for a lighter, cheaper, or local route. The premium route remains available for complex reasoning, but it no longer receives every intermediate step. The exact savings depend on the selected routes and local operating costs. The decision is measurable because the workflow has been separated.
| Workflow decision | Premium-only pattern | Deliberate-routing pattern |
|---|---|---|
| Evidence collection | Raw evidence sent directly to the premium route | Deterministic tools collect evidence first |
| Evidence packaging | Premium model summarizes routine material | Suitable lighter or local route compresses routine material |
| Complex decisions | Premium route receives every task | Premium route remains available for depth and final synthesis |
| Repeated context | Resent without inspection | Reusable context is retained or cached where appropriate |
| Cost visibility | Monthly bill reviewed after the fact | Request-level usage reviewed while routing evolves |
The salary comparison is a warning, not a slogan
Human-workflow comparisons can be useful because they force teams to ask what the system is supporting. They become misleading when they are treated as universal proof. A customer-support classifier, a code-review agent, a research assistant, and an always-on multi-agent operations system have different token shapes, output requirements, and risk profiles.
The responsible conclusion is narrower: sustained AI usage can become a major operating expense. In some workloads it can rival or exceed the cost of the supported work. Teams should measure the actual workflow before that happens, not after the invoice turns into a surprise.
Reduce waste without reducing quality
Cost discipline is not a race to the weakest model. Use premium models for high-value reasoning. Use smaller or local models where the task permits it. Reuse stable context where provider caching fits. Let deterministic tools handle mechanical operations. Inspect request-level usage so the routing policy can improve with evidence.
Continue with How Kaptain Reduces Unnecessary Token Spend or the model-routing guide.
Sources
- OpenAI API Pricing; Anthropic Pricing; Gemini Developer API Pricing. Accessed June 2, 2026.
- OpenAI API Pricing, GPT-5.5 standard processing listing. Accessed June 2, 2026.