Good Enough AI Already Won the Cost War
"Good enough" AI models made the whole industry affordable in 2026. That victory didn't make it manageable, because the model was never the expensive part.
Enterprise token costs fell from $18.40 to $6.07 per million in four months, and open-weight models took 38% of the volume, so the cost war is over and "good enough" won it. That victory didn't make AI manageable: 47% of people at work route around the budget they were given. The model was never the expensive part.
A few weeks ago I sat with the actual math behind why my own multi-model pipeline works the way it does. And around the same time, a research report landed that explained the same math at industry scale. Enterprise blended token costs fell from $18.40 to $6.07 per million tokens between January and April 2026, drawn from 2.4 billion API calls in AI.cc's 2026 AI API Infrastructure Report. So the shift is easier to see side by side than to summarize in a sentence:
- 1Blended token cost: $18.40 to $6.07 per million tokens, January to April 2026 (AI.cc).
- 2Open-weight share: 11% to 38% of enterprise token volume in one year (AI.cc).
DeepSeek's V4-Flash now matches a paid flagship's intelligence score at a fraction of the price. And even Microsoft's own AI division has leaned its public strategy toward "good enough" capability priced low, ahead of chasing the top of any benchmark chart, per an industry analysis of AI vendor strategy published in July 2026.
Because if the AI cost war was ever a fight over which model to buy, somebody already won it. "Good enough" models did.
And I want to sit with that for a second, because I don't think it means what most people are about to assume it means. "Good enough" is doing what the label promises, mostly. Models that needed a six-figure contract two years ago now run for lunch money, but what's left over sits somewhere between irrelevant and still the whole game for the hardest slice of tasks, which is a less satisfying answer than either side of that argument wants to give you. So set that tension aside for now: a cheap model still decides things, and the deciding has no vendor, no price tag, and no owner in most companies.
Affordable isn't manageable
A few weeks ago I wrote about a friend of mine at a European marketing agency's China office. Her agency's French headquarters signed a partnership with Anthropic and handed her office a token budget, generous on paper, tight in practice. And she burned through it fast enough that she started weighing a personal account to cover the shortfall.
I'm not going to resolve that story today either, and I've stopped feeling like I need to. It was never really about her, or her employer, or which vendor happened to sign the contract. It's one visible version of something that's apparently everywhere:
47%
And her office had a deal. By the time it mattered, price wasn't the thing deciding whether the budget held. I keep coming back to that, because it's such a clean disproof of the assumption everyone's currently making about what fixes this.
The model was never the point
The model is infrastructure. The system around it is the product. And I made that argument building The Seven Dimensions of Enterprise AI, long before this year's price collapse made it easy to prove with a chart. Priced this low, "good enough" models make the argument truer than ever. Because once every serious option clears the same floor of capability at close to the same low price, model choice stops being a decision worth making.
Nobody wins by picking the model anymore, the way nobody wins a coin flip by calling heads with more conviction.
That's the sentence I'd want someone to remember from this whole piece, if they only remembered one.
What actually fixes this?
But ask five people what a company should do about this and several of the five answers will be right, for different companies, at the same time, which is an uncomfortable thing to hear when you wanted one clean recommendation. Deployment support is a real answer: it's the same instinct behind the forward-deployed-engineer hiring surge I covered a few weeks ago, postings up more than 800% year over year by early 2026, per SkillScouter. And visibility tooling is a real answer too. An entire product category now exists to detect this kind of unmanaged usage, built because nobody could see it happening until it already had. Training helps somewhere. Governance helps somewhere else. So none of them alone closes the distance, and a company that picks one and calls the problem solved has chosen a symptom to treat and is hoping it picked the right one.
Autonomous driving runs into the same shape of problem at a different scale. A perception agent, a prediction agent, a planning agent, and a control agent each own one piece of a moving car, and none of them drives it alone. Get the handoff between them wrong by fifty milliseconds and the car brakes for a pedestrian who already crossed.
The car works because four narrower jobs each get done right, in relation to the others.
So that's the case for building across all seven dimensions of a system. Data Architecture, Domain Logic, Orchestration, Models & Agents, Governance & Containment, Human Judgment, Calibration: each one covers a different way a system, or a company's use of one, can fail, and no downstream dimension compensates for a decision made poorly further up.
What I built instead of writing
And TikSense is where I've built this. Two examples, chosen because they sit on genuinely different fronts, and neither alone would have been enough.
The business call
A routing step decides whether a creator counts as proven, emerging, or aspirational, meaning commercial intent without an actual sales record behind it. And a person decided what counts as a track record, encoded it once, and it shapes every downstream analysis before the analysis starts. It's the least glamorous afternoon of the whole build, and also one of the only decisions that's never needed revisiting since, a person's judgment, frozen into the system on purpose.
The authority boundary
TikSense runs two tiers of agents. The first tier classifies and scores; none of those agents can resolve a conflict between their own outputs. And a second, senior tier activates on three conditions: two tracks disagreeing past a threshold, confidence dropping below floor on a high-potential creator, or a data point in what's scraped that falls outside every track's scoring domain. It resolves what it can itself. So a human only sees what that tier can't authorize, which by design stays rare. That's the Agentic Authority Escalation Model, doing its actual job on live creator data, past the diagram stage.
And neither of those is the whole fix. That was never the point. The system is where all of this lives, and a system doesn't get built one afternoon and one lever at a time. I know that firsthand, mostly because I tried to skip steps early on and had to come back for them later.
Somewhere between a low-cost model and an outcome a TikTok seller can use, someone has to own the whole picture instead of patching one hole at a time. Most companies haven't decided that's one job yet, and most haven't decided whose job it would be either. The AI cost war is over. "Good enough" models won it, and it turns out that was never the expensive part.
This lives wherever you already are, Substack, X, or LinkedIn, drawn from building TikSense.