Asger Risom
OpenAI has released a new flagship AI model amid growing pressure from rivals and mounting scrutiny over safety and performance.
OpenAI has released a new flagship AI model amid growing pressure from rivals and mounting scrutiny over safety and performance.
The company says GPT-5.2 reflects months of development, even as its timing fuels debate about whether OpenAI is losing ground in the AI race.
Safety comes first
OpenAI placed heavy emphasis on safety alongside the model’s launch, as it continues to face lawsuits over ChatGPT’s role in mental health crises.
Fidji Simo, OpenAI’s CEO of applications, said the company has improved what it calls “safe completions,” designed to provide helpful responses without worsening emotional distress.
“On the safety side, as you saw through the benchmarks, we are improving on pretty much every dimension of safety,” Simo said, adding that models are released only when OpenAI is confident its safeguards meet internal standards.
The rollout coincided with a new lawsuit tied to a murder-suicide in Connecticut. OpenAI called the case “incredibly heartbreaking” and said it is continuing to improve ChatGPT’s ability to recognize distress and guide users toward real-world support.
Racing the market
GPT-5.2 arrives less than a month after OpenAI released GPT-5.1, underscoring the pace at which major AI labs are shipping updates.
The company is under pressure from Google and Anthropic, both of which have gained momentum with enterprises, particularly in software development and technical workflows.
In a briefing with reporters, OpenAI said the new model delivers stronger performance across professional “knowledge work,” including law, finance, and accounting, as well as coding and mathematical reasoning.
Not a reaction launch
Simo pushed back on suggestions that GPT-5.2 was rushed out in response to Google’s Gemini 3 Pro model, which prompted OpenAI CEO Sam Altman to declare a company-wide “code red” last month.
“I would say that [the code red] helps with the release of this model, but that’s not the reason it is coming out this week in particular,” she said. “It has been in the works for a while.”
She added that GPT-5.2 had been built over many months and tested by select customers before Altman’s internal directive.
Enterprise testing
OpenAI said GPT-5.2 had been in the hands of “alpha customers” for several weeks, including Harvey, Notion, Box, Shopify, and Zoom.
Those users reported stronger performance in using software tools and writing and debugging code. Coding benchmarks matter more now than general chat quality because enterprises increasingly judge models on whether they can replace or accelerate real engineering work, not just converse fluently.
OpenAI hopes the gains will help it claw back share in a segment where Anthropic’s Claude models have proven popular.
Measuring progress
On OpenAI’s GDPval benchmark, which tests complex professional tasks, GPT-5.2 met or exceeded human expert performance 70.9% of the time.
That marks a substantial jump from earlier OpenAI models, according to the company, though executives declined to detail the exact training changes behind the improvement.
Aidan Clark, OpenAI’s vice president of research for training, said the company made gains across pretraining and other stages, but did not elaborate further.
Sources: OpenAI, Fortune, The Information
