OpenAI releases 'GPT-5.2' just under a month after the emergency 'Code Red' was issued, will it be a counterattack against Google's fierce pursuit?



OpenAI has released GPT-5.2, the latest family of AI models for ChatGPT. This release was based on a ' Code Red ' declaration issued internally by CEO Sam Altman in early December 2025, and was reportedly launched less than a month after the Code Red was issued in response to competitive pressure from Google's Gemini 3.

GPT-5.2 is here | OpenAI
https://openai.com/ja-JP/index/introducing-gpt-5-2/

Update to GPT-5 System Card: GPT-5.2 | OpenAI
https://openai.com/ja-JP/index/gpt-5-system-card-update-gpt-5-2/

Advancing science and math with GPT-5.2 | OpenAI
https://openai.com/ja-JP/index/gpt-5-2-for-science-and-math/

GPT-5.2 is available in three models: GPT-5.2 Instant, GPT-5.2 Thinking, and GPT-5.2 Pro. Instant is a speed-optimized model for everyday queries like information search, writing, and translation, Thinking excels at complex structured tasks like coding, analyzing long documents, mathematics, and planning, and Pro is the top-of-the-line model designed to provide the highest accuracy and reliability for challenging problems.

GPT-5.2 Instant is a fast, reliable, lightweight model suitable for everyday work and learning. It demonstrates clear improvements in tasks such as information search, procedural explanations, technical documentation, and translation. It maintains the warm, conversational tone of its predecessor, providing clear explanations that present important information first. Safety has also been enhanced, significantly reducing inappropriate responses related to suicide, self-harm, and emotional dependency. The API costs $1.75 (approximately ¥270) for input and $14 (approximately ¥2,180) for output per million tokens.

GPT-5.2 Thinking is the flagship model in the series, ideal for specialized tasks, and excels in complex reasoning, coding, and long-text processing. It boasts significant improvements in tool utilization and multi-stage project processing, delivering superior performance in spreadsheets and code creation. In terms of long-text processing, it achieved nearly 100% accuracy in information retrieval up to 256k tokens in the OpenAI MRCRv2 benchmark. It also boasts improved visual comprehension, with error rates for chart reasoning and UI comprehension reduced by half. Key benchmark results include a 70.9% score on GDPval , which evaluates knowledge-based tasks; an 80.0% score on SWE-bench Verified , which measures coding ability; and a 40.3% score on FrontierMath . It's priced in the same range as the Instant model.



OpenAI claims that GPT-5.2 Thinking outperforms Claude Opus 4.5 and Gemini 3 Pro in various benchmarks.



GPT-5.2 Pro is the most advanced, reliable, and top-of-the-line model, specialized for challenging scientific or specialized questions. It boasts the highest quality inference capabilities, reducing critical errors and improving performance in complex domains like programming. It has proven successful in supporting scientists' research, including solving open problems in statistical learning theory. Developers can configure inference parameters via API, including the xhigh setting, which prioritizes quality. Benchmarks have achieved high scores: 93.2% on GPQA Diamond , a graduate-level science Q&A test; 100.0% on HMMT , a mathematics competition; and 54.2% on ARC-AGI-2 , an abstract reasoning test. The price is $21 (approximately ¥3,270) for input and $168 (approximately ¥26,000) for output per million tokens.



GPT-5.2 will be gradually rolled out to all ChatGPT models on paid plans, including Plus, Pro, Go, Business, and Enterprise, from December 12, 2025. The developer API will also be available on the same day under the names 'gpt-5.2' for the Thinking model, 'gpt-5.2-chat-latest' for the Instant model, and 'gpt-5.2-pro' for the Pro model. The previous GPT-5.1 model will remain available to paid users as a legacy model for three months, after which it will be discontinued. While the current version of GPT-5.2 is available for coding purposes on Codex, a version optimized for Codex is expected to be released within the next few weeks.

Max Schwarzer, product lead at OpenAI, said that GPT-5.2 has made significant improvements in code generation and debugging, and that coding startups such as Windsurf have reported it to be the best performing agent. He also highlighted that the Thinking model significantly reduces hallucinations compared to its predecessor, reducing errors by 38%. He also said that the vision for the future is to make ChatGPT the best personalized assistant possible.

The release of GPT-5.2 comes as Google's latest AI model, Gemini 3, has taken the top spot in multiple AI benchmarks and begun to expand its share in the consumer market. Amid growing concerns about ChatGPT's declining traffic and the loss of market share to Google, OpenAI's CEO, Sam Altman, declared a 'Code Red' in an internal memo in early December, taking emergency measures to counter Google's offensive.

OpenAI CEO Sam Altman declares 'Code Red,' prioritizes ChatGPT improvements and postpones advertising in response to Gemini's success - GIGAZINE



Under this 'Code Red,' Altman focused internal resources on improving ChatGPT's core experience, temporarily postponing other initiatives such as plans to introduce advertising, and announced a company-wide policy to focus on strengthening the product's competitiveness. OpenAI Chief Product Officer Fiji Simo also explained that Code Red focused resources on specific areas.

As a result, OpenAI launched the new model, GPT-5.2, exceptionally quickly, just under a month after releasing the previous version, GPT-5.1. Although some employees called for a delay in the release, saying that more time should be spent on improving the model, it is reported that the priority was to regain control from Google, and the release of GPT-5.2 was forced.

Altman commented, 'Even without new features like sophisticated file output, I feel that GPT-5.2 is the biggest upgrade we've been waiting for in a long time.'



Related Posts:

in AI,   Software, Posted by log1i_yk