OpenAI’s latest model will change the economics of software

WHEN OpenAI announced a new generative artificial intelligence (AI) model, called o3, a few days before Christmas, it aroused both excitement and scepticism. Excitement from those who expected its reasoning capabilities to be a big step towards superhuman intelligence (some reckoned it would be a bigger deal than OpenAI’s launch of ChatGPT in 2022). Scepticism because OpenAI did not release it to the public and had every incentive to overplay the firm’s pioneering role in AI to curry favour with Donald Trump, the then incoming president.

Yet since then, one point of consensus has emerged. The model, as well as its predecessor, o1 (o2 was skipped because that is the name of a European mobile network), produces better results the more “thinking” it does in response to a prompt. More thinking means more computing power – and a higher cost per query. As a result, a big change is afoot in the economics of the digital economy, which was built on providing cheap services to large numbers of people at low marginal cost, thanks to free distribution on the Internet. Every time models become more expensive to query, the zero-marginal cost era is left further behind.

Investors value OpenAI like a tech darling: it is worth US$157 billion, going by a recent fundraising. They hope that thanks to the success of products like ChatGPT, it will become the next US$1 trillion tech giant. But the higher costs of state-of-the-art models, as well as other pressures from suppliers, distributors and competitors, suggests model-making may not confer the sort of monopoly-like powers enjoyed by the tech giants.

“One very important thing to understand about the future: the economics of AI are about to change completely,” said veteran AI researcher Francois Chollet on X, the day o3 was made public.

Chollet has helped drum up excitement about o3. In June, he launched a US$1 million prize for models that could run a gauntlet he had created five years earlier called the “Abstraction and Reasoning Corpus”, or ARC. It is a plethora of simple-looking visual-reasoning puzzles intended to be “easy for humans and impossible for modern AI”. (Easy being relative – in a study in September, humans recruited online scored 60-70 per cent on the test.) The prize wasn’t just challenging for its own sake. Chollet said beating an ARC task was a “critical” step towards building artificial general intelligence, meaning machines beating humans at many tasks.

Six months later, OpenAI aced the test. Its o3 model achieved a breakthrough score of 91.5 per cent. Its success in the challenge showed a step-change in AI’s ability to adapt to novel tasks, Chollet said. The new model is not just better; it is different. Like o1, it uses a “test-time compute” approach, which produces better results the more time that is spent on inference (when a trained AI model answers queries). Rather than simply producing an answer as quickly as it can spit it out, o3 is built to – in effect – think harder about the question.

Higher costs

That is where the higher costs come in. Chollet set a limit of US$10,000 on the amount that contestants can spend on computing power to answer the 400 questions in his challenge. When OpenAI put forward a model under the limit, it spent US$6,677 (about US$17 per question) to score 82.8 per cent. The score of 91.5 per cent, achieved by o3, came from blowing the budget. The company didn’t reveal the amount spent, but said that the expensive version of the process used 172 times the amount of “compute” as the cheaper approach – suggesting around US$3,000 to solve a single query that takes humans seconds.

To be sure, past AI models already challenged the low-marginal-cost norm of the software industry, because answering queries required substantially more processing power than using equivalent tools like a search engine. But the costs of building large language models and running them were small enough in absolute terms that OpenAI could still give away access for free.

With the latest models that is no longer the case. OpenAI restricts the “pro” version of the o1 model to users on its US$200-a-month subscription tier (and loses money, according to Sam Altman, its boss, because customers are spending more on queries than the company had budgeted for). Pierre Ferragu of New Street Research, reckons that OpenAI may charge as much as US$2,000 a month for full access to o3.

The power of such models relies on them bringing a version of the sector’s “scaling laws” closer to the end user. Until now, progress in AI had relied on bigger and better training runs, with more data and more computer power creating more intelligence. But once a model was trained, it was hard to use extra processing power well. As o3’s success on the ARC challenge shows, that is no longer the case. Scaling laws appear to have moved from training models to inference.

Such developments change the economics facing model-makers, such as OpenAI. The new models’ dependence on more processing power strengthens their suppliers, such as Nvidia, a maker of specialist AI chips. It also benefits the distributors of AI models, notably the cloud-service providers Amazon, Microsoft and Alphabet (Google’s parent company). And it may help justify the fortunes that these tech giants have invested in data centres because more inference will need more computing power. OpenAI will be squeezed from both sides.

Then there is competition. Google has released its own reasoning model, called Gemini 2.0 Flash, and other tech firms probably will too. Open-source models are expected to follow. Customers will be able to draw on multiple models from different providers. And although generative AI models may improve a little through their interactions with customers, they lack true network effects (where the service gets a lot better the more users it has), unlike the products Google and Facebook made in the past era.

High marginal costs mean the model-builders will have to generate meaningful value in order to charge premium prices. The hope, says Lan Guan of Accenture, a consultancy, is that models such as o3 will support AI agents that individuals and companies will use to increase their productivity. Even a high price for use of a reasoning model may be worth it compared with the cost of hiring, say, a fully fledged maths PhD. But that depends on how useful the models are.

Different use cases may also lead to more fragmentation. Jeremy Schneider of McKinsey says providing AI services to corporate customers will require models that are specialised for the needs of each enterprise, rather than general-purpose ones such as ChatGPT.

Instead of being dominated by one firm, some expect model-making to be more like a traditional oligopoly, with high barriers to entry but no stranglehold – or monopoly profits. For now, OpenAI is the leader, but one of its main rivals, Anthropic, is reportedly raising money at a US$60 billion valuation, and xAI, majority-owned by Elon Musk, is worth US$45 billion. That suggests there are high hopes for them, too. With o3, OpenAI has demonstrated its technical edge, but its business model remains untested.

©2025 The Economist Newspaper Limited. All rights reserved