After escalating a ‘Code Red’ warning last week to fix ChatGPT’s deteriorating user experience, OpenAI has launched a new model in response to Google’s Gemini 3.
The AI giant has released GPT-5.2, a new frontier model for professional knowledge work, long-running agents, and complex workflows. OpenAI, in a blog post, said the model improves reasoning, tool use, long-context understanding, vision tasks, and coding.
The company is now rolling out the model to ChatGPT paid plans, available to developers through its API.
GPT-5.2 comes in three variants—Instant, Thinking, and Pro.
Instant targets everyday queries and lightweight tasks, while Thinking is designed for deep work such as document analysis, coding, multi-step reasoning and planning. Pro is positioned as the most reliable option for complex domains where accuracy outweighs latency.
GPT-5.1 will remain available to paid ChatGPT users for three months before being retired.
In the API, GPT-5.2 is priced at $1.75 per million input tokens and $14 per million output tokens. OpenAI said its greater token efficiency often results in lower total cost to achieve a desired quality level than GPT-5.1. The company added that there are no current plans to deprecate GPT-5.1, GPT-5, or GPT-4.1 in the API.
According to the company, GPT-5.2 Thinking outperformed industry professionals in 70.9% of tasks measured by GDPval, a benchmark covering well-specified work across 44 occupations. The company said it is the first time one of its models has matched or exceeded expert humans on this evaluation. It also reported a 30% relative reduction in response-level errors compared with GPT-5.1 Thinking, based on de-identified ChatGPT queries.
Across benchmarks, GPT-5.2 demonstrated gains in software engineering, mathematics, scientific reasoning, and abstract problem solving, OpenAI stated in its blog post. It set new highs on SWE-Bench Pro, GPQA Diamond, FrontierMath Tier 1–3, ARC-AGI-2, and internal spreadsheet-modelling tasks, it added.
OpenAI said the model also maintains coherence over hundreds of thousands of tokens, achieving near-perfect accuracy on long-context MRCR tests—measuring the model’s ability to understand ordering in natural text—at up to 256,000 tokens.
It reported significant advances in visual understanding. Error rates dropped on chart-reasoning and GUI-interpretation tasks, and the model showed stronger spatial mapping of image components, the company stated.
OpenAI said this improves practical use cases such as analysing dashboards, diagrams, and product interfaces.
OpenAI said GPT-5.2 builds on the safe-completion framework introduced with GPT-5, with improved responses in areas such as self-harm, emotional reliance on AI, and mental-health indicators. The company is expected to roll out an age-gated model to apply content restrictions for users under 18 automatically in the first quarter of 2026.
ALSO READ: Broadcom Reveals $21 Billion Google TPUs Order from Anthropic
