Gartner says agent costs will rise fivefold by 2028 even as tokens get cheaper. Here is how to keep them in check

Onyx So Gartner just dropped this thing saying agent costs are going to rise fivefold by twenty twenty-eight, even though token prices are in free fall. That's… a lot of people read that and went 'wait, what?'

Echo Yeah, and I think that headline is doing real work that the actual story doesn't entirely earn. The claim is that agents burn five to thirty times more tokens per task than a basic chatbot. That's real. But then they jump to 'therefore costs rise fivefold across the board by twenty twenty-eight' and that's where I start squinting.

Onyx Okay, but—

Echo Their model has three hundred plus inputs. It's an econometric forecast. That's not nothing, but agent cost drivers in production are still basically a black box for most teams. We're talking about routing decisions, context window management, retry logic, whether you're even measuring per-task cost at all. Most enterprises don't have that instrumented yet.

Onyx Right, but the direction is not wrong. And here's what actually matters to me—Gartner's own survey found that twenty-one percent of enterprises running orchestration platforms have zero automatic spend controls. Zero. So even if the fivefold is a model output and not ground truth, the fact that a fifth of teams can't even stop an agent from burning tokens is the real signal.

Echo Fair.

Onyx The token price fell. That's true. But labs didn't pocket the savings—they poured it into Fable and Claude Opus Five Point Five, which cost more per token than the previous tier. And simultaneously, agents don't work like chatbots. A chatbot answers a question. An agent chains reasoning, calls tools, retries, builds context. That's five to thirty times the tokens for the same outcome. So yeah, the economics break in a weird way.

Echo I don't disagree with the mechanism. What I'm saying is, the fivefold assumes enterprises keep adopting more complex agents at the rate Gartner's model predicts. But there's also a cost-cutting phase happening right now—they even mention it, 'belt-tightening phase.' So the forecast is jagged, not smooth. And if teams get smarter about routing and observability, the slope could be much shallower than fivefold.

Onyx Okay, so that's actually where the five recommendations come in. The thing that stood out to me is they're not saying 'agents are doomed' or 'go back to chatbots.' They're saying: observe what you're spending, measure value per outcome, tier your models so you don't default to Fable for everything, set execution standards before you ship, and use proprietary data to stay ahead. Those are all executable moves.

Echo Sure.

Onyx And none of them are new. But the fact that Gartner is spelling them out as 'here's how to not get killed by the Inference Paradox' tells you something about where the field is. Observability, routing discipline, value metrics—this is the boring infrastructure layer that actually determines whether your agent system survives contact with finance.

Echo That's the part I actually buy. Inference tiering is not complicated. You route simple lookups and basic QA to Opus Five Point Five or Claude Fable Five Point One, you reserve the expensive stuff for actual reasoning. But here's where I get skeptical again: how many enterprises are actually going to enforce that? The default is always 'use the best model we have.' Changing that requires governance and discipline. And most teams don't have that muscle yet.

Onyx They will, or they'll burn. I mean, we've talked about this before—the boring control layer is the product. If you can't answer 'why did this agent spend thirty bucks on this task,' you're not shipping agents in production, you're running an experiment on someone else's dime.

Echo Okay yeah. That's fair. The observability piece is real. And if Gartner is right that we're in a belt-tightening phase, the next wave of agent deployments will absolutely have cost gates baked in from day one. So maybe the fivefold lands higher in less disciplined orgs and lower in the ones that build this upfront.

Onyx Exactly. And that's not a Gartner forecast—that's just market selection. The teams that treat agent cost as a feature, not a surprise, are the ones who scale. The rest learn the hard way.

Echo One thing I'd push back on slightly: Gartner says proprietary data and fine-tuning are the move to stay ahead of margin compression. And yeah, that's true in principle. But fine-tuning at scale is expensive and slow. Most teams aren't going to have the data, the infra, or the patience to do it well. So I'd put that one further out on the adoption curve than they're implying.

Onyx Fair. That's more of a year-two or year-three move once you've got observability and routing locked. But right now, if you're shipping agents, the bet is observability and tiering first, everything else after.

Echo Yeah. And if you don't have those, the fivefold might actually be conservative.

Onyx So my read is: the Inference Paradox is real, the fivefold is a plausible forecast given the assumptions, but the actual cost curve depends entirely on whether teams build observability and enforce routing discipline in the next few months. If they do, they can blunt it. If they don't, Gartner's probably right.

Echo I'd say sixty-forty the fivefold lands a bit lower than predicted by end of twenty twenty-eight, but only because cost pressure is going to accelerate observability adoption faster than Gartner's model assumes. The mechanism they're pointing at though? That's solid.

Onyx Alright. So if you're running agents in prod right now, the move is: instrument per-task cost, set up tiering so you're not defaulting to the expensive model, and track value per outcome. That's not revolutionary, but it's the difference between scaling and exploding.