OpenAI launches GPT-5.5, Agentic Intelligence and Computer Automation

Performance in coding and computer use

The analyzes conducted on standardized benchmarks highlight significant results in the field of software development. On Terminal-Bench 2.0, which evaluates command-line workflows, the model achieved an accuracy of 82.7%. In the SWE-Bench Pro test, which focuses on solving real problems on GitHub, the system solves 58.6% of tasks in a single pass.

This operational effectiveness extends to “computer use”: the model can observe the screen, type and interact with software with an accuracy that allows it to be used in areas such as financial analysis and operations research. According to data from Artificial Analysis, the system offers advanced performance at approximately half the cost compared to competing models in the same category, thanks to more efficient token management.

Business applications and scientific research

Internal adoption of the technology has already shown measurable impacts. In the finance industry, automating the review of thousands of tax forms accelerated turnaround times by two weeks. In marketing and communications, automated agents now manage complex data request flows, autonomously filtering out low-risk cases and delegating only critical exceptions to human supervisors.

In the scientific field, progress is evident in GeneBench, a test for the analysis of genetic data and quantitative biology. The model demonstrates greater persistence across research cycles that include gathering evidence, testing assumptions, and interpreting complex statistical results—tasks that would normally require several days of work for human experts.

Efficiency and safety protocols

Despite the increase in model complexity, the latency per token remains in line with previous versions (GPT-5.4), ensuring a response speed compatible with intensive professional use.

The release is accompanied by an updated security framework. The model has undergone targeted testing to prevent abuse in sensitive areas such as cybersecurity and biology, using feedback from around 200 external partners. Currently, the technology is being rolled out to Plus, Pro and Business users, while API access will be extended in the near future once specific security protocols for large-scale integrations are finalized.

API Availability, Plans and Costs

GPT-5.5 is currently rolling out to Plus, Pro, Business and Enterprise users within ChatGPT and Codex. In particular, the function “GPT-5.5 Thinking” It is accessible at all paid membership levelsWhile the GPT-5.5 Pro version, optimized for very high precision tasks, is reserved for Pro, Business and Enterprise users.

As for integration for developers, gpt-5.5 APIs will soon be available with a 1 million token context window. The standard model will cost $5 per 1 million tokens for input and $30 for output. The most advanced version, gpt-5.5-pro, will be priced at $30 and $180 per million tokens, respectively. There are “Priority” processing options (at 2.5x the cost) and half the “Batch” rates for non-urgent processes.

In the Codex environment, the model supports a 400,000 token context window and is also available in “Fast” mode, which increases generation speed by 50%. Although nominal costs are higher than the GPT-5.4 version, the architecture has been optimized to be significantly more efficient: in fact, the system requires a lower number of tokens and attempts to correctly complete the same task, balancing the economic investment with greater operational effectiveness.