Reasoning AI models have moved from a narrow technical idea to one of the most watched trends in generative AI. OpenAI’s o1 helped set off the shift, and rival labs quickly followed with their own attempts to make models that can work through harder tasks by checking their own answers along the way.
The result is a fast-moving race with real stakes. These systems may point toward a new stage of AI development, but the source article makes clear that their value, cost, reliability, and openness are still unsettled questions.
The rush after OpenAI’s o1
OpenAI’s o1 is described as a so-called reasoning model. After its release, competing AI labs began introducing their own reasoning systems, suggesting that the category had become a strategic priority across the industry.
In early November, DeepSeek, an AI research company funded by quantitative traders, launched a preview of its first reasoning algorithm, DeepSeek-R1. That same month, Alibaba’s Qwen team unveiled what it claims is the first “open” challenger to o1.
This wave is tied to a broader problem in generative AI: the old approach of simply scaling models through “brute force” is no longer producing the same improvements it once did. That creates pressure to find new techniques that can keep progress moving.
The business incentive is large. According to one estimate cited in the source, the global AI market reached $196.63 billion in 2023 and could be worth $1.81 trillion by 2030. In that context, a model category that promises better performance on difficult tasks is not just a technical story; it is a market story.
What makes reasoning models different
The key distinction is how these systems handle a task. Unlike most AI models, o1 and other reasoning models attempt to check their own work as they proceed. That process can help them avoid some failures that normally trip up models.
The tradeoff is time and compute. A model that spends more effort working through a problem may take longer to deliver an answer, and it may require more computing resources to run. That makes the reasoning model trend both promising and expensive.
OpenAI has claimed reasoning models can “solve harder problems” than earlier models and mark a step change in generative AI development. The company also envisions future reasoning models “thinking” for hours, days, or even weeks on end.
In that future, usage costs would rise, but OpenAI argues the potential payoff could include areas such as breakthrough batteries to new cancer drugs. The source article presents that vision as ambitious, while also showing that experts are divided on how much confidence to place in it today.
The cost problem is already visible
Reasoning AI models are not cheap to use. In OpenAI’s API, the company charges $15 for every ~750,000 words o1 analyzes and $60 for every ~750,000 words the model generates. The source says that is 6x the cost of OpenAI’s latest “non-reasoning” model, GPT-4o.
O1 is available in ChatGPT for free, but with limits. OpenAI also introduced a more advanced o1 tier, o1 pro mode, that costs $2,400 a year.
Guy Van Den Broeck, a professor of computer science at UCLA, told TechCrunch, “The overall cost of [large language model] reasoning is certainly not going down.” That statement captures one of the core concerns around this model category: better answers may come with higher bills.
For companies building AI products, that matters. If reasoning models need more compute and take longer to answer, the cost structure can shape where they are practical. They may be attractive for difficult problems where accuracy is valuable, while being harder to justify for everyday tasks where cheaper models are good enough.
Reliability remains an open question
The case for reasoning models depends on whether their extra work produces meaningfully better results. The source article suggests the answer is not yet settled.
Costa Huang, a researcher and machine learning engineer at the nonprofit org Ai2, notes that o1 isn’t a very reliable calculator. The article also says cursory searches on social media turn up a number of o1 pro mode errors.
Huang told TechCrunch, “These reasoning models are specialized and can underperform in general domains.” He added, “Some limitations will be overcome sooner than other limitations.”
Van den Broeck raises a deeper objection. He argues that reasoning models are not performing actual reasoning and are therefore limited in the kinds of tasks they can solve successfully. As he put it, “True reasoning works on all problems, not just the ones that are likely [in a model’s training data].”
That distinction is important. A model can appear stronger on certain tasks without possessing a general ability to reason across every kind of problem. The source article frames this as one of the main challenges still facing the field.
The transparency question
Ameet Talwalkar, an associate professor of machine learning at Carnegie Mellon, says the early reasoning models are “quite impressive.” But he also warns against accepting confident predictions about how far this approach will take the industry.
Talwalkar told TechCrunch that AI companies have financial incentives to present optimistic views of future capabilities. His concern is that the field could focus too narrowly on one paradigm while overlooking the need for concrete results.
He also worries that big labs may keep important advances closed. In his view, competitive secrecy can make it harder for the broader research community to examine and build on new ideas.
The likely path, based on the source, is continued investment and rapid development. OpenAI, DeepSeek, Alibaba, VCs, founders, and adjacent industries are all moving around the idea of reasoning AI. But the future of these models depends on more than hype: costs, reliability, openness, and real-world performance will decide how far the trend can go.