OpenAI introduced GPT-5.2

OpenAI has unveiled its new flagship —GPT-5.2, releasing it amid increasing competition from Google. The company calls the new product the most advanced AI it has ever developed, and is targeting it at both developers and everyday professional use.

GPT-5.2 will be available to paid ChatGPT users and developers via API in three versions. Instant — a speed-optimized model for routine queries: information retrieval, texting and translation. Thinking — a model that does better with complex structured work: programming, analyzing long documents, math and planning. Pro — a senior version designed to maximize accuracy and reliability for difficult tasks.

«We built 5.2 to give people even more economic value» — said Fiji Simo, OpenAI’s product director, speaking to reporters. «It’s better at creating tables, assembling presentations, writing code, analyzing images, understanding long context, working with tools and linking together complex, multistep projects».

The release of GPT-5.2 comes amid an arms race with Google and its Gemini 3, leading the LMArena rankings in most tests (except coding — where it’s still stronger than Anthropic Claude Opus-4.5)

The Information reported earlier this month that CEO Sam Altman had issued an internal «red alert» — memo related to declining ChatGPT traffic and concerns that the service was losing audience to Google. The document called for a shift in priorities: postpone initiatives such as running ads and focus on improving the user experience.

The document said the company would be able to shift its priorities: postpone initiatives such as running ads and focus on improving the user experience.

GPT-5.2 — is OpenAI’s attempt to regain the lead, despite reports that some employees wanted to delay the release to improve the model. And while part of OpenAI’s plans have been focused on developing consumer scenarios with ChatGPT personalization, the launch of GPT-5.2 looks like an increased focus on enterprise customers.

The company is clearly betting on developers and an ecosystem of tools, aiming to become a mainstream platform for building AI applications. Earlier this week, OpenAI released new data showing a dramatic increase in enterprise use of its tools over the past year.

In the meantime, Gemini 3 is becoming more deeply integrated into Google’s ecosystem and Google Cloud, enabling multimodal and agent-based workflows. This week, Google unveiled MCP-managed servers that make it easier for agents to connect to services like Maps and BigQuery. (MCP — is a system of connectors between AI and external data and tools.)

OpenAI says GPT-5.2 shows new records in coding, math, science, computer vision, long context handling, and tool usage. The company says it will enable more robust agent-based workflows, production code, and complex systems that work with big data and real-world context.

The company says it will enable the creation of more robust agent-based workflows, production code, and complex systems that work with big data and real-world context.

These capabilities put GPT-5.2 in direct competition with Deep Think-mode Gemini 3, which Google is promoting as a major breakthrough in reasoning — especially in math, logic and science. OpenAI’s charts show that GPT-5.2 Thinking outperforms Gemini 3 and Anthropic Claude Opus 4.5 in almost every test, from SWE-Bench Pro (engineering problems) and GPQA Diamond (PhD level in the sciences) to abstract reasoning and pattern finding (ARC-AGI).

Study leader Edain Clark explained that high math scores — are not just about being able to solve equations. Mathematical reasoning — a measure of how well a model follows multi-step logic, holds numerical accuracy and avoids small errors that can accumulate.

«This is all critical for different tasks»— Clark said. «For example, financial modeling, forecasting, data analysis».

OpenAI product manager Max Schwarzer said GPT-5.2 greatly improves code generation and debugging and is able to explain complex math and logic step-by-step. Startups Windsurf and CharlieCode, he said, are capturing the «cutting-edge quality of agent programming» and efficiency gains in complex multistep tasks.

And he said that GPT-5.2 is a great way to improve the quality of agent programming» and increase efficiency in complex multistep tasks.

Besides programming, Schwarzer noted that GPT-5.2 Thinking responses contain 38% fewer errors than the previous version. This makes the model more reliable for everyday tasks: decision-making, research, and writing.

GPT-5.2 — is not a revolution, but rather an evolution of the ideas of the last two updates. GPT-5, released in August, was «a reboot» creating a single system with the ability to switch between fast mode and deep Thinking. GPT-5.1 in November made the model warmer, more conversational, and better suited for agent-based scripting and coding. GPT-5.2 reinforces these trends, becoming a more stable base for real-world use.

Stakes for OpenAI are at their highest right now. The company has invested about $1.4 trillion in AI infrastructure development over the coming years &#8212 a commitment it made when it still retained first mover advantage. Now that Google has accelerated markedly, however, that bet may explain Altman’s «red alert»

Focus on reasoning models also carries risks. Thinking and Deep Research use more computational resources than conventional chat models. By emphasizing these models, OpenAI risks getting caught in a vicious cycle: spending more resources to win ratings, and even more — to keep these expensive models running.”

At the same time, the OpenAI risks getting caught in a vicious cycle: spending more resources to win ratings, and even more — to keep these expensive models running.

Some reports suggest that OpenAI’s computing costs are already higher than the company has previously claimed. As TechCrunch wrote, most of the cost of the inferenc — computation required to run a trained model — OpenAI pays for in cash, not cloud credits. That means the cost of the computations exceeded what the partnerships could cover.

With all the attention to reasoning, there is no new image generation model in today’s release. In a memo «red alert» Altman allegedly cited image as a key priority — especially after the Google Nano Banana (a near-viral version of the Gemini 2.5 Flash Image model) received a lot of publicity after its release in August.

Last month, Google released the Nano Banana Pro (Gemini 3 Pro Image) — an improved version with more accurate text, enhanced world knowledge and frighteningly realistic photos. It’s also more deeply integrated into Google’s products, as demonstrated by the automatic generation of presentations through Google Labs Mixboard.

OpenAI reportedly plans to release the new model in January — with improved image generation, faster speed and an enhanced «personality». However, the company gave no official confirmation on Thursday.

OpenAI also said it is launching new security measures related to the use of AI in the context of mental health and age verification for teens, but those changes were not the centerpiece of the presentation.