Released on April 23, 2026, GPT-5.5 is a frontier AI model that represents the first fully retrained base model since GPT-4.5. It is primarily designed as an agentic system capable of executing complex, multi-step tasks independently, which includes planning, utilizing external tools, browsing, and verifying its own work without the need for continuous human supervision. The model's training and capabilities are heavily concentrated in four main domains: agentic coding, computer use, knowledge work, and early scientific research.A defining feature of GPT-5.5 is its massive 1,000,000-token context window, allowing it to process large datasets, extensive codebase repositories, and lengthy documents within a single prompt. The model is offered in two primary variations: the standard GPT-5.5, tailored for default production workloads, and GPT-5.5 Pro, a higher-compute tier designed for complex research, advanced mathematical reasoning, and high-stakes tasks.The model demonstrates state-of-the-art performance across multiple industry benchmarks. It scored 82.7% on Terminal-Bench 2.0, which evaluates complex command-line workflows and tool coordination. In knowledge work, it achieved 84.9% on GDPval, a benchmark measuring output quality across 44 professional occupations. For autonomous computer use, the model scored 78.7% on OSWorld-Verified, showcasing its ability to natively read screens and navigate desktop interfaces. In coding, it reached 58.6% on SWE-Bench Pro and 73.1% on the internal Expert-SWE test, proving highly capable of resolving real-world software issues.Despite its increased intelligence and capabilities, GPT-5.5 successfully matches the per-token latency of its predecessor, GPT-5.4. This performance is largely attributed to co-design efforts with NVIDIA GB200 and GB300 systems, as well as infrastructure optimizations that were authored by the AI itself. While the base API pricing has increased—costing $5 per 1 million input tokens and $30 per 1 million output tokens for the standard tier—the model is highly token-efficient. It requires significantly fewer output tokens to complete tasks, and a prompt caching feature can reduce repeated input costs by 90%.There are a few notable limitations. The model exhibits a high hallucination rate on the questions it gets wrong, meaning it tends to be confidently incorrect rather than flagging its own uncertainty. To mitigate this, it is recommended to pair the model with verification tools and higher reasoning effort settings. Additionally, it can sometimes be overly verbose when answering straightforward queries.In terms of security, GPT-5.5 is rated "High" for cybersecurity and biological capabilities under the Preparedness Framework. It features stricter safety classifiers by default, though verified defenders can access cyber-permissive features through a Trusted Access program. The model is currently available to premium subscribers in chat and coding interfaces, but broader API access was initially delayed to implement necessary safety safeguards for large-scale serving. Become a supporter of this podcast: https://www.spreaker.com/podcast/tech-talk-daily--6886557/support.
