Trends · 5 minute read
The Commoditization of Model Capability and What Replaces It
Frontier model capability is converging while prices fall, which means a product whose advantage is its model has an advantage that expires. What does not commoditise is proprietary data access, evaluation infrastructure, workflow fit, and the operational discipline to run a system reliably.
Frontier model capability is converging and prices keep falling, which makes any advantage built on model choice temporary. This piece covers what replaces it, drawing on FISTA Solutions' AI enablement delivery work.
What is commoditising and what is not?
The split is clearer than the discourse suggests.
| Commoditising | Not commoditising |
|---|---|
| Frontier capability | Proprietary data access |
| Price per token | Evaluation for your task |
| General reasoning | Domain workflow fit |
| Basic tool use | Operational reliability |
| Prompt techniques | Organisational adoption |
| Model hosting | Regulatory position |
Why do capability advantages expire?
Because the gap between providers closes in months, not years.
A capability that one provider has exclusively tends to appear elsewhere within a release cycle or two. A product built on that exclusivity has a window, and the window is shorter than the time required to build a defensible business.
The practical consequence is that any strategy sentence beginning with the name of a model is fragile. The strategy has to be about what surrounds it. See the shift from model choice to system design.
What does proprietary data actually give you?
Answers others cannot produce, regardless of which model they use.
An organisation's operational history, its customer interactions, its documented procedures, and its domain corpus are not available to competitors at any price. A system grounded in that material produces output that a general tool cannot.
This is the most durable advantage available, and it depends on data quality rather than model choice. Teams with poor data hygiene find this out during their first serious retrieval project. See why data quality decides AI outcomes.
Why does evaluation compound?
Because it converts every future model release from a risk into an opportunity.
A team with a representative evaluation suite can test a new model against their actual workload, see whether quality improved and cost fell, and switch with evidence. That takes days.
A team without one has to decide by feel, which means they either switch carelessly or do not switch at all. Both are expensive, and the cost grows each time a better or cheaper option appears. See how to build an agent evaluation harness.
What does falling cost enable?
Use cases that were uneconomic become viable, which is where most of the value arrives.
A workflow that costs more to automate than to do manually is not deployed. As price falls, the set of viable workflows expands, and the organisations positioned to take advantage are those with the integration and evaluation work already done.
It also means cost projections made today are pessimistic. Building a business case that only works at current prices is conservative; building one that requires prices to fall is not.
How should systems be designed for this?
With the model as a replaceable component.
That means an abstraction over the provider interface, prompts and evaluation versioned as artefacts, and caution about building on capabilities only one provider offers. Where you do use a proprietary capability, know what it would cost to leave.
Portability is not free, and a thin abstraction that leaks is worse than none. The test is whether you could run your evaluation suite against a different provider next week. See how to run a model migration.
What about the providers themselves?
They compete on capability, price, and increasingly on the surrounding platform.
For buyers, the useful consequence is that switching costs are the thing to watch. Provider-specific tooling, fine-tuned models, and proprietary formats all raise the cost of leaving, which is a commercial position rather than a technical one.
Evaluate those features on their merits, and price the lock-in explicitly rather than discovering it at renewal.
What is the counter-argument?
The counter is that frontier capability still matters enormously for the hardest tasks, and that is correct. Research, complex reasoning, and long-horizon agent work all benefit from the best available model. The argument here is about durable advantage, not about whether capability matters — it does, for everyone, equally.
What does this change for engineering teams?
It shifts effort from prompt tuning toward the surrounding system: retrieval, evaluation, observability, and cost control. Those are engineering disciplines that transfer across model generations.
It also means treating the model as a dependency with a version and a migration plan, the way you would treat a database engine.
What does this change for buyers?
It means discounting vendor claims about which model they use, and asking instead about data handling, evaluation evidence, and what happens when a better model appears.
A vendor who cannot switch models quickly will be slow to pass on capability and price improvements you could otherwise have.
What should leaders do about it now?
Fund evaluation infrastructure before funding model experimentation. It is the asset that makes every subsequent decision cheaper and safer.
Then audit where your advantage actually comes from. If the answer is a model, the advantage has an expiry date and the work is to build something that does not.
Does this apply to open-weight models?
It accelerates the trend. Capable open-weight models put a floor under what is freely available, which compresses pricing and makes self-hosting viable for a wider set of workloads.
The operational cost of running them is the trade, and for most organisations a hosted frontier model remains cheaper all-in. See the economics of open weight models.
How will you know if this is happening?
Watch for a competitor matching your capability within a quarter, for cost per task falling without any change you made, and for a new release making a deferred use case viable. All three confirm the trend.
How FISTA Solutions reads this
FISTA Solutions builds and operates production AI systems through AI agents, AI enablement, and forward deployed engineering: the model kept replaceable behind an abstraction, and evaluation built first so every future model release becomes an opportunity rather than a risk, decisions documented with their reasoning, and handover that leaves your team able to maintain what was delivered. The record is 150+ projects for 50+ companies across 12+ countries.
To discuss what this means for your roadmap, message FISTA on WhatsApp, or read why evaluation is the new moat.
Share-ready article cover
Download the generated social format.
Clear answers
Questions raised by this field note.
Straightforward guidance for evaluating scope, fit, and the next step.
01Does model choice still matter?
For specific tasks and cost profiles, yes. As a durable competitive position, no — a capability advantage measured in months cannot support a product strategy measured in years.
02What replaces it as a differentiator?
Access to data others do not have, evaluation that proves quality for your specific task, workflow integration that fits a real process, and the operational discipline to run it reliably at volume.
03Why is evaluation an asset?
Because it compounds. A team with a good evaluation suite can adopt a new model in days and know whether it helped. A team without one cannot safely change anything, which is a growing cost.
04What does falling cost change?
It makes previously uneconomic use cases viable and makes today's cost projections pessimistic. It also means cost-based competitive advantages erode, since everyone's costs fall together.
05How should architecture respond?
By keeping the model replaceable. Abstract the provider interface, keep prompts and evaluation versioned, and avoid building on capabilities only one provider offers unless the benefit is substantial.
Continue exploring
Related capabilities
Start with the hard problem
Need the outcome owned, not merely analyzed?
Tell us where delivery is constrained. We’ll map the fastest credible path from intent to verified production.