OpenAI reasoning model shows an image shift for GenAI vendor

Developers and deep research

Compared with o1, o3-mini supports more developer features such as function calling, structured output and developer messages. Function calling is the ability of a large language model (LLM) to identify a predetermined function to interface with external tools or data sources; structured output is when an LLM generates outputs in an existing format such as JSON; developer messages are instructions to an LLM from a developer.

The o3-mini model is available for ChatGPT Plus, Team and Pro users. Enterprise users will gain access this month. Users of the free ChatGPT model can try it by selecting 'Reason' in the message composer.

On Sunday, two days after releasing o3-mini, OpenAI also launched deep research in ChatGPT, a new agentic mode that conducts research on the internet for complex tasks.

The deep research feature is for users who perform knowledge work in industries such as finance, science, policy and engineering. The AI agent works within ChatGPT to complete multi-step research, which takes five to 30 minutes, OpenAI said. The capability is now available to Pro users and will soon be available to Plus and Team users.

We're seeing a lot more capabilities when we give models time to think, and also when we give them multimodal capabilities such as being able to see and type.

The independent generative AI vendor released the updated reasoning model and deep research capability amid continuing excitement -- and skepticism -- about DeepSeek. The market ferment led several other top generative AI vendors, including Ai2, Mistral and Chinese tech giant Alibaba, to quickly release new models of their own that use reinforcement learning and time-test scaling.

Time-test scaling is a technique that enables models to receive extra processing time to produce better results.

"We're seeing a lot more capabilities when we give models time to think, and also when we give them multimodal capabilities such as being able to see and type," said Omdia analyst Bradley Shimmin.

Changing perceptions

That DeepSeek, using far less compute power than established generative AI vendors and at far lower cost, was able to train its new V3 model using time-test scaling indicated that OpenAI's image as the leading generative AI innovator is changing, some observers said. Some have speculated that DeepSeek has not been fully transparent about how many compute resources and AI chips it used in building V3.

While the vendor has been first in the market with some technologies, it has lagged in others.

For example, its deep research capability is similar to rival Anthropic's computer use tool, introduced last October.

And it also appears that OpenAI is later to the market with an agentic capability that uses computers on behalf of users. However, Shimmin said OpenAI is ahead with other capabilities. For instance, in October the vendor released Swarm, a system that helps AI agents collaborate and perform more complex tasks.

As for o3-mini and the deep research function, OpenAI continues to show how capable its models are in science, technology and engineering applications, said Dion Hinchcliffe, an analyst at Futurum Group.

"Until recently, OpenAI was kind of unchallenged in that regard," he said, adding that DeepSeek has shifted that perception.

Meanwhile, OpenAI has set o3-mini apart from other LLMs because o3-mini also includes structured outputs and messages, which let developers provide differentiated outputs. Developers can now ask the model to provide outputs in a programming format like JSON, instead of only human-like output.

A major difference between o3-mini and the original o1 model is that o3-mini fact-checks itself, reducing hallucinations. However, this process causes the model to be slower in producing outputs. The slower output likely will not deter enterprises, though, said Andy Thurai, an analyst with Constellation Research.

"Given the cost efficiency, performance, and more developers-friendly approach, [o3] should gain much traction within both innovators and enterprises," Thurai said.

Open source and changing strategy

But the fast-proliferating advancements in the open source community, especially in light of DeepSeek's instant success when it became the most downloaded app on the Apple Store last month, may be a big challenge for OpenAI.

"Open source models are catching up and have been catching up steadily for a year," Hinchcliffe said.

While OpenAI has a history as an AI innovator, DeepSeek's open source approach is good for generative AI users, he said.

"It gives DeepSeek the short-term edge and makes people trust it," Hinchcliffe said.

The recent moves by OpenAI also give insight into the AI strategy the vendor is taking toward its enterprise customers, Shimmin said.

"I think OpenAI is much more of a data play than an enterprise play," he said. The vendor wants to position itself more like Meta and TikTok, which gain revenues from the data they gather from others using its system, he added. OpenAI's close partnership with Microsoft also supports the generative AI vendor's strategy of trying to capture data to build data.

OpenAI also revealed on Monday that it is partnering with SoftBank Group to establish a venture company in Japan to market enterprise AI technologies.

OpenAI did not immediately respond to a request for comment for this story.

Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.