The promise of AI-assisted software development has been framed largely in terms of productivity gains—faster coding cycles, reduced manual effort, and the ability to scale engineering output with fewer resources. But a new prediction from Gartner suggests a less discussed reality is quickly taking shape: the economics of AI coding may soon rival, and even exceed, the cost of human developers themselves.
According to the firm, by 2028, the cost of using AI coding tools—driven primarily by token consumption in large language models (LLMs)—is expected to surpass the average salary of a developer. This shift signals a fundamental recalibration in how enterprises evaluate the return on investment from generative AI in software engineering.
At the centre of this transformation is a deceptively simple concept: tokens. These units of data, processed by AI models for every prompt, response, and iteration, are fast becoming the new currency of software development. As organisations move from experimentation to scaled deployment of AI coding agents, token usage is exploding—often without the governance structures needed to keep costs in check.
“Organisations are rapidly moving from experimentation to scaled deployment of AI coding agents, but many are underestimating the financial impact of rising token consumption,” said Nitish Tyagi, Sr. Principal Analyst at Gartner. His observation highlights a growing disconnect between perceived productivity gains and the underlying cost mechanics driving AI adoption.
From seats to consumption: a pricing model shift with consequences
One of the most significant contributors to this cost surge is the industry-wide transition from seat-based licensing to consumption-based pricing. Unlike traditional software models—where costs are predictable and tied to the number of users—AI coding tools now charge based on usage, measured in tokens.
This shift introduces a level of variability that many enterprises are unprepared for. Without clear visibility into how tokens are consumed across different development tasks, forecasting costs becomes a challenge. Even more concerning is the lack of transparency from vendors on how token usage is calculated and billed.
The result is a growing sense of unease among software engineering leaders. Budgets allocated for AI initiatives are often exhausted sooner than expected, and the link between spending and tangible business outcomes remains difficult to quantify.
The governance gap: where costs spiral
While pricing models play a critical role, the way AI coding tools are used within organisations is equally responsible for escalating costs. In many cases, developers are given significant autonomy in how they interact with AI agents—optimising for speed and convenience rather than efficiency.
This behaviour, while understandable, leads to several inefficiencies. Large, unstructured context windows increase token consumption. Repetitive or poorly optimised prompts add unnecessary overhead. And the absence of structured feedback loops means there is little opportunity to refine usage patterns over time.
Compounding the issue is the relative immaturity of cost optimisation features within AI coding tools themselves. Vendors are still evolving their platforms, and built-in mechanisms to control or reduce token usage remain limited.
The paradox of adoption: efficiency drives more usage
Ironically, the very success of AI coding tools in improving developer productivity is contributing to rising costs. As developers become more familiar with these tools, their usage intensifies. What begins as occasional assistance quickly becomes an integral part of the development workflow.
Light users transition into heavy users, and reliance on AI grows. Each interaction, each iteration, and each refinement consumes tokens—incrementally driving up overall spend.
At the same time, broader industry dynamics are pushing costs higher. Increased infrastructure investments and the ongoing challenge of achieving profitability in AI model development are expected to keep pricing under upward pressure.
Rethinking the operating model for AI-driven development
To navigate this evolving landscape, Gartner emphasises the need for a disciplined and structured approach to AI adoption in software engineering. The focus is shifting from unrestricted usage to governed efficiency.
A key recommendation is the establishment of a use-case-driven framework that clearly defines when and how AI coding agents should be used. Not every task requires full automation, and organisations must distinguish between developer-led, hybrid, and fully agent-driven workflows.
Equally important is aligning model selection with task complexity. Smaller, less resource-intensive models can handle routine, high-frequency tasks more cost-effectively, while advanced models should be reserved for complex, high-value work. This approach, often referred to as intelligent model routing, can significantly optimise token usage.
Another critical lever is context engineering—the practice of refining inputs to AI systems to include only relevant information. By reducing unnecessary data and summarising inputs effectively, developers can minimise token consumption without compromising output quality.
Finally, governance must be embedded into the development lifecycle itself. Token thresholds, monitoring systems, and regular usage reviews should become standard practice, ensuring that cost control is not an afterthought but an integral part of engineering workflows.
The road ahead: balancing innovation with economics
The rise of AI coding agents marks a pivotal moment in the evolution of software engineering. But as the technology matures, so too must the frameworks that support its adoption.
The narrative is no longer just about what AI can do—it is increasingly about what it costs to do it. As token consumption becomes a defining factor in AI economics, organisations will need to strike a careful balance between innovation and financial discipline.
If Gartner’s prediction holds true, the industry may soon face a striking reality: the cost of augmenting developers with AI could rival the cost of employing them. And in that moment, the question will no longer be whether to use AI in coding—but how to use it wisely.