When AI Gets Cheaper but Costs More

When AI Gets Cheaper but Costs More

Why cheaper models will not necessarily mean lower enterprise bills

Agentic AI is currently breaking seat-based economics. Even as token prices fall, total enterprise AI bills are rising because machine-paced workflows consume more calls, context, tools, reasoning, and retries. The enterprises that succeed will manage the quality-adjusted cost per accepted outcome, not cost per token.

In all our client conversations in the past few months, a concern has surfaced in almost the same words: “Our AI bills are rising faster than expected. What should we do?” The question has come from technology leaders, finance teams, and business executives alike. It is becoming the first real operating test of enterprise AI maturity.

The concern is not that AI has stopped becoming cheaper, but the fact that cheaper intelligence is being consumed in much larger quantities. A single employee request can now trigger a chain of planning, retrieval, model calls, tool use, validation, and retries. The price of each step may fall while the total cost of the workflow rises.

For instance, Uber offers a useful warning. According to its executives, the company exhausted its 2026 AI budget in four months as adoption of coding agents expanded across its engineering organization. Whether every enterprise reaches that level is not the point. The case shows how quickly a fixed annual budget can become obsolete when usage shifts from human-paced assistance to machine-paced work.

1. The First Era of AI Was Priced for Humans

Generative AI entered the enterprise largely as a productivity tool sold by the seat. That model worked because human attention imposed a natural limit on consumption. An employee could ask only so many questions, review only so many answers, and run only so many sessions in a day.

Agentic AI changes the underlying economics. A single instruction can now start a workflow that plans, retrieves documents, calls several models, invokes external tools, tests its own output, and retries failed steps. The consumer is no longer only the employee. It is also the software acting on the employee’s behalf, at machine speed and with far less natural friction.

Recent changes reflect this shift. GitHub moved Copilot toward AI Credits in June 2026, explicitly distinguishing between a quick interaction and a long-running agentic coding session. Anthropic introduced paid usage credits that allow subscribers to continue beyond included limits at standard API rates, with organization- and user-level controls. These are not isolated price changes. They are signs that flat-rate access is giving way to hybrid economics: subscriptions for baseline use, metering for heavier consumption, and controls for enterprise buyers.

Public-market scrutiny may accelerate the transition, but the structural driver is more basic: a fixed seat price does not map cleanly to machine-paced work. The relevant management question is therefore changing from “Which model has the lowest published rate?” to “What does a trusted business outcome cost once every model call, tool call, retry, and review step is counted?”

2. Cheaper Tokens Can Still Produce Higher Bills

At first glance, sustained cost pressure appears unlikely. Stanford’s AI Index found that the cost of querying a model with GPT-3.5-equivalent performance fell from about $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024, a decline of more than 280 times. More recent research continues to show rapid improvement in capability per dollar.

The paradox is that enterprises do not purchase benchmark-equivalent tokens in isolation; they purchase business workflows. As AI workflows become increasingly autonomous, overall consumption is driven less by the price of an individual model call and more by the number of planning steps, retrieval operations, tool invocations, validation cycles, and retries required to complete the task

Named enterprise cases demonstrate this pattern. Uber reportedly responded to rapid coding-agent adoption with per-employee spend caps. AT&T has described using a portfolio of smaller, fine-tuned models while a larger model controls more complex cases, materially reducing API costs. GitHub’s shift to credit-based metering reflects the same underlying reality at the provider level: a multi-hour autonomous session cannot be priced as if it were a single chat request.

The cost base also extends beyond tokens. Search, code execution, retrieval, vector databases, observability, orchestration, regional processing, and human review all add to the bill. Failed runs consume resources without creating value. Low-quality outputs create rework that may never appear on a provider invoice but still belongs in the enterprise cost model.

3. Orchestration Becomes the Economic Control Point

As model capabilities converge for routine tasks, enterprise advantage moves from access to orchestration. The control layer decides whether a model is needed, what context to send, which model should respond, how many steps an agent may take, when it should stop, and when a person must intervene. The economic control point is not the model; it is the workflow’s permission to consume.

1. Eliminate unnecessary model work. Use deterministic software, rules, search, and conventional automation where generation or reasoning is not required. The cheapest model call is the one the system never makes.
2. Reuse work already performed. Exact caching, semantic caching, and provider prompt caching can prevent repeated processing of the same question, instruction, document, or conversation prefix. In a documented production deployment, semantic caching lifted the cache hit rate from 18% to 67% and cut monthly LLM spend by roughly 73%.
3. Route tasks to the right model. Use smaller or lower-cost models for routine extraction, classification, and summarization, escalating only when complexity, uncertainty, or risk justifies it.
4. Constrain agent behavior. Set explicit limits on steps, tool calls, retries, context growth, execution time, and reasoning effort. Unbounded autonomy creates unbounded cost variance.
5. Govern at the outcome level. Attribute model, tool, infrastructure, review, rework, and failure costs to the workflow and business owner that generated them. Use AI observability capabilities such as runtime tracing, token attribution, execution analytics, and workflow-level cost monitoring to identify cost drivers, benchmark workload efficiency, and continuously optimize AI cost and performance.

An example of “Reuse” is caching rather than recomputing. This is the single highest-impact move available today, as it removes queries from the meter entirely. Across large user populations, roughly a third of LLM queries are semantically similar to questions already answered; for instance, a helpdesk fields “How do I install the VPN app?” and “Guide me through company VPN setup” as different questions and pays full price for near-identical answers each time. Semantic caching converts each query into a vector embedding, checks a similarity index of past questions, and serves the stored answer instantly when a close match exists; as a result, the query never reaches the model. A Japanese firm is currently using this process to rationalize its AI costs.

4. What Buyers Should Demand and What Businesses Should Measure

Buyer demands and financial measurement are one argument, not two. Enterprises need provider economics that can be translated into business economics. Contracts and management consoles should expose usage by model, team, workflow, and user; support pooled allowances and granular budgets; and provide advance notice of changes to rates, models, cache terms, and included credits.

Efficiency features, including prompt caching, batch processing, smaller-model tiers, routing support, and exportable usage data, should be part of the core product. Model portability also matters because it creates negotiating leverage and lets teams adopt a better cost-quality option without rebuilding the application. The goal is not constant provider switching. It is credible freedom to switch when economics or risk requirements change.

The enterprise metric should be the quality-adjusted cost per accepted outcome, not cost per token or even cost per task. A task can be cheap and still fail quality checks, create compliance exceptions, or shift rework downstream. The numerator should include model, data, tools, infrastructure, review, rework, and failed runs. The denominator should count only outcomes accepted at the required quality, risk, and service level.

The supporting dashboard should track premium-model share, average agent steps and tool calls, cache and reuse rate, human-review and rework rate, and cost variance by workflow. These measures connect architecture decisions to financial results and allow comparison with the human or conventional-automation baseline on the same terms.

The next phase of enterprise AI will be shaped by two truths:

1. Intelligence per unit will continue to become cheaper, and machine consumption will continue to expand. Enterprises that treat the first trend as a guarantee of lower bills will be surprised by the second. Those who engineer the full workflow for the required quality at the lowest total cost will scale without losing financial control
2. The winners in the AI race will not be the companies that use the fewest tokens. They will be the companies that spend the least to produce a trusted business outcome.

By Akshay Khanna, Managing Partner, and Sahil Chaudhary, Associate Research Director, Avasant