Anthropic may look stronger than OpenAI on the usual pre-IPO scoreboard.
It has reported stronger private-market momentum. It appears closer to near-term operating profit. Its gross margin is reported above OpenAI’s. It has deep enterprise relationships, a strong reputation with developers, and Claude Code has become one of the clearest examples of an AI product that customers already pay for at scale.
On the surface, that looks like the cleaner business.
But the more useful question is not which company looks better today. It is which company gets stronger as AI does what everyone expects it to do: improve and get cheaper.
On that question, OpenAI and Anthropic are not the same kind of company.
Anthropic mainly sells access to frontier intelligence. OpenAI sells that too, but it also controls a mass consumer interface used by hundreds of millions of people. That difference matters because falling AI costs do not affect both businesses in the same way. For a frontier model seller, cheaper AI erodes the price of the thing being sold. For a consumer platform, cheaper AI lowers the cost of serving users whose attention can be monetised through advertising, commerce, subscriptions, and referrals.
The same cost curve can damage one business model and strengthen another.
That is the central divergence.
The price of AI work is falling extremely fast. Depending on the benchmark and the task, equivalent-quality AI has been getting cheaper by orders of magnitude. One widely cited estimate puts constant-capability inference cost decline at roughly tenfold per year. Other measurements show even sharper falls for some tasks. Work that once cost tens of dollars per million tokens has moved toward cents.
This is not a normal software pricing cycle. It is the economics of a manufactured input. As models improve, hardware scales, inference systems are optimised, and competition intensifies, the unit cost of producing cognitive work falls. The unusual part is that the manufactured product is not a phone, chip, battery, or solar panel. It is intelligence delivered through computation.
That distinction matters for valuation.
If a company sells a scarce software product with durable pricing power, investors can imagine software margins. But if a company sells a manufactured input on a steep cost curve, the better analogy is not classic enterprise software. It is a commodity producer with a premium tier on top. The premium may be valuable, but it is constantly under attack from the next cheaper substitute.
This is the problem Anthropic has to solve.
Also Read: Your offshore vendor’s AI is running on your code: Do you know which one?
At the moment, frontier models are not good enough for many complete workflows. They still make errors. They need supervision. They lose context. They often complete pieces of work rather than whole jobs. Because of that, each capability improvement increases demand for the newest model. Buyers want the best system because the current one is still not quite enough.
That creates a temporary premium market.
But the demand for capability is bounded by the task. Once a model can perform a defined job to a competent standard, a better model adds less value. If an AI system can take a set of accounting records, identify what matters, apply the rules, produce the filing, flag the judgment calls, and explain the output, then the buyer no longer needs the newest frontier model for that task. The job is done.
At that point, the buyer has a different question: what is the cheapest model that clears the bar?
That is the good-enough threshold. Once a task crosses it, the task leaves the premium market. It falls into the commodity market, where open-weight models, older frontier models, and cheaper specialist systems compete for the work. The frontier model may still be better in a general sense, but better no longer matters enough to command a large price premium for that specific job.
This is the structural trap for a business built around selling frontier access.
Improving the model conquers more tasks. But conquering a task means that task eventually stops needing the frontier. Over time, the frontier-only market does not automatically expand. It may narrow unless new categories of work open faster than old ones commoditise.
Anthropic is not blind to this. Claude Code is important because it moves the company higher up the stack. It is not merely selling tokens. It is selling a specific outcome inside a valuable workflow. That is the right direction. Application-layer products are less exposed than raw model access because customers are paying for the completed job, integration, and workflow value rather than just the intelligence underneath.
But the valuation question remains. How much of Anthropic’s future value comes from durable application products, and how much still depends on frontier access retaining premium pricing?
OpenAI has a different problem and a different opportunity.
It also sells model access. It also faces inference costs. It also competes in the frontier race. But it owns something Anthropic does not: a consumer destination at enormous scale.
That changes the effect of falling AI costs.
Also Read: AI is answering your customers before they ever click, and it may never mention you
For OpenAI, cheaper inference reduces the cost of serving free and low-paying users. If those users can be monetised through advertising, commerce, referrals, subscriptions, enterprise conversion, or in-chat purchasing, then falling AI costs widen the gap between the cost of serving attention and the value of monetising it.
That is not the economics of a pure model vendor. It is the economics of a platform.
