MOUNTAIN VIEW, 21 AUG 2026 — Google DeepMind has released two generative media models with unusually specific prices attached: Nano Banana 2 Lite at US$0.034 per thousand images with about four seconds of latency, and Gemini Omni Flash for video and conversational editing at US$0.10 per second of output.
Three and a half cents for a thousand images is the number that changes what people build.
The two models
Nano Banana 2 Lite is described as the company's fastest and most cost-efficient image generation model. Gemini Omni Flash generates video and supports conversational editing of it.
Run the arithmetic before reacting to it
At US$0.034 per thousand, a single image costs roughly three thousandths of a cent. A product generating one image per user action can serve a hundred thousand actions for about three dollars and forty cents.
That is below the threshold where generation cost appears in a product decision at all. It stops being a line item to manage and becomes something closer to bandwidth — a cost you assume rather than model. Anyone who has been rationing image generation behind a paywall, a credit system or a queue can stop, and the engineering built to enforce that rationing becomes dead weight.
The video pricing has the opposite effect. At ten cents per second — six dollars a minute — cost remains a significant factor. This prices video as a feature a user must request deliberately, not something a product can generate speculatively. That asymmetry — images effectively free, video distinctly not — will shape what gets built more than either model's quality does.
The four-second latency deserves as much attention as the price, because it decides where generation can sit in a product. Four seconds is too slow to feel instant and fast enough to sit inside a request a user has consciously made — the same range as a slow page load. That places it comfortably in an editing tool or a configurator and awkwardly in anything expected to feel live. Products that need generation to disappear into the interaction still cannot have it.
The new capability here is conversational editing.
Generating video from a prompt is a solved capability with several credible providers. Changing a generated video by describing the change is a different problem.
The reason it matters is that generation alone is nearly useless for real work. Nobody gets what they wanted on the first attempt, and the standard remedy — regenerate with a modified prompt — produces a different video rather than the same video with one thing altered. Continuity is what makes iteration possible, and its absence is why most video generation demonstrations show a single striking clip rather than a finished sequence.
Whether this delivers that is not established by an announcement. The specific test is whether asking for one change leaves everything else intact, and that is the property every provider claims and few demonstrate over multiple rounds.
One thing a low price does not fix is rights. Cheap generation multiplies the volume of imagery an organisation publishes, and every one of those images carries the same unresolved questions about training provenance and commercial indemnity that expensive generation did. A business publishing a hundred thousand generated assets has a hundred thousand times the exposure it had at one thousand, and the terms governing that are usually read once at procurement and never again.
Pricing has become the product announcement
Notice what these releases lead with: not a capability claim or a benchmark, but a price and a latency figure.
That is a market telling you where it is in its cycle. When capability differences were large, vendors announced capability. When several providers can produce an acceptable image from a prompt, the differences that remain are cost, speed and the willingness to commit to both in public.
We reported this month that Gemini 3.7 Flash's introductory pricing expires in December, which is the caveat to attach here. An introductory price is a customer acquisition instrument, and a product architecture built on a price that has an expiry date is carrying a risk its author may not have priced. Ask what the number becomes later, and assume you will not be told.
What this does to the regional creative economy
Southeast Asia's creative and marketing services sector is large, competitive and priced against volume, which makes it unusually exposed to this particular cost curve.
Agencies across the region produce enormous quantities of variant creative — the same campaign in six languages, a dozen aspect ratios and multiple regional treatments — and much of that work has been billed by the hour or by the asset. Generation at three thousandths of a cent does not eliminate that work; it removes the justification for pricing it by unit.
The firms that do well from this are the ones whose value was never the production. Art direction, cultural judgement about what works in Jakarta versus Manila, and the taste to reject nine images out of ten remain scarce. The firms in difficulty are those whose margin came from executing volume, and they are numerous here.
There is also a straightforward opportunity. Products that were uneconomic at previous prices, like personalised imagery for every user or listing, are now buildable by small regional teams without a capital budget. This levels the field in a market where compute has long been the barrier.
What we could not establish
Output quality relative to price. A cheap model that produces images people reject is not cheap, and nothing in a price announcement addresses whether the output is usable for a given purpose.
Other open questions include: whether these prices are introductory and when they change; rate limits and whether the throughput exists to use the price at volume; what the four-second latency measures and at what resolution; the video model's maximum output length and resolution; how conversational editing performs over repeated rounds; commercial usage and indemnity terms; and regional availability, which for generative media models has repeatedly lagged the announcement across ASEAN markets.
What to watch
Watch for the price expiring. Introductory pricing on generative media has become standard, and the date matters more to anyone building a business on it than the launch number does.
Then watch whether competitors match within weeks. If they do, this is a market clearing at commodity prices and the differentiation moves entirely to integration and rights. If they do not, somebody has a cost advantage worth understanding.
Finally, watch regional availability specifically. A price nobody in your market can pay is a press release, and the gap between announcement and availability in ASEAN has been the recurring disappointment of this category.
View author profile → · Editorial policy
