State of Generative AI: From Creation to Orchestration

Section 01

From Creation to Orchestration

The unit of analysis is the user job, inferred from the inputs attached to each successful inference. Models come and go around jobs; jobs are what people are actually trying to do.

The composition of image generation changed substantially: text-only creation lost share while multi-reference workflows expanded.

Text-to-image is declining as a workflow, not as a technology.

Section 02

The same users are changing how they work

Most of the growth in Orchestrate occurs within existing users. Audience replacement contributes more to the decline in Create.

For comparability, both populations are normalized across four prompted image jobs: Create, Transform, Compose and Orchestrate. Utility workflows are excluded.

This is descriptive evidence, not causal attribution. A ratio below 1 does not prove that the remainder is audience replacement, and Transform carries no ratio at all because the market moved only −0.04 pp — too little for the division to mean anything. The fixed cohort represents a declining share of linked image inference activity over the period, so it is used as a robustness check rather than as a replacement for the full population.

The fixed cohort contains users with linked image-inference activity in both November 2025 and July 2026. Inferences are linked to users through request_id, and only requests mapping unambiguously to one user are retained. Shares are measured from the first observed month to the last.

Section 03

Video is already reference-led

If image work is travelling from creation toward orchestration, video shows what the destination looks like — because video never made the journey. It started there.

Image-to-video held between 88.32% and 92.24% in every one of the nine months. Text-to-video never exceeded 6.60% and ended at 3.98%. The fastest-moving video job is the one that starts from existing footage: video-to-video went from 2.57% to 6.97% of video work.

These proportions describe the observed platform population. They are not a claim about the entire global video-generation market.

Section 04

Complex workflows are becoming more explicit

Reference counts say what people brought. Prompt length says how much direction they gave — and it grew inside every job, not only because the job mix changed.

The Transform figure is the most telling. The median instruction for changing a single image went from 80 to 242 characters. Eighty characters is a phrase — make it night-time. Two hundred and forty is a sentence with constraints in it: the same request type, three times the specification. Orchestration went from 324 to 909, with a ninetieth percentile of 3998 — multi-paragraph art direction attached to four or more images.

The increase occurs within jobs, not only because the job mix changed: every line above rises on its own, holding the job constant.

Section 05

Model relevance depends on the job

Now models — conditionally. For each job we can ask which models the work actually goes to, which is a different question from which model is used most overall.

The clearest result is a matched pair of opposites. Flux 2.0 over-indexes 2.78× on the blank canvas and collapses to 0.14× on four-plus-reference assembly. Seedream 4.5 is its mirror image, over-indexing 2.39× on assembly. Two models with comparable overall footprints and almost inverse profiles: an overall ranking would have placed them next to each other and said nothing useful about either.

Section 06

Methodology and limitations

Every figure in this report is computed from aggregate exports at build time. None is hardcoded, and automated assertions guard the citable ones.

Limitations

Observational, not a census

This is observational behaviour on a large multi-model platform, not a census of the global generative-AI market.

Supply expanded while demand shifted

Multi-reference capability was already present in the first observed month, but the number of models supporting it also expanded. The shift contains both changing user behaviour and expanding technical capability, and this dataset cannot separate them.

The taxonomy is structural, not semantic

Jobs are inferred from input counts: the taxonomy describes how users work, not what they create. It does not identify product photography, portraits, architecture, packaging, fashion, advertising or any other semantic category. A semantic classifier is deferred to a later iteration.

Composition, not volume

This report is about the composition of work — the mix of jobs and how it changes. It is not a report about how much work was produced, and it makes no absolute-volume claims.

Prompt length is not prompt quality

Length is a proxy for explicit direction and task complexity. This dataset contains no measure of output quality or user satisfaction, and no literal prompt content is exposed.

The specialization index is approximate

It is built from leading models per job and has a structural ceiling per job. Treat it as directional.

Pricing is out of scope

Subscription, credit-price and plan analysis sits outside this behavioral report — those are exactly the product decisions that would contaminate it.