OpenAI has announced a new version of its language model called GPT-5.5. According to the company, this iteration is primarily focused on improving so-called agentic capabilities, whereby a model not only generates answers but can also independently plan and execute steps to complete a task.
In its explanation, OpenAI states that GPT-5.5 performs better with complex workflows that require multiple steps. The model is reportedly better able to gather information, check intermediate results, and correct errors without constant human guidance. This shifts the focus from a question-and-answer interface to a system that can partially take over tasks. The Verge notes that this development fits within a broader trend in the AI market, where vendors are focusing on systems that not only respond but actually perform work.
A key focus of the announcement is software development. OpenAI states that GPT-5.5 handles larger codebases better and delivers more consistent results when debugging and restructuring code.
According to the company, progress has been made in maintaining context across longer interactions, which is relevant for projects consisting of multiple files and dependencies. The Verge notes that these improvements are specifically aimed at practical use within development environments, where AI increasingly acts as an active assistant rather than a standalone tool.
GPT-5.5 performs better in programming and workflow tests
OpenAI also highlights specific performance improvements. On internal and external benchmarks focused on software development and agent-based tasks, GPT-5.5 outperforms its predecessor GPT-5.4, particularly in realistic programming and workflow scenarios. The company emphasizes that the gains are primarily in reliability and the completion of full tasks, rather than in isolated, standalone responses.
In terms of efficiency, OpenAI reports that GPT-5.5 has response times comparable to its predecessor, while requiring fewer tokens to produce a result. According to the company, this indicates optimizations in both the model architecture and the underlying infrastructure. The model also supports larger context windows, allowing it to handle large datasets and codebases more effectively.
The organization further states that additional safety measures have been implemented, with specific attention to risks in cybersecurity and the life sciences. How these measures have been implemented in practice is described only in general terms. The Verge notes that OpenAI provides little insight into the model’s technical architecture, making it difficult to independently verify performance claims.
GPT-5.5 is being made available within ChatGPT for paying users and business customers. According to OpenAI, API support will follow shortly. With this, the company is taking the next step in developing models that not only respond to input but also play a more active role in performing digital tasks.
Read also: DeepSeek V4 drastically reduces the cost of agentic AI