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

OpenAI launched its latest frontier model, GPT-5.2, on Thursday amid growing competition from Google, making it its most advanced model yet and designed for developers and everyday professional use.

OpenAI’s GPT-5.2 comes to paid ChatGPT users and developers via the API in three flavors: Instant, a speed-optimized model for routine queries like information searching, writing, and translation; Thinking, which excels at complex structured work such as programming, analyzing long documents, mathematics, and planning; and Pro, the cutting-edge model that aims to provide maximum accuracy and reliability for difficult problems.

“We designed 5.2 to unlock more economic value for people,” Viji Simo, chief product officer at OpenAI, said Thursday during a briefing with reporters. “She’s better at creating spreadsheets, creating presentations, writing code, making sense of images, understanding long-form context, using tools and then connecting complex, multi-step projects.”

GPT-5.2 is in the middle of an arms race with Google’s Gemini 3, which tops the LMArena leaderboard across most benchmarks (apart from programming – which Anthropic’s Claude Opus-4.5 still has under lock and key).

Early this month, Information pIt has been reported CEO Sam Altman issued an internal “code red” memo to employees amid this Low ChatGPT traffic And fears of losing consumer market share to Google. Code Red called for a change in priorities, including stalling on commitments like serving ads and instead focusing on creating a better ChatGPT experience.

GPT-5.2 is a push for OpenAI to regain leadership, even as some employees It is said Request to postpone the release of the model so that the company can have more time to improve it. Despite indications that OpenAI will focus its attention on consumer use cases by adding more customization and personalization to ChatGPT, the launch of GPT-5.2 looks to boost its enterprise opportunities.

The company specifically targets developers and the tool ecosystem, with the goal of becoming the default foundation for building AI-powered applications. Earlier this week, OpenAI has released new data The exposure of organizations using their own AI tools has risen dramatically over the past year.

This comes as Gemini 3 becomes more closely integrated into Google’s product cloud ecosystem for multimedia and proxy workflows. Google launched this week Managed MCP servers This makes Google and cloud services like Maps and BigQuery easier for agents to connect to. (MCPs are the links between AI systems, data, and tools.)

OpenAI says GPT-5.2 sets new benchmark scores in programming, mathematics, science, vision, long-contextual reasoning, and tool use, which the company claims could lead to “more reliable agent workflows, production-level code, and complex systems operating across large contexts and real-world data.”

These capabilities put it in direct competition with Gemini 3’s Deep Think mode, which is described as a major logic advance targeting math, logic, and science. On OpenAI’s benchmark chart, GPT-5.2 Thinking outperforms Gemini 3 and Anthropic’s Claude Opus 4.5 on almost every reasoning test listed, from real-world software engineering tasks (SWE-Bench Pro) and doctoral-level scientific knowledge (GPQA Diamond) to abstract reasoning and pattern detection (ARC-AGI sets).

Research leader Aidan Clarke said strong results in mathematics were not just about solving equations. He explained that mathematical logic is a proxy for whether a model can follow multi-step logic, maintain consistency of numbers over time, and avoid subtle errors that can worsen over time.

“These are all really important characteristics across a wide range of different workloads,” Clark said. “Things like financial modeling, forecasting, and doing data analysis.”

During the press conference, Max Schwarzer, OpenAI’s product lead, said GPT-5.2 “makes significant improvements to code generation and debugging” and can handle complex arithmetic and logical operations step-by-step. Coding startups, such as Windsurf and CharlieCode, report “state-of-the-art proxy coding performance” and measurable gains in complex, multi-step workflows, he added.

Beyond programming, Schwarzer said GPT-5.2’s thinking responses contain 38% fewer errors than its predecessor, making the model more reliable for decision-making, research, and everyday writing.

GPT-5.2 appears to be less of a reinvention and more of an amalgamation of the last two OpenAI upgrades. GPT-5, dropped in August, was a reset that laid the foundation for a unified system with a router to switch the model between the fast default model and a deeper “thinking” mode. November’s GPT-5.1 focused on making this system warmer, more conversational, and more suitable for proxying and programming tasks. The latest model, GPT-5.2, seems to improve on all of these advances, making it a more reliable foundation for production use.

For OpenAI, the stakes have never been higher. The company has made commitments worth $1.4 trillion to build out AI infrastructure over the next few years to support its growth — commitments it made when it still had a first-mover advantage among AI companies. But now that Google, which initially lagged behind, is moving forward, this bet may be what prompts Altman’s “code red.”

OpenAI’s renewed focus on inference models is also a risky flexibility. The systems behind the deep reasoning and searching modes are more expensive to run than standard chatbots because they consume more computation. By doubling down on this type of model with GPT-5.2, OpenAI may have created a vicious cycle: spend more on compute to win the leaderboard, then spend more to keep these high-cost models running at scale.

OpenAI is reportedly already spending more on compute than it previously let on. like TechCrunch reported Recently, most of OpenAI’s inference spending — the money it spends on compute to run a trained AI model — is paid in cash rather than through cloud credits, suggesting that the company’s compute costs have grown beyond what partnerships and credits can support.

During the call, Simo suggested that as OpenAI expands, it is able to offer more products and services to generate more revenue to pay for additional computing.

“But I think it’s important to put that in the big bracket of efficiency,” Simo said. “You get a lot more intelligence today for the same amount of computing and the same amount of dollars that you did a year ago.”

Despite its focus on logic, one thing absent from today’s launch is a new image generator. Altman reportedly said in his code-red memo that generating images will be a major priority going forward, especially after Google’s Nano Banana (the alias for Google’s Gemini 2.5 Flash Image model) had a major viral moment following its release in August.

Last month, Google launched the Nano Banana Pro (also known as Gemini 3 Pro Image), an upgraded version that features better text display, global knowledge, and Strange, realistic and unedited atmosphere To her pictures. It also integrates better across Google products, as demonstrated over the past week as it appeared in tools and workflows like Google Labs Mixboard for automatically creating presentations.

OpenAI is said to be planning to release another new model in January with better visuals, improved speed, and better personality, though the company did not confirm those plans on Thursday.

OpenAI also said Thursday that it is rolling out new safety measures around the use of mental health and age verification for teens, but it did not spend much of the launch time touting those changes.

This article has been updated with more information about the state of compute efficiency in OpenAI.

Got sensitive advice or confidential documents? We report on the inner workings of the AI ​​industry – from the companies shaping its future to the people affected by its decisions. Connect with Rebecca Bellan on rebecca.bellan@techcrunch.com Or email Brandom at russell.brandom@techcrunch.com. For secure communication, you can contact them via Signal at @rebeccabellan.491 And Russell Brandom.49.