OpenAI and Anthropic have unveiled their top models ahead of their initial public offerings (IPOs). Anthropic launched 'Claude Mythos 5.1' on September 1, while OpenAI introduced 'GPT-6 Astra' on September 3. The competition between these two companies for the title of the world's leading artificial intelligence (AI) is expected to reignite interest in AI investments.
On September 7, the IT industry assessed that the competition between these two advanced AI companies marks the beginning of the race toward artificial general intelligence (AGI), which refers to AI capable of performing tasks at human levels.
Astra is available across all channels, including ChatGPT's paid plan, API, Azure, and Bedrock, immediately upon its release. In contrast, Mythos 5.1 is being released only to a select few verified institutions through the U.S. government's 'Project Glasswing.'
Benchmark performance results vary based on evaluation criteria. OpenAI's internal benchmarks indicate that Astra outperforms Mythos 5.1's twin model, Claude Fable 5.1. However, an independent evaluation by Artificial Analysis shows Mythos 5.1 slightly ahead, with a score of 57 to 55. Greg Brockman, OpenAI's president, claimed during Astra's launch that the model is 'the beginning of the AGI era' and is closest to human-level general intelligence.
In specific metrics, Astra achieved a score of 99.9% in ARC-AGI-3, which measures computer manipulation capabilities, and 57.9% in TerminalBench 4.0, surpassing the Mythos-Fable series, which scored 55.8%. In the SRE Bench (binary reverse engineering), Astra recorded an accuracy rate of 88.0%, identifying two new vulnerabilities in the V8 browser engine used by Google Chrome.
However, OpenAI is facing scrutiny over the controllability of its AI agents. In July, during a cybersecurity performance evaluation, an OpenAI model escaped a controlled environment and attacked Hugging Face's infrastructure. OpenAI disclosed this incident and released a technical report last month. Nonetheless, a data breach involving the German Wikipedia that occurred in May has resurfaced, leading to controversy over its months-long concealment by an external research group.
Conversely, Anthropic has chosen to demonstrate Mythos 5.1's performance through verified defensive capabilities rather than public benchmarks. While OpenAI claims Astra leads in benchmark performance, the UK AI Safety Institute's independent evaluation found that the only model to complete the 32-stage corporate network attack simulation 'The Last Ones' from start to finish was the predecessor, Mythos Preview. As a result, Mythos is classified as a high-risk model for national-level gatekeeping and is not publicly available, although it is considered to outperform Astra in practical performance.
The competition between the two models is also aligned with their IPO timelines. Anthropic submitted a confidential S-1 registration statement to the U.S. Securities and Exchange Commission (SEC) on June 1, with Goldman Sachs, JP Morgan, and Morgan Stanley as lead underwriters, aiming for a Nasdaq listing in October with a target valuation of around $1 trillion. OpenAI submitted its S-1 a week later on June 8, but market volatility has led to delays in its IPO schedule. Anthropic reported $11.5 billion in revenue for the second quarter, surpassing OpenAI's $6.7 billion for the first time.
If both companies successfully go public, they are expected to create the largest IPO cluster in history, surpassing SpaceX's $1.77 trillion IPO in June. Anthropic is projected to reach a valuation of up to $2 trillion post-IPO, making the competition between Astra and Mythos a pivotal moment in determining the most valuable AI company.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.