OpenAI returns to Google with GPT-5.2 after ‘code red’ memo

OpenAI launched its newest frontier mannequin, GPT-5.2, on Thursday amid rising competitors from Google, pitching it as its most superior mannequin but and one designed for builders {and professional} on a regular basis use.

OpenAI’s GPT-5.2 is coming to paid ChatGPT customers and builders through the API in three flavors: Prompt, a speed-optimized mannequin for routine queries like info search, typing, and translation; Pondering, which excels in complicated structured work similar to coding, analyzing lengthy paperwork, arithmetic and planning; and the Professional, the high-end mannequin aimed toward delivering most accuracy and reliability for powerful issues.

“We designed 5.2 to unlock much more financial worth for individuals,” Fidji Simo, OpenAI’s chief product officer, stated Thursday throughout a convention name with reporters. “It is higher at creating spreadsheets, constructing displays, writing code, perceiving pictures, understanding lengthy context, utilizing instruments, after which connecting complicated multi-step tasks.”

GPT-5.2 lands in the course of an arms race with Google’s Gemini 3, which is main LMARena’s leaderboard in most benchmarks (besides encoding – which Anthropic’s Claude Opus-4.5 nonetheless has locked down).

Earlier this month, The Info reported that CEO Sam Altman issued an inner “code purple” memo to employees amid declining ChatGPT site visitors and issues that it’s dropping shopper market share to Google. Code Pink known as for a shift in priorities, together with stalling commitments like introducing adverts and as an alternative specializing in creating a greater ChatGPT expertise.

GPT-5.2 is OpenAI’s push to retake the lead, though some workers reportedly requested for the mannequin’s launch to be pushed again so the corporate might have extra time to enhance it. And regardless of indications that OpenAI will focus its consideration on shopper use circumstances by including extra customization and personalization to ChatGPT, the launch of GPT-5.2 seems to be set to extend its entrepreneurial alternatives.

The corporate is particularly focusing on builders and the instrumentation ecosystem, aiming to grow to be the default basis for constructing AI-powered functions. Earlier this week, OpenAI launched new knowledge displaying that enterprise use of its AI instruments has grown dramatically over the previous yr.

This comes as Gemini 3 is tightly built-in into Google’s product and cloud ecosystem for multimodal and agent workflows. Google this week launched MCP managed servers that make its Google and Cloud companies like Maps and BigQuery simpler to connect with brokers. (MCPs are the connectors between AI programs and knowledge and instruments.)

OpenAI says GPT-5.2 units new benchmarks in coding, arithmetic, science, imaginative and prescient, long-context reasoning, and gear use, which the corporate claims can result in “extra dependable agent workflows, production-scale code, and complicated programs that run on massive contexts and real-world knowledge.”

These capabilities put it in direct competitors with Gemini 3’s Deep Suppose mode, which has been touted as a serious reasoning breakthrough aimed toward math, logic and science. On OpenAI’s benchmark chart, GPT-5.2 Pondering comes out forward of Gemini 3 and Anthropic’s Claude Opus 4.5 in practically each reasoning take a look at listed, from real-world software program engineering duties (SWE-Bench Professional) to PhD-level scientific information (GPQA Diamond) and summary sample reasoning (AGIARC).

Analysis chief Aidan Clark stated stronger ends in maths aren’t nearly fixing equations. Mathematical reasoning, he defined, is an indicator of whether or not a mannequin can observe multi-step logic, hold numbers constant over time, and keep away from refined errors that may compound over time.

“These are all properties that actually matter throughout a variety of various workloads,” Clark stated. “Issues like monetary modeling, forecasting, knowledge evaluation.”

In the course of the convention, OpenAI product supervisor Max Schwarzer stated that GPT-5.2 “makes substantial enhancements to code era and debugging” and might step by means of complicated math and logic step-by-step. Coding startups like Windsurf and CharlieCode, he added, report “state-of-the-art agent coding efficiency” and measurable features in complicated, multi-step workflows.

Past coding, Schwarzer stated GPT-5.2’s considering responses comprise 38% fewer errors than its predecessor, making the mannequin extra dependable for on a regular basis decision-making, analysis and writing.

GPT-5.2 seems to be much less of a reinvention and extra of a consolidation of the final two OpenAI enhancements. GPT-5, which dropped in August, was a reset that laid the groundwork for a unified system with a router to change between a quick default sample and a deeper “Pondering” mode. November’s GPT-5.1 centered on making that system hotter, extra conversational, and extra appropriate for agent and coding duties. The most recent mannequin, GPT-5.2, seems to allow all of those developments, making it a extra dependable base for manufacturing use.

For OpenAI, the stakes have by no means been greater. The corporate has made commitments to the tune of $1.4 trillion to construct AI infrastructure over the subsequent few years to help its development — commitments it made when it nonetheless had first-mover benefit amongst AI corporations. However now that Google, which lagged behind at first, is transferring ahead, that wager could possibly be the one which pushes Altman’s “code purple.”

OpenAI’s renewed deal with reasoning fashions can also be a harmful bend. The programs behind its Pondering and Deep Analysis modes are dearer to run than customary chatbots as a result of they chew up extra calculations. By doubling that sort of mannequin with GPT-5.2, OpenAI can create a vicious cycle: spend extra on computation to win the leaderboard, then spend much more to maintain these high-cost fashions working at scale.

OpenAI is already stated to be spending extra on computing than it beforehand allowed. As TechCrunch not too long ago reported, most of OpenAI’s inference bills — the cash it spends on computation to run a skilled AI mannequin — are being paid in money fairly than by means of cloud loans, suggesting that the corporate’s computational prices have grown past what partnerships and loans can subsidize.

In the course of the name, Simo advised that as OpenAI scales, it is ready to provide extra services to generate extra income to pay for the extra accounts.

“However I feel it is vital to place it within the huge arc of effectivity,” Simo stated. “You are getting, at present, much more intelligence for a similar quantity of computation and the identical quantity of {dollars} as you probably did a yr in the past.”

For all its deal with reasoning, one factor lacking from at present’s presentation is a brand new picture generator. Altman reportedly stated in his code purple memo that picture era can be a prime precedence transferring ahead, particularly since Google’s Nano Banana (the nickname for Google’s Gemini 2.5 Flash Picture mannequin) had a viral second after its launch in August.

Final month, Google launched Nano Banana Professional (aka Gemini 3 Professional Picture), an upgraded model with even higher textual content rendering, world information and an eerie, real-life, unedited vibe to its photographs. It additionally integrates higher into Google merchandise, as demonstrated over the previous week because it seems in instruments and workflows like Google Labs Mixboard for automated presentation era.

OpenAI reportedly plans to launch one other new mannequin in January with higher pictures, improved velocity and higher character, although the corporate didn’t affirm these plans on Thursday.

OpenAI additionally stated Thursday that it’s introducing new safety measures round the usage of psychological well being and age verification for youngsters, however did not spend a lot time outlining these adjustments.

This text has been up to date with extra details about the state of OpenAI’s computational effectivity.

Do you will have a delicate tip or confidential paperwork? We’re reporting on the inside workings of the AI ​​trade – from the businesses shaping its future to the individuals affected by their choices. Contact Rebecca Bellan at rebecca.bellan@techcrunch.com or Russell Brandom at russell.brandom@techcrunch.com. For safe communication, you possibly can contact them through Sign at @rebeccabellan.491 and russellbrandom.49.