Coding Agents More Expensive Than Programmers: Gartner's Prediction for 2028

By 2028, the cost of tokens consumed by a single developer to power coding agents could exceed what that same developer earns in a month. This is the prediction from Gartner, which sets the benchmark at a global average monthly salary of around $2,000.

The reference is to the global average, not the six-figure annual salaries typical of U.S. developers, which remain well beyond the reach of token costs. The driving force behind this escalation, according to Gartner, is the shift by vendors from a seat-based licensing model to a consumption-based model: you pay for what you use, and the use of agents is growing rapidly.

From $100 to $5,000

Monthly costs per developer are shifting from a range of $20-100 to a range of $2,000-5,000, with extreme cases reaching up to $20,000 per month. These numbers shift the AI agent spending from the tools category to the much heavier personnel category.

Data collected by Gartner Peer Insights provide a measure of where we are today: 23% of technical leads spend between $200 and $500 per month per developer on tokens for agents such as Claude Code, Cursor, and OpenAI Codex, while 6% exceed $2,000 monthly.

The critical point is that spending does not automatically translate into work done. Nitish Tyagi, senior principal analyst at Gartner, is clear: "There is no direct relationship between increased token consumption and increased productivity gains." Yet vendors, Gartner notes, have not yet made mature cost optimization features available, and they tend to encourage tokenmaxxing, the idea that more tokens equal more productivity.

Where the Overtaking Has Already Happened

In some markets, the prediction has already become reality. In India, the cost of tokens for coding now exceeds the salary of a developer with four to six years of experience. The problem is exacerbated because the price of tokens does not vary based on the geographical location of the consumer: the same amount weighs very differently on a paycheck in Bangalore compared to one in Silicon Valley.

The countermeasures suggested by Gartner go toward spending management rather than outright cuts. Practical recommendations include introducing consumption thresholds, automated token monitoring, context engineering to reduce the context sent to models, and model routing: small models for simple and recurring tasks, frontier models only for complex, high-value work. Additionally, tasks can be classified into three operational regimes, from those developer-driven to those fully entrusted to the agent, choosing the degree of autonomy based on complexity.

The way to measure results is also changing. The old metric of lines of code written loses meaning when an agent can generate an entire Python library almost instantly. Tyagi estimates that assisted basic development, without autonomous agents, can already lead to a productivity gain of up to 20%. And a high cost is not in itself a waste: he cites the case of a developer at an Indian IT firm who was burning $20,000 in tokens per month, until internal analysis showed that they were working on modernizing legacy systems, justifying the expense.

The condition that Tyagi sets as decisive remains: "Without a governed engineering operating model, costs can grow faster than the productivity gains that these tools promise to deliver."