BT Explainer: OpenAI’s GPT-5.5 brings autonomy into focus, takes on Anthropic's Mythos

OpenAI has officially launched GPT-5.5, its newest flagship large language model, marking a significant evolution in the interaction between artificial intelligence and software systems across various enterprise applications. This release follows extensive speculation about the next-generation model and is timed with escalating competition from rivals such as Anthropic and Google. Positioned as more than just an incremental upgrade from GPT-5.4, the new model features a re-architecture aimed at enhancing 'agentic' capabilities. This allows GPT-5.5 to independently analyze challenges and perform complex tasks with minimal user prompts. Amelia 'Mia' Glaese, OpenAI's Vice President of Research, described it as the most proficient model yet, particularly emphasizing its advancements in coding tasks. OpenAI co-founder Greg Brockman highlighted the model's reduced need for human intervention, stating that GPT-5.5's intuitive design enables it to tackle ambiguous problems and determine subsequent steps without detailed instructions. This shift towards greater autonomy distinguishes it from previous models, which often required structured prompts to operate effectively. GPT-5.5 excels in managing intricate workflows, from debugging extensive codebases to navigating through various documents and spreadsheets. OpenAI asserts that this makes the model particularly valuable for coding, enterprise software applications, and scientific research—domains that Brockman identified as experiencing significant bottlenecks. In addition to the standard model, OpenAI has introduced GPT-5.5 Pro, designed for scenarios that demand heightened accuracy, including legal analysis and high-stakes business decisions. While the base model caters to a wide range of applications, the Pro version focuses on providing deeper analytical capabilities and optimized performance for multi-step processes. CEO Sam Altman expressed on social media that OpenAI aims to ensure equitable access to advanced technology, reinforcing their commitment to iterative deployments. Although GPT-5.5 is already a sophisticated tool, Altman anticipates rapid enhancements as part of their safety strategy. Benchmark tests indicate that OpenAI is regaining competitive ground in the race for advanced AI models. GPT-5.5 has outperformed accessible competitors like Claude Opus 4.7 and Gemini 3.1 Pro in most evaluated categories, especially in agentic tasks, cybersecurity challenges, and intricate mathematical problems. Notably, on the Terminal-Bench 2.0 test—designed to gauge a model's task completion abilities in a terminal environment—GPT-5.5 achieved 82.7% accuracy, surpassing Opus 4.7 and slightly edging out Claude Mythos Preview. Despite its advancements, GPT-5.5 Pro scored lower than some rivals in pure reasoning tests. While it achieved 43.1% on Humanity’s Last Exam, it trailed behind Opus 4.7 and Mythos Preview. This suggests that, while OpenAI has advanced in practical, task-oriented intelligence, competitors continue to excel in abstract reasoning. Among the noteworthy updates in GPT-5.5 is its enhanced efficiency, delivering improved intelligence without increased latency compared to its predecessor. OpenAI claims this was achieved through a combination of advanced hardware and custom algorithms, resulting in over 20% faster token generation. Currently, GPT-5.5 is not available via API, with OpenAI highlighting that API deployments necessitate distinct safety measures. Access is currently restricted to paid tiers of ChatGPT, including Plus, Pro, Business, and Enterprise, with the Pro tier reserved for higher-paying subscribers. As for pricing, costs for developers are expected to significantly rise once API access becomes available, with input and output costs for GPT-5.5 approximately double those of GPT-5.4. On the safety front, GPT-5.5 has been classified as 'High' risk in both biological and cybersecurity fields. OpenAI has instituted stricter controls for general users while permitting verified security professionals to access advanced capabilities with fewer limitations, acknowledging the dual nature of powerful models—increasing vulnerability detection while also posing risks for misuse.