OpenAI Launches New GPT-5.5 Model as Anthropic Faces Security Controversy
by Kong Byeongseon
Published 24 Apr.2026 10:10(KST)
OpenAI Unveils New AI Model 'GPT-5.5'
Counters, "Anthropic Model Shows Signs of Memorization"
Unauthorized Access Detected for Anthropic's New Model
OpenAI is accelerating its competition with Anthropic by consecutively releasing new artificial intelligence (AI) models. This move is seen as an effort to solidify its position as the industry leader by directly comparing the performance of its new AI models to those of Anthropic. Meanwhile, Anthropic, which had been closing in on OpenAI, is now focusing on damage control after becoming embroiled in a security controversy just ahead of the official launch of its high-performance AI model.
On April 23 (local time), OpenAI introduced its new generative AI model, 'GPT-5.5', during a video briefing, describing it as "the smartest and most intuitive model to use so far." Greg Brockman, President of OpenAI, stated, "This model can accomplish much more with fewer instructions," and added, "When it encounters an ambiguous problem, it figures out for itself what to do next." He also emphasized, "To me, GPT-5.5 feels like it lays the foundation for how we will use computers in the future and how large-scale agent-based computation will work."
OpenAI released its previous model, 'GPT-5.4', on March 5 and unveiled the new model in less than two months. OpenAI explained that the new model offers improvements not only in coding but also in document writing, online information retrieval, data analysis, and scientific and technological research. On the 'Terminal Bench 2.0' metric, which evaluates the ability to perform complex command tasks, GPT-5.5 scored 82.7%, surpassing GPT-5.4’s 75.1%. The company also said that on its internal 'Expert-SWE' metric, which assesses performance on lengthy coding assignments, GPT-5.5 produced better results than GPT-5.4.
When unveiling the new model, OpenAI focused on comparisons with Anthropic. According to the performance report released by OpenAI, GPT-5.5 scored 84.9% on the 'GDPval' metric, which measures knowledge work capabilities. This outperformed Anthropic’s generative AI model 'Opus 4.7', which scored 80.3%. However, in the coding metric 'SWE-Bench Pro', GPT-5.5 recorded 58.6%, falling behind Opus 4.7’s 64.3%. OpenAI argued that "Anthropic’s model shows signs of data memorization," meaning that Opus 4.7 produces answers not by understanding and programming problems, but by reciting information memorized during training.
OpenAI Enveloped in 'Crisis Theory' Amid Anthropic’s Pursuit... Successive Releases of New AI Models
OpenAI has also found itself caught up in 'crisis theory' due to Anthropic’s aggressive pursuit. According to data analysis platform Mobile Index, the number of monthly active users (MAU) of ChatGPT in Korea last month reached 15.34 million, up about 7% from 14.29 million in January. In contrast, Anthropic’s MAU in Korea last month was about 600,000, representing a nearly fourfold increase over the same period and demonstrating rapid growth. Sam Altman, CEO of OpenAI, implemented an emergency management system in December last year, conscious of the competition.
Accordingly, OpenAI is now focusing on 'separating the wheat from the chaff.' Last month, OpenAI withdrew its video generation AI model 'Sora' to concentrate on more profitable ventures. On April 21, the company unveiled a new image generation AI model, 'ChatGPT Image 2.0.' ChatGPT Image 2.0 is recognized for resolving longstanding issues in generative AI models, such as blurry text and images irrelevant to instructions. Based on its successive AI model launches, OpenAI aims not just to provide simple tools but to build an AI agent ecosystem.
Signs of External Access to Mythos Detected... Concerns Over Large-Scale Hacking
Meanwhile, Anthropic, which had been persistently catching up, now faces a crisis amid a series of security controversies. According to U.S. media outlets including Bloomberg on April 22, Anthropic discovered indications that unauthorized external users had accessed its internal systems to reach its new high-performance AI model 'Mythos,' and has launched an investigation. While Mythos is currently being tested and provided to only about 40 institutions, this incident involved unauthorized external access. Due to Mythos’s outstanding capabilities, some argue that this is not just a simple security breach but could pose large-scale hacking and other security risks.
Mythos was also revealed after an internal report from Anthropic was leaked. On March 27, U.S. business outlet Fortune reported that internal documents from Anthropic’s online repository had been leaked, exposing details about Mythos’s performance. According to these documents, Mythos outperformed its predecessor, 'Opus 4.6,' in all areas, including software coding and academic reasoning. Most notably, its exceptional performance in the field of cybersecurity has made governments and companies around the world uneasy. Mythos can not only detect security vulnerabilities autonomously, but also generate tools to execute attacks. On April 12, the Cloud Security Alliance in the U.S. pointed out that "before a hacker attacks, vulnerabilities in programs or code must be immediately checked through AI agents" in reference to Mythos.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.