The competitive battle between AI model builders has shifted. Where OpenAI was the leader for years, it now has to yield to Google. GPT-5.2 is a new model that is notably focused on knowledge work, but also includes improvements for coding, spreadsheets, presentations, and visual recognition. The new version is available to paying ChatGPT users starting today.
OpenAI’s announcement probably comes sooner than the company had previously suggested. The company recently declared a “code red” due to Google Gemini’s stunning performance compared to GPT-5 and 5.1. GPT-5.2 aims to turn the tide with a series of improvements focused primarily on business applications.
The new model comes in three variants, as was the case with GPT-5.1. GPT-5.2 Instant focuses on quick daily tasks. The Thinking variant is intended for more complex tasks such as extensive coding sessions and spreadsheet models. According to OpenAI, the Pro version offers the highest quality for difficult problems that require extra computing power. Despite the hefty price tag of $200 per month, this option would be unprofitable for the company behind ChatGPT, mainly because such high-end users make full use of their subscription.
Performance surpasses predecessor
On GDPval, an evaluation of knowledge work across 44 professions, GPT-5.2 Thinking scores 70.9 percent. This means that in more than two-thirds of cases, the model performs equally well or better than industry experts. Unfortunately, we cannot compare this with the latest models from Google and Anthropic, as GDPval is a proprietary benchmark from OpenAI. For software engineering, the model scores 55.6 percent on SWE-Bench Pro, a test that covers four programming languages. If we look at SWE-Bench’s site ourselves, we see that the regular “Bash Only” test places Claude 4.5 Opus and Gemini 3 Pro ahead of GPT-5.2.
The improvements in GPT-5.2 are also evident in the details. The model hallucinates 30 percent less than its predecessor, GPT-5.1. For spreadsheets and presentations, OpenAI claims better formatting and more advanced functionality. Users will need to be patient, though: complex generations can take several minutes. This has been a known fact since the emergence of OpenAI’s o-models (o1, o3) and was further emphasized by DeepSeek-R1 earlier this year: “inference time compute,” or long thinking by AI models, almost always yields better end results.
Last week, Google launched Gemini 3 with Deep Think mode, which promises similar reasoning capabilities. AI models are once again in a fierce competition, with Anthropic’s Claude and xAI’s Grok 4 also being serious options that are regularly updated with new versions.
Business integrations take center stage
As mentioned, OpenAI emphasizes the increased value for business users. This is also an area that the company has been leaning on more recently, although it does not yet seem entirely sure of its revenue model for consumers. According to the company, a ChatGPT Enterprise user saves an average of 40 to 60 minutes per day. Heavy users claim savings of more than ten hours per week.
The platform integrates with tools from partners such as Notion, Box, and Shopify. For data scientists, there are specific improvements in tool calling and document analysis through collaborations with Databricks and Hex. Coding platforms such as JetBrains and Augment Code report measurable progress in agentic coding performance.
The new version also supports scientific work. On GPQA Diamond, a graduate-level benchmark, GPT-5.2 Pro scores 93.2 percent. For mathematics, the model achieves 40.3 percent on FrontierMath Tier 1-3, an evaluation for expert-level problems.
Pricing and availability
In the API, GPT-5.2 costs $1.75 per million input tokens and $14 per million output tokens. That is more expensive than GPT-5.1, but according to OpenAI, it is more efficient due to better token efficiency. The Pro variant costs $21 per million input tokens.
OpenAI built the model together with partners Nvidia and Microsoft. Azure data centers and Nvidia’s H100, H200, and GB200-NVL72 GPUs formed the basis for the training infrastructure. According to both parties, the collaboration should lead to faster development cycles for future models.
The price will remain the same for ChatGPT subscribers. GPT-5.1 will remain available for three months under legacy models, after which it will be phased out. The rollout will be phased in order to keep the service stable.