Gartner just put a number on something every engineering leader already feels in their gut. By 2028, AI coding costs will pass the average developer salary, pushed up by rising token consumption and the shift to consumption based pricing.[1]
The tool your developer uses is on track to cost more than the developer.
Here is the part most people will not say plainly. No developer is going to stop that. Engineers do not care what they spend. They never have and they never will. I’m not taking a shot at them, it’s just true, and once you accept it, everything about how you manage AI cost changes.
Why doesn’t cost-conscious training work?
Right now most of the industry is treating this as a discipline problem. The thinking goes that if we train developers to be more cost conscious, ship a few dashboards, and remind them often enough, they will start making cheaper choices.
That’s the wrong target. Engineers are paid to ship. Ask one what their token spend was last week and you will get a shrug, and they're not wrong to shrug, it isn't their job.
Why developers have no incentive to cut AI costs
Gartner says the same thing in cleaner language. Developers optimize for speed and convenience over cost efficiency, and token discipline will not emerge through developer choice alone.[1]
Asking an engineer to police token usage is asking them to work against the exact thing they were hired to do. Ship fast and move the product forward. Nobody gets promoted for using fewer tokens.
None of this means engineers don't care about efficiency, plenty do. It means efficiency isn't the lever they're incentivized to pull, and a training deck doesn't change an incentive.
Who actually owns AI cost management?
The people who care about spend are the people developers work for. Finance, engineering leaders, whoever signs off on the budget and has to explain it when it runs dry early.
A developer working on their own is not going to wake up one morning and decide to go save the company money, any more than most people volunteer to cut a budget that isn't theirs.
Most AI cost tools miss this. They surface spend to the engineer and hope a little visibility (or guilt) does the rest. The tool sits unused, aimed at the one person who was never going to act on it, while the real causes go untouched. Gartner names three of them clearly.[1]
- Ungoverned autonomy in agent driven workflows
- Context windows that balloon and inflate every single call
- No feedback loop, so waste compounds quietly where nobody is watching
Every one of those is a systems and leadership decision, not a developer habit. That includes context engineering. Gartner recommends it as something leadership mandates and enforces by default, the same way you'd set a resource limit or a rate cap, not as a discipline individual engineers are expected to maintain call by call.
The mandate belongs to the system that runs the request, not the person typing the prompt.
Where should AI cost controls should live?
If the developer is the wrong layer, the right one is wherever every request already flows through. Tokens get consumed at a single point, and that's where visibility and limits belong. Watch spend there, set thresholds there, and cost gets managed by default instead of by willpower.
That also frees developers to do the thing you actually hired them for. Nobody has to change habits or sit through a frugality training that was never going to stick. The system carries the discipline so people don't have to.
How Revenium controls AI token costs
Revenium is the AI Economic Control System that sits at the layer where AI cost controls already flow. For the person who owns the budget, it delivers real-time visibility. Spend is broken out by team, feature, and workflow. Overrun alerts fire before the quarter closes—not after.
If a single agent run starts consuming 10x its normal token budget, that triggers a threshold alert the same afternoon, before it becomes a line item someone notices in next month's invoice.
For the developer, it adds nothing to their workflow. No dashboard to check, no habit to build. The control lives in the system; the spend stays visible and bounded regardless of what any individual engineer does that day.
How to actually control your AI coding costs
If you run a budget that AI touches, don’t hand cost discipline to your engineers and expect it to stick. Own it yourself. Start here: can you see your token spend broken out by team, in real time, today? If not, that's the gap, and no amount of developer training closes it.
Then put controls at the layer every request already passes through, with visibility and limits on by default. That's the part you can act on without asking a single engineer to change how they work.
AI spend is going to keep climbing. Gartner has the trajectory right. The fix was never going to come from the person writing the code. Get it right, and an overrun becomes an alert instead of a quarter-end surprise, and no developer is ever asked to think about tokens.