This is why OpenAI’s experiments with advertising and in-chat commerce matter. They are not just incremental monetisation ideas. They are attempts to shift the company away from selling tokens and toward monetising the interface where users already spend time, search for answers, compare products, and make decisions.
If ChatGPT becomes a meaningful consumer gateway, then cheaper AI helps OpenAI twice. It lowers the cost of each interaction, and it increases the number of interactions that can be economically served. More usage is no longer only a cost burden. It becomes monetisable surface area.
That is the flywheel OpenAI is trying to build.
The contrast is simple. A tenfold annual fall in inference cost lowers the price of the thing Anthropic mainly sells. The same tenfold fall lowers the cost of the thing OpenAI can give away to attract and monetise users.
One business sells the deflating input. The other may use the deflating input to build a larger platform.
That does not make OpenAI’s outcome guaranteed. Advertising inside an AI assistant could damage user trust. Commerce may not convert at scale. Referral economics may be weaker than expected. Regulators may limit parts of the model. Users may resist a shift from neutral assistant to monetised shopping interface. The consumer flywheel is still a thesis, not a proven revenue engine.
OpenAI also carries its own financial pressure. Serving hundreds of millions of users is expensive, even when unit costs are falling. Infrastructure commitments are large. The company still has to prove that mass usage converts into durable economics rather than just enormous demand for subsidised computation.
But structurally, OpenAI has more ways to benefit from AI deflation.
Also Read: Beyond the chatbot: How Gen Z pioneers are leading ASEAN’s new AI revolution
Anthropic’s strongest counterargument is that frontier capability may stay scarce. If the best models remain meaningfully better, if enterprise customers are deeply locked into workflows, if safety, reliability, compliance, and data integration matter more than raw token price, then premium pricing can survive longer than the simple commodity story suggests. In that version of the market, Anthropic’s enterprise concentration is not a weakness. It is evidence of pricing power.
The second counterargument is that new frontier-only work may open faster than existing work commoditises. If each generation of models enables qualitatively new tasks — autonomous agents, longer-horizon reasoning, richer multimodal work, or entirely new software workflows — then demand for the newest model may keep expanding.
That is the key uncertainty.
The question is whether frontier intelligence remains a scarce product, or whether it becomes a rapidly cheapening input.
If it remains scarce, Anthropic’s model is stronger than the deflation argument implies. If it becomes a cheap input, value moves elsewhere: to compute and energy capacity, to distribution, to consumer interfaces, and to applications that convert cheap intelligence into specific outcomes.
That is why the IPO filings matter. The headline valuation will get the attention. The more important signals will be beneath it.
The first signal is revenue quality. Is reported run-rate revenue gross or net of reseller and partner pass-through? A large gap between gross and net would make the top line look stronger than the underlying economics.
The second signal is gross margin. If margins rise sharply while inference prices keep falling, that supports the view that frontier labs can retain pricing power. If margins remain compressed, the commodity interpretation gains strength.
The third signal is revenue mix. How much comes from raw model access, and how much comes from higher-stack products? For Anthropic, Claude Code and similar workflow products matter because they reduce dependence on frontier access alone. For OpenAI, advertising, commerce, subscriptions, and platform monetisation matter because they show whether consumer distribution can become a real economic engine.
The fourth signal is customer behaviour. If enterprises keep paying for the newest model even after cheaper alternatives become good enough for many tasks, lock-in is stronger than expected. If customers shift workloads aggressively to lower-cost models once capability thresholds are crossed, the frontier premium decays.
The useful conclusion is not that OpenAI is certain to beat Anthropic.
It is that the normal scoreboard may be measuring the wrong thing. Revenue growth, private valuation, filing sequence, and near-term profitability describe the present. They do not answer the more important question: what happens as the product gets cheaper?
Anthropic may be ahead on today’s visible metrics. But if most of its value remains tied to selling frontier access, then it is exposed to the falling price of its own output. OpenAI may look less clean financially today, but if it turns cheaper AI into cheaper user acquisition, cheaper user service, and more monetisable attention, then the same deflation becomes an advantage.
The AI market is usually described as a race to build the best model.
That may be the wrong race to watch.
The durable value may not sit with whoever produces the frontier model at any given moment. It may sit with whoever owns the interface, the workflow, the distribution, and the customer relationship once intelligence itself becomes cheap.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post The real difference between OpenAI and Anthropic is what happens when AI gets cheaper appeared first on e27.