McKinsey flags a thirtyfold cost gap in AI agents

On September 26, McKinsey warned that the cost of agentic AI can vary by as much as 30 times for the same intended task, because different agents can take very different routes to reach an outcome. The finding shifts attention away from a model’s advertised token price and towards the computing, tool use and supervision needed to complete actual work.

The warning comes as companies move beyond isolated chatbots and trials. Agentic systems can break down a request into several stages, call external tools or enterprise systems, review intermediate results, revise their approach and continue until they consider a task finished. Every added stage can trigger further model calls and consume additional tokens.

The cost is in the workflow, not only the token

This distinction matters because AI has become much cheaper on a per-token basis. Stanford’s 2025 AI Index found that the cost of querying a model with GPT-3.5-level performance on a standard benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024. OECD analysis published in July 2026 similarly estimated that quality-adjusted text-to-text model prices fell by nearly 80% between January 2024 and April 2026.

Lower unit prices, however, do not automatically lower an organisation’s bill. Coforge told a recent executive briefing that the unit cost of tokens needed for the same AI intelligence had fallen roughly 100-fold over seven months, while token consumption had grown about 8,000-fold. Those figures are the company’s observations, not an economy-wide measure, but they illustrate why cheaper inference can encourage much broader use.

• Measure the cost of a completed task rather than token consumption alone.
• Track model calls, tool use, computing resources, orchestration and human review.
• Compare agents that pursue the same business objective through different workflows.
• Reserve more capable and costly models for work that requires them.

Coding agents face particular scrutiny

McKinsey’s warning is particularly relevant to software development teams, where demand for agentic coding tools can be token-intensive. A 2026 working paper involving researchers from the University of Michigan, Stanford University, All Hands AI, Google DeepMind, Microsoft AI and MIT found that agentic coding tasks used around 1,000 times more tokens than the code-reasoning and code-chat tasks used for comparison. It also found up to a 30-fold difference in total token consumption across runs of the same task.

That research does not establish that every coding agent has the same cost profile, nor does it show that high token use is always wasteful. A longer workflow may produce a more useful result in some cases. But it does show why an apparently similar assignment can produce sharply different bills when an agent retries, explores alternatives, invokes tools or carries out longer reasoning chains.

Budgets are becoming a management issue

McKinsey’s 2026 AI survey found that about one-third of respondent organisations expected to invest more than 10% of their technology and communications budgets in AI, while 60% planned to increase AI spending in the following year. About one-fifth said AI expenditure was already putting pressure on operating costs. McKinsey executives described those expenses as increasingly tangible and visible.

A separate McKinsey Enterprise AI FinOps survey from May 2026, covering 75 qualified respondents across five industries, found that 93% reported exceeding their AI budgets. McKinsey also cited Menlo Ventures data indicating that enterprise spending on large language models roughly tripled in the 12 months to the end of 2025. These surveys describe respondents and market estimates, rather than a full census of businesses.

Productivity claims need a cost test

The attraction of agents remains clear: McKinsey has previously said agentic AI could reduce human time spent on some transformation-office tasks by 35% to 40%, and by more than 70% in certain cases. Yet time saved is not by itself a return-on-investment calculation. Companies still need to test whether the output is reliable enough, whether staff time is genuinely redeployed, and whether the value created exceeds the full cost of the workflow.

For technology leaders, the practical question is therefore less about finding the lowest published token price than about identifying the least expensive dependable route to a defined result. Firms can mix models, using smaller or open-weight options for some workloads and frontier systems for harder tasks. In the coming 12 months, McKinsey says measuring both the benefits and costs of agentic AI will become a central management challenge, including for European companies balancing productivity goals with tighter operational controls.