OpenAI fires back at Google with GPT-5.2 after ‘code red’ memo

OpenAI launched its latest frontier model, GPT-5.2, connected Thursday amid expanding contention from Google, pitching it arsenic its astir precocious exemplary yet and 1 designed for developers and mundane nonrecreational use. 

OpenAI’s GPT-5.2 is coming to ChatGPT paid users and developers via the API successful 3 flavors: Instant, a speed-optimized exemplary for regular queries similar information-seeking, writing, and translation; Thinking, which excels astatine analyzable structured enactment similar coding, analyzing agelong documents, math, and planning; and Pro, the top-end exemplary aimed astatine delivering maximum accuracy and reliability for hard problems. 

“We designed 5.2 to unlock adjacent much economical worth for people,” Fidji Simo, OpenAI’s main merchandise officer, said Thursday during a briefing with journalists. “It’s amended astatine creating spreadsheets, gathering presentations, penning code, perceiving images, knowing agelong context, utilizing tools and past linking complex, multi-step projects.”

GPT-5.2 lands successful the mediate of an arms contention with Google’s Gemini 3, which is topping LMArena’s leaderboard crossed astir benchmarks (apart from coding – which Anthropic’s Claude Opus-4.5 inactive has connected lock).

Early this month, The Information reported that CEO Sam Altman released an interior “code red” memo to unit amid ChatGPT postulation decline and concerns that it is losing user marketplace stock to Google. The codification reddish called for a displacement successful priorities, including stalling connected commitments similar introducing ads and alternatively focusing connected creating a amended ChatGPT experience. 

GPT-5.2 is OpenAI’s propulsion to reclaim leadership, adjacent arsenic immoderate employees reportedly asked for the exemplary merchandise to beryllium pushed backmost truthful the institution could person much clip to amended it. And contempt indications that OpenAI would absorption its attraction connected user usage cases by adding much personalization and customization to ChatGPT, the motorboat of GPT-5.2 looks to beef up its endeavor opportunities. 

The institution is specifically targeting developers and the tooling ecosystem, aiming to go the default instauration for gathering AI-powered applications. Earlier this week, OpenAI released caller data showing endeavor usage of its AI tools has surged dramatically implicit the past year. 

This comes arsenic Gemini 3 has go tightly integrated into Google’s merchandise and unreality ecosystem for multimodal and agentic workflows. Google this week launched managed MCP servers that marque its Google and Cloud services similar Maps and BigQuery easier for agents to plug into. (MCPs are the connectors betwixt AI systems and information and tools.)

OpenAI says GPT-5.2 sets caller benchmark scores successful coding, math, science, vision, long-context reasoning, and tool-use, which the institution claims could pb to “more reliable agentic workflows, production-grade code, and analyzable systems that run crossed ample contexts and real-world data.”

Those capabilities enactment it successful nonstop contention with Gemini 3’s Deep Think mode, which has been touted arsenic a large reasoning advancement targeting math, logic, and science. On OpenAI’s ain benchmark chart, GPT-5.2 Thinking edges retired Gemini 3 and Anthropic’s Claude Opus 4.5 successful astir each listed reasoning test, from real-world bundle engineering tasks (SWE-Bench Pro) and doctoral-level subject cognition (GPQA Diamond) to abstract reasoning and signifier find (ARC-AGI suites). 

Research pb Adain Clark said that stronger mathematics scores aren’t conscionable astir solving equations. Mathematical reasoning, helium explained, is simply a proxy for whether a exemplary tin travel multi-step logic, support numbers accordant implicit time, and debar subtle errors that could compound implicit time. 

“These are each properties that truly substance crossed a wide scope of antithetic workloads,” Clark said. “Things similar fiscal modeling, forecasting, doing an investigation of data.”

During the briefing, OpenAI merchandise pb Max Schwarzer said GPT-5.2 “makes important improvements to codification procreation and debugging” and tin locomotion done analyzable mathematics and logic step-by-step. Coding startups similar Windsurf and CharlieCode, helium added, study “state-of-the-art cause coding performance” and measurable gains connected analyzable multi-step workflows.

Beyond coding, Schwarzer said that GPT-5.2 Thinking responses incorporate 38% less errors than its predecessor, making the exemplary much dependable for day-to-day decision-making, research, and writing. 

GPT-5.2 appears to beryllium little a reinvention and much of a consolidation of OpenAI’s past 2 upgrades. GPT-5, which dropped successful August, was a reset that laid the groundwork for a unified strategy with a router to toggle the exemplary betwixt a accelerated default exemplary and a deeper “Thinking” mode. November’s GPT-5.1 focused connected making that strategy warmer, much conversational, and amended suited to agentic and coding tasks. The latest model, GPT-5.2, seems to crook up the dial connected each of those advancements, making it a much reliable instauration for accumulation use. 

For OpenAI, the stakes person ne'er been higher. The institution has made commitments to the tune of $1.4 trillion for AI infrastructure buildouts implicit the adjacent fewer years to enactment its maturation – commitments it made erstwhile it inactive had the first-mover vantage among AI companies. But present that Google, which lagged down astatine the start, is pushing ahead, that stake mightiness beryllium what’s driving Altman’s ‘code red.’ 

OpenAI’s renewed absorption connected reasoning models is besides a risky flex. The systems down its Thinking and Deep Research modes are much costly to tally than modular chatbots due to the fact that they chew done much compute. By doubling down connected that benignant of exemplary with GPT-5.2, OpenAI whitethorn beryllium mounting up a vicious cycle: walk much connected compute to triumph the leaderboard, past walk adjacent much to support those high-cost models moving astatine scale.

OpenAI is already reportedly spending much connected compute than it had antecedently fto on. As TechCrunch reported recently, astir of OpenAI’s inference walk – the wealth it spends connected compute to tally a trained AI exemplary – is being paid successful currency alternatively than done unreality credits, suggesting the company’s compute costs person grown beyond what partnerships and credits tin subsidize.

For each its absorption connected reasoning, 1 happening that’s absent from today’s motorboat is simply a caller representation generator. Altman reportedly said successful his codification reddish memo that representation procreation would beryllium a cardinal precedence moving forward, peculiarly aft Google’s Nano Banana (the nickname for Google’s Gemini 2.5 Flash Image model) had a viral infinitesimal pursuing its August release.

Last month, Google launched Nano Banana Pro (AKA Gemini 3 Pro Image), an upgraded mentation with adjacent amended substance rendering, satellite knowledge, and an eerie, real-life, unedited vibe to its photos. It besides integrates amended crossed Google’s products, arsenic demonstrated implicit the past week arsenic it pops up successful tools and workflows similar Google Labs Mixboard for automated presumption generation.

OpenAI reportedly plans to merchandise different caller exemplary successful January with amended images, improved speed, and amended personality, though the institution didn’t corroborate these plans Thursday.

OpenAI besides said Thursday it’s rolling retired caller information measures astir intelligence wellness usage and property verification for teens, but didn’t walk overmuch of the motorboat pitching those changes.