Models are becoming a commodity, workflows aren't

If you wanted to summarize what's happening in the AI video generation market right now in a single line, it would be this:

"Models are becoming a commodity, faster and faster — and what makes a product distinct is moving up the stack, above the model."

Back in 2024, the market split on which model could move a person convincingly, or imitate camera work. Through 2025, those gaps narrowed quickly. By 2026, whatever one model does well, another one catches up to within two or three months. The gap didn't disappear — the duration that any given gap stays open did.

This post is an attempt to map where the industry is going, written from the side that builds tools on top of these models.

Models are becoming a commodity

The trend over the past year is clear.

• New models ship on a shorter cycle. When one company releases a new model, a comparable one tends to land elsewhere within the same quarter.
• Per-second pricing has dropped. Producing the same result costs anywhere from half to a quarter of what it cost six months ago.
• Open source has caught up. Quality that used to only exist in closed models is now reachable through open ones like Wan or HunyuanVideo.

When these three happen at once, a product strategy that leans on a single model gets fragile. "Today we're best because we use model X" becomes a different name next quarter.

We've seen this in other industries. Databases. Cloud. At first the market split on which engine was fastest, which provider was cheapest. Over time, the engine and the provider became commodities, and the differentiation moved up — into operations, into pipelines, into tools.

AI video models are heading the same way.

So why are outcomes so different from the same model?

The interesting thing is what's happening alongside the commoditization.

Two people on the same model, the same price, the same free credits, working on the same topic — they produce results of very different quality. The difference doesn't come from the model.

When you look at where the difference does come from, it usually ends up in the same place.

• Which prompts they tried, and in what order
• How they kept one character looking like one person across cuts
• Which generations they kept and which they threw out
• How they stitched cuts together

That's not the model's territory. That's workflow.

A model gives you one cut. A film is not one cut. A film is a sequence of cuts arranged into a story. The path from "a good cut" to "a good film" isn't something the model is responsible for.

So the more the model becomes a commodity, the more the real question becomes: what do you put on top of it?

Model routing — not one model, several

One shift happening on the tooling side is model routing.

Within a single piece of work, different scenes go to different models depending on which one is strongest there. Action-heavy cut to Kling, quiet atmospheric cut to Veo, multi-shot composition to Seedance, cost-sensitive filler cut to open-source Wan. The user can pick model per cut, or the tool can route automatically based on the cut's intent.

Once that pattern settles in, "which model is best?" stops being the most important question for the user. The important question becomes "which model is this cut on?" — and more importantly, "do I have to make that choice every time, or does something do it for me?"

The tool's responsibility moves from calling one model to selecting and combining among several.

Consistency — what models don't solve

Separate from the commoditization story, there's a problem one model alone can't solve.

Character consistency.

A character that looked great in one cut comes back subtly different in the next — hair length shifts, the outfit reads as a different garment. For a single still image, getting one good frame is enough. For video, several cuts need to read as the same person to the viewer.

This is being addressed partially inside models: richer conditioning inputs, dedicated character embeddings, adapter approaches like LoRA. But one shot at a model isn't yet reliable at holding one identity across an entire piece.

So consistency has been settling into the tooling layer as a systems problem.

• Define characters at the project level, not per cut.
• Map each character to a trained adapter (LoRA or similar), and combine it automatically at generation time.
• Feed earlier cuts of the same character back in as references for new ones.

This kind of infrastructure doesn't automatically get cheaper when model prices drop. It's something the tool side has to actually build and run.

So where do the tools go?

Pulling it together, the center of gravity in this industry is shifting in a fairly clear direction.

Models are becoming more alike, and cheaper, more quickly. A product that's just a thin wrapper over one model loses meaning fast. What holds value is being the one that gives the user a good result without making them think about "is this on model X or model Y."

Workflow, consistency, sequence — that part doesn't commoditize. The steps a user goes through to make not one cut but one film don't get cheaper because model pricing fell. If anything, as models get cheaper and faster, users generate more, compare more, choose more. They need a tool that holds that whole process.

The responsibility surface of the tool is growing.

• Automating model selection (routing)
• Preserving characters, styles, and world rules at the project level
• Keeping the flow from scenario → storyboard → cuts → timeline coherent
• Recording and comparing across the many attempts a user makes

This kind of work keeps its value across whichever model arrives next quarter. You can swap out a model. You can't swap out the body of work the user has built around it.

Closing

"Which model is best?" is becoming a question asked on a quarterly rhythm — sometimes monthly. When a question has to be re-asked that often, the weight of the question itself is going down.

Models are becoming a commodity. What you put on top of them is not.

From the side that builds the tools, we're standing on that upper layer. The faster models get better, the more clearly the problems above them come into view, not the other way around. The better a model gets at one cut, the more the user wants to make one film, not one cut.

We don't know which model is shipping next quarter. But when it does ship — making sure the user doesn't lose their character while moving to it, doesn't lose the flow of their scenario — that part is on the tool.