AI tokens got cheaper, but the deployment bill didn't

For enterprise technology leaders, falling AI prices sound like good news. In practice, the total bill often tells a different story.

Recently released models, such as GPT-5.6 and Claude Opus 5.5, can write and debug code, analyze documents and use tools to complete multi-step tasks, all at substantially lower token prices. A year ago, Claude Opus 4.1 cost $15 per million input tokens and $75 per million output tokens; Opus 5.5 costs $4 and $20 per million, respectively -- a reduction of approximately 73%. OpenAI cut the price of GPT-5.6 Luna, its fastest and most affordable model, by 80% on July 30, and the price of GPT-5.6 Terra by 20%, leaving Luna at $0.20 per million input tokens and $1.20 per million output tokens.

Yet, lower prices have not eliminated budget surprises. Uber reportedly exhausted its full-year AI budget by April as engineering teams accelerated adoption of AI coding tools. The price of a computation unit is falling, but the amount of computation behind a business request can grow considerably. The apparent contradiction becomes easier to understand when we look at what enterprises are asking AI to do now.

Cheaper tokens, more work behind every request

A simple chatbot interaction might involve sending a question and receiving an answer. An enterprise agent handling a business process may need to make repeated model calls, retrieve records, inspect results, revise its approach and verify completion.

That many steps change the economics. Consider an illustrative example: If the effective token price falls by 50%, but a workflow consumes four times as many tokens, its token cost doubles. If more employees run that workflow more frequently, spending multiplies across the entire head count.

The cost multiplier: From answering to executing

Consider a familiar business request: "Which completed jobs haven't been invoiced?" The sentence is short. Executing it may require the agent to do the following:

• Retrieve completed jobs from a field-service application.
• Identify matching customers across operational and billing records.
• Check whether invoices exist and whether they cover the completed work.
• Investigate inconsistent names, missing identifiers and partial invoices.
• Recover from failed calls or incomplete responses.
• Produce a verified list of exceptions.

If the request includes creating the missing invoices, then the workflow also needs approval, controlled writeback and confirmation that the transactions succeeded.

Each additional model call processes billable input and output tokens. Retries add calls, and records returned by tools can become input to later calls. Cached input is cheaper but still billed. Reasoning models also generate "thinking" tokens before they answer, and those are billed as output tokens -- the more expensive side of the price list -- so a harder question can cost far more than its short answer suggests.

Clio, a legal software company, reported 38% fewer prompt tokens with no quality loss in multi-step document analysis tests when code handled intermediate tool operations. Fewer tokens came from changing how the workflow ran.

Managing the cost of execution

Routing easy steps to smaller models, caching repeated context and using batch discounts all help manage costs. They can reduce token consumption or token charges, but they don't show how much a finished task costs, so CIOs must still measure the entire workflow.

Follow these steps to help manage costs:

1. Map the workflow before you price it

For example, accounts payable teams follow an invoice from receipt through purchase-order matching, delivery verification, any exceptions review and payment approval. Count the model calls and include what happens when records are missing. Use code to check totals and match purchase-order numbers, and use the model to interpret and explain discrepancies. Budget for the exception path and the clean record match.

2. Test the cost of a correct result

For customer returns, test requests involving split shipments, expired return windows, missing receipts and unavailable order records. Compare models on whether they apply the return policy correctly, what each completed case costs and how much employee correction it requires. Run cases repeatedly to expose expensive retries and repeat the tests when the model or return policy changes in the future.

3. Set limits that stop unproductive runs

For an enterprise coding agent, track spending per accepted change, including failed attempts. If the agent repeatedly runs the same failing test or cannot access a dependency, stop it and notify an engineer of the blocker. Set spending, retry and runtime limits before rollout. Measure changes that pass review and testing. Token consumption alone tells you how much activity you paid for.

The bottom line

CIOs must reset AI budgets around completed workflows now. The 2026 research preprint The Price Reversal Phenomenon: When Cheaper Reasoning Models Cost More tested eight models across 12 tasks. In 32% of model-pair comparisons, the lower-priced model incurred higher total inference costs. Listed token prices are not a reliable basis for a budget forecast.

Before expanding deployment, CIOs should measure cost per correctly completed task, including failed attempts, infrastructure and human review. Next, CIOs need to set spending and retry limits, and track completion rates and correction time against the existing process. Finally, CIOs can only scale when the results justify the expense.