AI image and video 2026: the models are done, now taste counts

TL;DR. Generation costs cents. According to Google’s price list, an image costs between 3.36 and 13.4 cents, video around 10 cents per second. Text in images and believable motion are largely solved. The bottleneck is no longer rendering but selection: when an image costs a few cents, the only expensive part left is deciding which one fits the brand. And the models are unstable. OpenAI shut down Sora in 2026, and prices and model names change constantly. If you chain yourself to one model, you build on sand. The pipeline has to be model-agnostic.

The price list that turns the calculation around

If you last looked at the cost of image and video generation in 2024, you should recalculate. Google’s price list shows where things stand: Nano Banana Pro, Google’s strongest image model, costs 13.4 cents per image at 1K or 2K resolution. Nano Banana 2 is at 6.7 cents, Nano Banana 2.1 and the Lite variant at 3.36 cents per 1K image. Through the batch API, most of these prices halve. For video, Google quotes around 10 cents per second at 720p for Gemini Omni Flash.

The consequence for budgeting: rendering itself is no longer a significant cost block. Ten candidates for one motif cost $1.34 with Google’s most expensive model. A complete set of campaign motifs costs less to generate than an hour of a person’s time to review it.

What is actually solved in 2026, and what is not

The two classic sources of embarrassment in image generation, illegible text and implausible physics, have largely been cleared. Google presents Nano Banana Pro as a model for correctly set, legible text in images, in several languages too. That is a vendor claim. In video, too, bodies and objects now mostly move believably.

And yet “largely solved” does not mean “solved”. Engraved digits, handwriting, fine scales, anything that looks like text but is not, remain a reliable source of errors. Add hands, reflections and fabrics. The models are good enough to convince most of the time. The problem is the cases that slip through when nobody looks.

The end of Sora: why your pipeline must not have a favourite model

On March 24, 2026, OpenAI announced it would discontinue Sora. The app and website ran until April 26; the API with all Sora 2 models was switched off on September 24. There is no successor in the API. Anyone who had optimised their video production for Sora prompts, Sora quirks and Sora prices had to move in the middle of the financial year.

This is not an isolated case but the pattern of the market. Google deprecates too: according to its price list, the image model Gemini 2.5 Flash Image will be switched off on March 15, 2027, and users are to move to Nano Banana 2 Lite. New models appear every few months, and prices keep changing.

The business consequence is uncomfortable but clear: everything you build in knowledge and processes has to sit above the model layer. Your design rules, your quality criteria, your approval processes are the assets that last. The specific model behind them is replaceable infrastructure, like a hosting provider. A model-agnostic pipeline defines what a good result is and can switch renderers without the brand noticing.

The new bottleneck: nobody checks what comes out

Adoption is there. According to Salesforce’s State of Marketing report 2026, which surveyed about 4,450 marketers in October and November 2025, 75% of marketing organisations use at least one form of AI, for example to personalise content or generate images. The production side has long been scaling.

The control side lags behind. When an image costs a few cents, rendering ten candidates and throwing away nine is economically trivial. What is expensive is no longer generation but the decision: which of the ten fits the brand? Who makes that decision, by which criteria, and how does it scale across hundreds of motifs? The answer cannot be that the managing director takes a quick look. The output is too high for that. It has to be: codified taste plus automated checking.

How we solve it ourselves: a taste codex and three judges

For our own channels we run an image pipeline on this principle. From a set of proven reference images we derived a taste codex: a few design axes, such as lighting, materiality and visual calm, plus hard rules that must never be broken. Several candidates are created per motif. Then three independent, automated checks decide, each with rejection as the default.

1. 1Taste codexdesign axes, hard rules
2. 2Render candidatesseveral variants per motif
3. 3Three judgestopic, taste, AI artefacts
4. 4Approvalonly if all three pass

The three questions are: does the image fit the topic? Does it fit the codex? And does it show no typical AI artefacts, the so-called slop? Only a candidate that passes all three checks goes live.

The errors that candidates fail on repeat themselves: illegible pseudo-text in the image, AI-typical fake handwriting that looks real at first glance, and a sterile catalogue look that is clean but interchangeable. When the reasons for rejection are named precisely, the loop of rejecting and re-rendering converges quickly.

What belongs in such a codex? In our case, a few clearly checkable points:

• Hard exclusions. No text from the image model, no redrawn logos, no faces of real people, no colours outside the brand palette.
• Design axes with examples. For each axis, such as light or materiality, one good and one bad reference image. A judge can only check against something it can see.
• Typical artefacts as a list. Hands, reflections, fabrics, pseudo-text, duplicated objects. This list grows with every rejection.
• Clear ownership. Who changes the codex is decided by a person, not by the judge.

Three lessons carry over:

1. Plan motifs without text. Image models do not handle engraved digits and real handwriting reliably. Set text afterwards rather than hoping for the next model update.
2. Clichés fail on taste. What the first prompting reflex delivers looks like everyone else’s.
3. Rejection by default is the lever. A judge who waves things through when in doubt is worthless. Three judges who reject when in doubt turn cheap candidates into brand quality.

Labelling: what has applied since August

On top of the quality question comes a legal one. Since August 2, 2026, the transparency duties of Article 50 of the AI Act apply. Anyone who publishes images, videos or audio that deceptively imitate real people, places or events must label them as artificially generated. Providers of systems that generate such content must also mark their output in a machine-readable way. For systems already on the market before August, a transition period runs until December 2, 2026. For the pipeline, this means labelling belongs as a fixed step before approval, just like the three judges.

What this means for your company

1. Treat generation as raw material, not a cost factor. At a few cents per image, overproduction with hard selection is the right mode, not the one laboriously prompted single image.
2. Invest in selection, not in rendering. Write down what “good” means for your brand, with hard exclusion criteria, not as one person’s gut feeling. Only that document makes automated checks possible and your quality independent of who happens to be checking.
3. Stay model-agnostic. The end of Sora will not be the last of its kind. Anchor your knowledge in rules, review processes and reference sets that survive a model change, and treat the model as a replaceable supplier.
4. Measure first, then scale. Check your published content once against your own rules. It takes an afternoon and shows at once where you stand.

And because AI content increasingly has to be labelled, our post on the invisible mark in AI text is worth a look. If you would like to go through your image or video pipeline with us, we are happy to talk.