GPT-5.2

Subscribe to the channel for more like this!In this video, Igor breaks down OpenAI's new GPT-5.2 model, which is now powering ChatGPT for both free...

OpenAI is testing a new model for Codex called

OpenAI announced GPT-5.2 on December 11, 2025, positioning it as the company's most advanced model and best offering yet for everyday professional use, according to CNBC. The model improved over predecessors at creating spreadsheets, building presentations, perceiving images, writing code, and handling long context. It became available the same day inside ChatGPT and OpenAI's API. The launch followed OpenAI's GPT-5.1 release and came amid a "code red" effort to refocus resources on ChatGPT after rivals Anthropic and Google shipped new models. CEO of applications Fidji Simo said the cod red signaled resource prioritization rather than a direct response to Google. Sam Altman told CNBC he expected OpenAI to exit the code red by January and that Gemini 3's impact was smaller than feared.

The llm-stats.com benchmark dashboard lists GPT-5.2 ranked 64th on its composite LLM Stats Score, with a blended price around $2.33 per million tokens against a score of 41.1. The page, captured October 17, 2026, reports GPT-5.2 scoring a 1.00/1 on AIME 2025 (rank 2) under its Thinking variant without tools, sourced from openai.com, placing it near the top on that mathematical-reasoning benchmark. Conversation-depth tracking shows GPT-5.2 holding roughly 14.8 at turn 1 and declining about 2 points to 12.8 by turns 31+, with a Quality Tracker reading of +3.54 sigma stable over 1,585 votes. Subcategory sigmas highlight gains in Chat (+10.86), Games (+5.32), Data Viz (+5.28), and Animations (+6.86), while Websites ran slightly below baseline.

A January 15, 2026 arXiv safety report (2601.10527) by researchers from Fudan University and collaborators evaluated GPT-5.2 alongside Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5 using a unified protocol spanning language, vision-language, and image generation. The abstract states GPT-5.2 demonstrates consistently strong and balanced safety performance across evaluations while other models showed pronounced trade-offs. The study aggregated results into safety leaderboards and model safety profiles across multiple modes, including benchmark evaluation, adversarial evaluation, multilingual evaluation, and compliance evaluation. It found both language and vision-language modalities degrade substantially under adversarial evaluation despite strong benchmark results, underscoring that frontier-model safety is inherently multidimensional and shaped by modality, language, and evaluation scheme.

BenchLM's model record for GPT-5.2, marked "Established" and released December 11, 2025, reports a 65/100 capability score (rank 37 of 232) with data dated September 15, 2026. The profile lists a 400K token context window, roughly 80 tokens/second throughput with a 91.08s first-token latency, and API pricing of $1.75 p Across capability axes, the page places GPT-5.2 at the 32nd percentile for Agentic, 46th for Coding, 85th for Knowledge, and 53rd for Multimodal work, with several categories marked Not eligible or Not ranked. Its strongest eligible category is Multimodal & Grounded at rank 23, while Agentic sits lowest at rank 104 of