Coverage of generative imagery still fixates on capability: which model renders hands correctly, which handles typography, which produces the most convincing photorealism this quarter. That race is real, but it has obscured a more consequential shift. The interesting story of the past eighteen months is not that the models got better. It is that access to them became a utility.
From Capital Expense to Metered Line Item
Three years ago, generating images at any serious volume implied hardware. An organisation either bought GPUs, rented dedicated instances, or accepted that image production stayed a manual craft. The capability sat behind a capital decision, which meant it sat behind a budget cycle, which meant most organisations simply did not have it.
That barrier dissolved quietly. Hosted models now price per image, in fractions of a cent to a few tens of cents depending on resolution and quality tier. There is no floor, no commitment, no minimum. A research group generating a few hundred illustrations a month pays less than it spends on coffee, and pays nothing in the months it generates nothing.
Economists have a name for what happens when a critical input becomes cheap and abundant: advantage stops accruing to whoever owns the input and starts accruing to whoever applies it well. The same transition happened to computing itself, to bandwidth, and to storage. Image generation completed it faster than any of them.
The Fragmentation Nobody Anticipated
What complicates the picture is that no single laboratory leads across the board. The frontier is contested by a dozen serious research groups on three continents, each strongest in a different dimension — one at photorealistic scenes, another at flat illustration, a third at rendering legible text inside an image, which remains genuinely difficult across the field.
Rankings shift quarterly, and the price difference between roughly equivalent outputs regularly exceeds fifty percent. An organisation that wires itself permanently to one provider is, in effect, betting its production pipeline on that provider staying ahead — a bet the past two years suggest nobody should make.
The market’s answer has been an aggregation layer. Published rates for the Nano Banana 2 API and competing image models now sit side by side under a single account with unified billing, which reduces switching from an engineering project to a configuration change. For teams without dedicated machine-learning staff, that reversibility matters more than squeezing the last few percent of quality from any one vendor.
The Cost Structure Practitioners Actually Encounter
There is a number missing from every pricing page, and it governs real spend: how many attempts precede an accepted result.
Nobody keeps the first image. Observed practice runs three to eight generations before something clears the bar, meaning the true cost per usable asset is several multiples of the quoted rate. Any forecast built on the headline figure is wrong in proportion to how exacting the standards are.
The correction is procedural rather than technical. Generate exploratory drafts at low resolution and low quality to choose a direction, then regenerate only the selected candidate at full settings. Groups that adopt this pattern report roughly halving expenditure with no discernible difference in what finally ships — because the overwhelming majority of generations are discarded within seconds of appearing.
What Has Not Been Automated
The limitation worth stating plainly is that none of this produces judgement. Generating fifty variations takes minutes; recognising which one communicates the intended idea to the intended audience still requires someone who understands both. The bottleneck moved from execution to direction, and direction has not become cheaper.
A second limit is more concrete. Text rendered inside generated images remains unreliable across every current model — if a graphic requires a label or a caption, it is added afterwards in a design tool. Fine repeated patterns, logos, and hands still betray synthetic origin under scrutiny.
And a boundary that credible organisations observe: generated imagery is used for concept, environment, and illustration. It is not used to depict a physical product a customer will receive, nor to represent people who do not exist as though they do. That line is drawn for commercial reasons more than ethical ones — audiences have become notably fluent at spotting the uncanny, and the credibility cost of being caught exceeds the production saving by a wide margin.
The Consequence
When a capability becomes a metered utility, the question facing an organisation stops being “can we afford this” and becomes “what would we do with it that is worth doing”. That is a harder question, and considerably more interesting than which model currently renders the best hands.
The groups extracting real value from generative imagery are not the ones that picked the best model. They are the ones that identified a specific, repeated visual bottleneck in their own work and pointed the capability at it — then kept the choice reversible, because the leaderboard will look different next quarter.
