AI model race heats up with OpenAI, Google, DeepSeek releases
The GenAI race intensifies with rapid model releases from LLM players, but enterprises are less focused on incremental upgrades and more interested in agentic AI.
AI vendors in the U.S. and China continue to improve their generative AI models as they compete in a close race to grab enterprise attention.
This week, OpenAI, Google and DeepSeek introduced new updates to their AI models.
First, Chinese AI startup DeepSeek released an upgrade to its V3 model Monday. The new DeepSeek-V3-0324 launched on Hugging Face and includes improvements in reasoning and coding capabilities.
Hours after DeepSeek's release, OpenAI launched its latest image-generating model within GPT-4o. The model can accurately render text and follow prompts precisely to generate images.
Following OpenAI's release, Google launched Gemini 2.5, a "thinking model" designed to tackle more complex problems than previous models could.
There is a need to outperform competitors. It's not surely driven by this massive leap in performance, but it's driven by incremental improvement over your competition.
Lian Jye SuAnalyst, Omdia
The release of these three models within the span of 24 hours shows the speed of innovation in the GenAI market.
"The release cycle has shortened so much," said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget. "There is a need to outperform competitors. It's not surely driven by this massive leap in performance, but it's driven by incremental improvement over your competition."
DeepSeek V3 upgrade
With DeepSeek-V3-0324, the performance is more superior than the original release of V3, even though it is an open source non-reasoning model, Su said.
The fact that the model is open and released under an MIT license makes it one of the most permissive open source models on the market, said Bradley Shimmin, an analyst at The Futurum Group.
"It's going to put a lot of pressure on the other frontier model makers to do a better job of not just paying lip service to being open with their community licenses, but to employ real, free and open software licenses like MIT," Shimmin said.
The fact that the model was released on Hugging Face is also noteworthy, said Constellation Research analyst Andy Thurai.
"Hugging Face doesn't randomly host models, but they do testing and validation -- or at least used to -- before they can host a model," Thurai said.
He added that the performance of V3, along with the vendor's models being hosted on almost every hyperscaler and cloud provider, increases the vendor's chance of adoption.
While the model's openness and performance make it viable and competitive, geopolitical concerns might prevent enterprises from adopting it, said David Nicholson, an analyst at Futurum Group
"The concern that there is a global threat associated with this stuff is going to determine how widely this is adopted," Nicholson said.
While Nicholson's testing shows that the model is on par with other GenAI from OpenAI and Anthropic, when it comes to code debugging and web generation, enterprises might still be cautious due to geopolitical tensions.
"They are all very hesitant to employ something that relies on a model that the U.S. government has in its sights," Nicholson said. "Right now, many of the performance benefits V3 shows are a bit of an academic exercise. It pushes the frontiers of what is possible. It's healthy for competition."
OpenAI's 4o image generation
DeepSeek has undoubtedly increased the competitiveness among AI vendors.
While OpenAI's GPT-4o image generation model does not directly compete with V3, the timing of the release following the V3 upgrade shows that OpenAI is paying attention to DeepSeek's actions.
OpenAI said the image generation model leverages GPT-4o's knowledge base and chat context. It is trained on the distribution of online images and text and can blend precise symbols with imagery.
GPT-4o image generation can follow detailed prompts with attention to detail. It can also analyze and learn from user-uploaded images, integrating the details of those images to create and transform images convincingly, OpenAI said.
According to William McKeon-White, an analyst at Forrester Research, the intriguing part of 4o image generation is the information density.
"The biggest announcement here is the overall information density that each image can convey to the point where it's starting to get more alarming," McKeon-White said.
For example, one of the images OpenAI showcased is an image of a woman writing on a whiteboard wearing a shirt with an OpenAI logo. The model later generated another image of the woman high-fiving a photographer in a selfie.
"Not only is there proper rendering of human beings and proper rendering of text, but there's background reflection, there's proper demonstration understanding of occlusion, more or less of how those movements would create shadow and adjust that reflection and consistency in the objects that are reflected too in a way that I wasn't expecting," McKeon-White said.
"Looking quickly at any of these images, it would not be immediately apparent that they were AI-generated, except for specific details and sections," he continued.
He added that the ability to create visual noise and render imperfection is also getting better.
The leap toward more realistic image generation means there will need to be new ways to determine what is AI-generated and what is real.
"We need people to be much more skeptical of images they see on the internet," McKeon-White said. "It comes down to human education problem, which I think is difficult."
Despite its realism, OpenAI said the model still contains limitations it is working to address. The AI vendor also revealed that all generated images come with C2PA metadata to identify them as coming from GPT-4o. It will also block requests for child sexual abuse materials and sexual deepfakes. There are also restrictions for images that include real people.
GPT-4o image generation is available to ChatGPT Plus, Pro, Team and free users. It is also available in Sora and accessible through DALL-E GPT.
Gemini 2.5
Shortly after OpenAI's release of image generation in 4o, Google released Gemini 2.5 thinking models. These new models can reason through "thoughts" before responding, according to Google.
Gemini 2.5 Pro Experimental is available in Google AI Studio and the Gemini App, and will soon be available in Vertex AI.
Google said the model scored well on different AI benchmarks for advanced reasoning, and excels at creating visually compelling web apps and agentic code applications.
"The idea is to make it a lot more logical when it comes to its reasoning and ... a more structured way of processing information," Su said.
The model has multimodal capabilities and a long context window of 1 million tokens, with 2 million to be added soon.
Lack of differentiation
While the introduction of the three models is representative of the competitiveness in the AI market, it also shows a lack of differentiation, Su said.
"In many ways, it points toward the state of the market where ... there's not much of a new direction," he said.
The lack of differentiation is leading to model fatigue, and many experts are focusing less on the rapid release of new models and more on their application in areas like agentic AI.
For instance, Microsoft introduced on Tuesday two reasoning agents: Researcher and Analyst in Microsoft 365 Copilot. The agents analyze information in users' work data.
Researcher helps users tackle complex multi-step research at work, while Analyst provides insights from raw data, Microsoft said. Microsoft built the agents on OpenAI's o3-mini reasoning model.
For enterprises, analysts say it is better to focus on agentic applications and domain-specific use cases to maximize the return on their AI projects.
"A lot of enterprises do care about the amount of investments that they need to make for AI; they do care about their ROI," Su said. "Different use cases may require different types of optimization."
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
