OpenAI will make its Deep Research model available to users on ChatGPT’s free tier soon, according to an X post by CEO Sam Altman. Deep Research is the colloquial name for large language models that use a “test-time” computing method to produce in-depth, detailed research reports. These reports can take anywhere from 30 minutes up to multiple hours to produce, and technologists have said Deep Research could replace humans in jobs like business analysis or finance. Of course, it is important to take these claims with a grain of salt.
Altman also said that OpenAI will consolidate its numerous different AI models. Many people have complained that everyday users do not understand the difference between its varying models and what they offer. “We hate the model picker as much as you do and want to return to magic unified intelligence,” Altman said. “We will next ship GPT-4.5, the model we called Orion internally, as our last non-chain-of-thought model. After that, a top goal for us is to unify o-series models and GPT-series models by creating systems that can use all our tools, know when to think for a long time or not, and generally be useful for a very wide range of tasks.”
At some point in the near future, users on the free tier will gain “unlimited” access to a new model that includes o3 built-in, though Altman says that paid users will get a higher level of intelligence, whatever that means.
Currently, Deep Research is only available to ChatGPT users who pay for its premium, $200-a-month tier. It has always been expected that OpenAI would make its bleeding edge models available to more users as it scales up its server infrastructure, and Altman has confirmed as much in another recent X post. In his typical fashion of writing everything in lowercase like a Gen Z girl, Altman wrote, “i think we are going to initially offer 10 uses per month for ChatGPT plus and 2 per month in the free tier, with the intent to scale these up over time.”
Large language models like ChatGPT are, in essence, probability machines, or a really powerful autocomplete. They ingest billions of pieces of human-written text and are capable of generating new texts with the verisimilitude of something written by a person. They do not actually “think” like a human does, however, and companies like OpenAI have had to find new techniques to improve their accuracy and capabilities. Test-time thinking forces the language model to fact-check itself before answering a question. For instance, if a user asks the model how much it would cost to replace every Uber in the United States with Waymos, the model would break the query down into multiple related questions—how many Ubers are on the road today, how much each Waymo vehicle costs, and so on and so forth.
The AI industry has been throwing spaghetti at the wall for some time now, trying to find the unlock that will make language models truly capable of replacing humans in a variety of tasks. Deep Research is the latest such attempt; another is AI “agents” that can take control of a user’s computer to actually complete tasks, like booking a flight. Agents could be particularly useful for demographics like the elderly, who may not be tech-savvy and could ask AI to help them navigate a computer. Early users of agents like OpenAI’s Operator say they are very slow and error-prone, however.
As impressive as Deep Research might be at producing lengthy research reports, the accuracy of AI language models remains a problem. During the Super Bowl, Google aired a commercial promoting its Gemini chatbot that included inaccurate information. X users have noticed if you ask ChatGPT what players are on various NFL teams, the chatbot will return incorrect information even though ChatGPT has access to rosters from NFL.com. And of course, Apple was forced to neuter its AI-based notification summaries after producing incorrect summaries of BBC news notifications.
Some proponents of AI have argued that sounding correct often means the text is correct, but a major problem with AI models seems to be that its users assume whatever they produce is correct even when it is not. Google repeatedly presenting incorrect information during demos of its AI is emblematic of this problem. It seems like leaders at the tech giants just glaze over and skim whatever their AI models produce without doing any work to fact-check it.
Meanwhile, OpenAI is dealing with other problems, namely a long-running legal battle with Elon Musk, which escalated this week after the early backer of OpenAI made a hostile $97 billion offer to acquire the AI company out from under Altman. Though Musk was able to take control of Twitter despite all odds, this situation is considerably different, as OpenAI is private and remains a non-profit. Although OpenAI is trying to convert into a for-profit company, for now it does not have any fiduciary duty to investors and likely does not have to consider acquisition offers.