Published on the 08/07/2026 | Written by Heather Wright

Token bills threaten productivity promise…

The economics of AI-assisted software development are heading for a shakeup with Gartner predicting AI coding costs will surpass the average developer salary by 2028 as token consumption rises and pricing models shift from seat-based to usage-based billing.

The prediction reflects one of the less discussed consequences of the generative AI boom. While early conversations focused on productivity gains, code generation speeds and developer efficiency, software engineering leaders (and indeed, all IT leaders) are now facing a new challenge: Controlling the cost of AI itself.

“Token discipline will not emerge through developer choice alone.”

At the same time, vendors are increasingly moving away from traditional seat-based licensing models towards consumption-based pricing, where organisations pay for the tokens consumed by AI models. As AI deployments scale, token consumption is emerging as a growing cost consideration across the industry. The combination of higher usage and metered billing is creating a cost trajectory that Gartner believes many organisations are underestimating.

On the coding front, Gartner says 40 percent of software engineering leaders are reporting that more than half of their teams are using AI tools to augment software development processes. As adoption grows, token consumption – the units increasingly used to measure and bill AI model usage – is rising rapidly.

Early signs of the cost impact are already emerging with Gartner Peer Insights data showing 23 percent of software engineering leaders are reporting costs of US$200 to US$500 per developer per month in token costs for AI coding agents such as Claude Code, Cursor and OpenAI Codex. For five percent of organisations, that figure has already soared to more than US$2,000 per developer per month.

“Organisations are rapidly moving from experimentation to scaled deployment of AI coding agents, but many are underestimating the financial impact of rising token consumption,” Nitish Tyagi, Gartner senior principal advisor, says.

““AI coding costs will continue to rise as infrastructure investment and profitability challenges push model pricing higher,” Tyagi says. “At the same time, as more developers adopt AI tools, light users are expected to rapidly become mainstream users as familiarity and reliance increase, driving further growth in token consumption and overall spend.”

The concern isn’t simply the price of large language models. Gartner argues that three common behaviours are driving much of the overspending: Giving AI agents too much autonomy, providing excessive amounts of context, and failing to establish feedback mechanisms to identify and prevent wasteful usage patterns.

“Token discipline will not emerge through developer choice alone,” Tyagi says, noting that developers will naturally optimise for speed and convenience, rather than token efficiency. Instead, organisations should treat token consumption as a governed part of their engineering operating model, the analyst company says in its How to Optimise Token Consumption for AI Coding Agents report.

That recommendation may be particularly relevant as organisations race to adopt increasingly capable AI coding agents. Unlike traditional coding assistants that suggest code snippets or provide developer support, agentic platforms can take on broader development tasks autonomously. Gartner points to increases of 100x over coding assistants experienced by software engineering leaders.  

Gartner says organisations need to establish clear rules around when AI agents should be used and how much autonomy they should be granted.

The report recommends categorising software development activities into three tiers: Developer-only work for sensitive, operationally risky work or work not worth the token cost; AI-assisted development where humans remain in control, and fully autonomous agent-led development. AI-assisted development should be the default zone for most day-to-day engineering because it improves throughput while maintaining human oversight and avoiding some of the introducing costly agentic loops which can occur when autonomous systems repeatedly call models and tools without adequate controls, Gartner says. Those with lower software engineering maturity and limited governance frameworks are also advised to prioritise assistive AI over autonomous development.

Governance is another area Gartner says requires urgent attention.

Less than one in three engineering organisations currently has formal AI governance policies in place. Gartner says engineering leaders should assess both the maturity of their software development practices and existing AI controls before expanding the use of autonomous coding agents.

“Organisations should introduce mechanisms such as token thresholds, escalation policies and automated monitoring to manage growth. Embedding these controls into engineering workflows ensures consistency and prevents uncontrolled costs.”

The report also highlights what Gartner calls context engineering as a critical discipline for controlling costs.

Every file, document, code repository extract or tool output included in an AI prompt increases token consumption. Gartner advises development teams to treat context as both a quality input and a cost driver, requiring developers to evaluate what information is essential, what can be summarised and what can be removed altogether.

“A mature context engineering practice should teach developers how to structure context so that it is not merely comprehensive, but optimised.”

Another recommendation is to avoid defaulting to premium AI models for every task. Instead, Gartner advocates model-routing approaches where smaller, lower-cost models handle routine work, with tasks escalated to more powerful and expensive models only when complexity or risk justifies the additional spending.

“AI coding agents are most cost‑effective when work is broken into smaller tasks that can be handled by smaller models, with escalation only when complexity demands it. Model choice should be driven by task complexity, context size, application criticality, and developer maturity.”

Also recommended is better visibility into how AI coding tools are being consumed, with regular reviews of high-token-consuming workflows as part of sprint retrospectives to identify inefficiencies, refine practices and promote knowledge sharing across engineering teams.

“Most organisations still lack the maturity and frameworks to effectively measure cost versus business impact,” says Tyagi.

The report stresses that higher costs aren’t necessarily bad if they bring measurable business benefits, such as profit, reduced risk, improved growth or support strategic goals.

“The goal is to use AI coding agents efficiently but not to restrict the use of them.”