OpenAI unveiled its latest frontier model, GPT-5.2, on Thursday as it faces growing competition from Google, which introduced the model as its most advanced yet, built for developers and everyday professionals.
OpenAI’s GPT-5. 2 is rolling out to ChatGPT paid users and developers through the API in three flavors: Instant, a quicker model designed for easy questions such as info lookup, writing, and translation; Thinking, better suited for complex work structures like coding or long documents analyzing, math and planning; and Pro – the high-end model hoping to provide pinpoint accuracy and higher reliability on complex problems.
“We built 5.2 to unlock additional economic value for people,” Fidji Simo, OpenAI’s chief product officer, said on Thursday during a briefing with reporters. “It does better when making spreadsheets, crafting presentations, coding, processing images and knowing long context, using tools or connecting up to complex multi-step projects.”
GPT-5.2 lands itself in the thick of an arms race right now, with Google’s Gemini 3 leading the pack on LMArena’s leaderboard on nearly every benchmark (except coding, which Anthropic’s Claude Opus-4.5 still has on lock).
Earlier this month, The Information reported that CEO Sam Altman sent an internal “code red” memo to employees, citing declining ChatGPT traffic and fears that the company is ceding consumer market share to Google. The code red prompted a change in priorities, delaying commitments such as enforcing ads to focus on creating a better ChatGPT experience.
GPT-5.2 is a gambit by OpenAI to reassert the virtues of dominance, despite reports that some employees seek to delay the release of such massive models so they have more time to improve them. Yet while all the signs had pointed to OpenAI moving to support consumer-centric use cases by bringing more personalisation and customisation to ChatGPT, GPT-5 exceeded even those expectations. 2, seeks to bolster enterprise prospects.
The company is going after the developer and tooling ecosystem, aiming to be the de facto platform on which companies use AI as the pillar of application development. OpenAI, an artificial intelligence research lab in San Francisco, said earlier this week that enterprise use of its AI tools has grown “by orders of magnitude” over the past year.
This is on the heels of Gemini 3 being more deeply integrated into Google’s product and cloud environment to support multimodal and agentic workflows. This week, Google began rolling out managed MCP servers that make it easier for agents to plug into its various services, including Maps and BigQuery, without relying on Google’s infrastructure. (MCPs are the intermediary between AI systems and data and tools.)
OpenAI says GPT-5. 2 sets new record scores in coding, math, science, vision, long-context reasoning and tool use, which the company says would make for “more reliable agentic workflows, production-grade code and complex systems that span many contexts and real-world data.”
Those abilities put it head-to-head with Gemini 3’s Deep Think mode, which has been advertised as a serious upgrade on reasoning skills covering math, logic and science. When run on OpenAI’s own benchmark chart, GPT-5. 2 Thinking loses to Gemini 3 and Anthropic’s Claude Opus 4.5 on almost all of the reasoning tests for which results are publicly available, across real-world software engineering (SWE-Bench Pro) and doctoral-level scientific knowledge (GPQA Diamond), abstract reasoning and pattern discovery tasks (ARC-AGI suites).
One leader of the research, Aidan Clark, said that higher levels of math are not simply about solving equations. Mathematical reasoning, he said, is a stand-in for whether a model can do multi-step logic and still be consistent with numbers over the long haul, avoiding the kinds of subtle errors that multiply over time.
“These are all properties that matter so much in a bunch of different workloads,” Clark said. “Things like financial modelling, forecasting, analysing data.”
OpenAI product lead Max Schwarzer mentioned GPT-5 in the briefing. 2 “provides significant enhancements to code generation and debugging, and can step through complicated math and logic one by one. Coding startups like Windsurf and CharlieCode, he said, have reported “state-of-the-art agent coding performance” and measurable improvement on complex multi-step workflows.
Beyond programming, Schwarzer added that GPT-5.2 Thinking responses have 38% fewer mistakes than the previous version, making the model more reliable for everyday decision-making, research, and writing.
GPT-5. 2 comes off as not so much a reinvention as a consolidation of OpenAI’s past two upgrades. The GPT-5 drop in August was a reset, setting the groundwork for a single system with an associated router to toggle the model between a fast default mode and a deeper “Thinking” mode. November’s GPT-5. One was about making this system warmer, more conversational, and tailored to agentic and coding-specific tasks. The latest model, GPT-5.2, seems to dial up all those improvements, rendering it a sounder foundation for production use.
For OpenAI, the risk is higher than ever. Into that, the company has pledged some $1.4 trillion to spend on AI infrastructure rollouts over the next few years to buttress its growth — pledges it made while still far ahead of others as a first mover among AI companies. But with Google, which was initially behind but is now gaining steam, that bet could be fueling Altman’s “code red.”
OpenAI’s pivot to reasoning models is also something of a risky flex. The systems behind its Thinking and Deep Research modes require more expensive computing than standard chatbots, because they chew through it. By doubling down on that model form with GPT-5. 2, OpenAI may also be creating a feedback loop: spend more to win the leaderboard and then even more to sustain running those high-cost models at scale.
OpenAI has already said it is spending more on compute than it previously disclosed. As TechCrunch recently reported, most of OpenAI’s inference spend — the money it shells out to run a trained AI model on compute (more vernacular here) — is paid for with cash rather than cloud credits, suggesting that OpenAI’s compute bills have grown beyond what partnerships and credits can offset.
Simo also mentioned during the call that as OpenAI scales, it can provide more products and services to generate more revenue, which they can use to pay for additional compute.
“But I think it’s important to put that into the grand arc of efficiency,” Simo said. “Today you are getting a lot more intelligence for the same computer and for the same dollars as you were getting a year ago.”
For all its emphasis on reasoning, one thing that is missing from today’s launch: a new image generator. Coincidentally, Battelle had quoted Altman as saying in his code red memo that image generation would be a significant focus for the company going forward, after an AI model developed by Google (dubbed Gemini 2.5 Flash Image and colloquially known as Nano Banana) went viral following its release in August.
Last month, Google unveiled an upgraded version of the model for text rendering, world knowledge, and a more uncanny-valley-real feel to its unedited photos, called Nano Banana Pro (aka Gemini 3 Pro Image). It also embeds more tightly into Google’s products, as you can see when it starts popping up in tools and workflows over the last week, such as Google Labs Mixboard for auto-generating presentations.
OpenAI will also release a new model in January, so the company says, that offers better images, improved speed, and personality — though it didn’t confirm those plans on Thursday.
OpenAI also announced Thursday that it’s rolling out additional safeguards, including age- and mental health-based use restrictions for teenagers. However, it didn’t spend much of the launch event pitching those changes.
