Get ready to redefine how you work with your computer. We’re thrilled to introduce GPT-5.5, our most intelligent and intuitive model to date, marking a significant leap toward a new paradigm for digital productivity. This groundbreaking AI is designed to understand your intent faster and shoulder more of your workload, transforming complex tasks into streamlined processes.
GPT-5.5 excels across a wide range of demanding activities, from writing and debugging code to conducting in-depth online research and analyzing intricate data. It autonomously handles tasks like creating professional documents and spreadsheets, operating software, and seamlessly transitioning between various tools until a job is finished. You can now delegate messy, multi-part assignments and trust it to independently plan, execute, check its own work, and navigate ambiguities.
This advanced capability shines particularly bright in agentic coding, sophisticated computer use, high-level knowledge work, and pioneering scientific research. GPT-5.5 delivers a substantial intelligence boost without sacrificing speed, matching its predecessor GPT-5.4 in per-token latency during real-world serving. Furthermore, it achieves higher performance while consuming significantly fewer tokens for the same tasks, making it both more powerful and remarkably efficient.
Our commitment to safety is paramount, and GPT-5.5 launches with our most robust set of safeguards yet. We rigorously evaluated the model across comprehensive safety frameworks, engaged internal and external red teamers, and conducted targeted testing for advanced capabilities. Crucially, we gathered invaluable feedback from nearly 200 trusted early-access partners before its public release.
Today, GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users within ChatGPT and Codex, with GPT-5.5 Pro also becoming available for Pro, Business, and Enterprise users in ChatGPT. API deployments, which require unique safety and security considerations for large-scale use, are being meticulously prepared. We anticipate bringing both GPT-5.5 and GPT-5.5 Pro to the API very soon.
Unleashing Unprecedented Coding Power
OpenAI is building the global infrastructure for agentic AI, empowering individuals and businesses worldwide to revolutionize their workflows. AI has already dramatically accelerated software engineering, and with GPT-5.5 now integrated into Codex and ChatGPT, this transformative power is rapidly expanding into scientific research and broader computer-based work. GPT-5.5 isn’t just smarter; it’s also remarkably more efficient, consistently achieving higher-quality outputs with fewer tokens and fewer retries.
Independent evaluations underscore GPT-5.5’s groundbreaking capabilities. On Artificial Analysis’s Coding Index, the model delivers state-of-the-art intelligence at half the cost of competitive frontier coding models. This efficiency paired with superior performance sets a new benchmark for AI-driven development.
GPT-5.5 stands as our strongest agentic coding model ever, proven by exceptional benchmark results. On Terminal-Bench 2.0, testing complex command-line workflows, it achieves a stellar accuracy of 82.7%. For real-world GitHub issue resolution on SWE-Bench Pro, GPT-5.5 reaches an impressive 58.6%, solving more tasks end-to-end in a single pass than previous models. Even on Expert-SWE, our internal evaluation for long-horizon coding, GPT-5.5 significantly outperforms GPT-5.4 while using fewer tokens across all these evaluations.
The model’s coding prowess shines particularly brightly in Codex, handling a wide spectrum of engineering tasks from implementation and refactoring to debugging and testing. Early tests indicate GPT-5.5 excels at crucial behaviors like maintaining context across large systems, reasoning through ambiguous failures, and seamlessly propagating changes throughout an entire codebase. This means more effective and less error-prone development cycles for engineers.
Industry leaders are already witnessing this paradigm shift. Dan Shipper, Founder and CEO of Every, praised GPT-5.5 as “the first coding model I’ve used that has serious conceptual clarity.” He recounted how the model successfully re-architected a critical system to fix a post-launch bug, a task that GPT-5.4 could not manage. Similarly, Pietro Schirano, CEO of MagicPath, saw GPT-5.5 merge a complex branch with hundreds of changes into a substantially altered main branch in just 20 minutes, noting, “It genuinely feels like I’m working with a higher intelligence.”
Knowledge Work Reinvented
The same inherent strengths that make GPT-5.5 an extraordinary coding assistant also empower it for virtually any task on a computer. With its superior ability to comprehend user intent, the model navigates the entire knowledge work loop with greater fluidity. This includes efficiently finding information, discerning critical details, skillfully employing various tools, and transforming raw data into polished, actionable results.
Within Codex, GPT-5.5 significantly surpasses previous models in generating complex documents, detailed spreadsheets, and compelling slide presentations. Alpha testers reported its exceptional performance in areas like operational research, sophisticated spreadsheet modeling, and converting disorganized business inputs into well-structured plans. Combined with Codex’s integrated computer use capabilities, GPT-5.5 truly feels like it’s collaborating with you, interpreting on-screen information and navigating interfaces with remarkable precision.
Teams at OpenAI are already leveraging these advancements in their daily workflows; over 85% of the company utilizes Codex every week across diverse functions. For instance, the Communications team used GPT-5.5 in Codex to analyze six months of speaking request data and validate an automated Slack agent to handle low-risk requests automatically. In Finance, Codex powered the review of an astounding 24,771 K-1 tax forms, totaling 71,637 pages, accelerating this daunting task by two weeks. Similarly, an employee on the Go-to-Market team successfully automated weekly business reports, saving 5-10 hours every week.
For ChatGPT users, GPT-5.5’s advanced “Thinking” capabilities unlock faster and more insightful assistance for challenging problems. It provides smarter, more concise answers that streamline complex professional tasks, excelling in coding, research, information synthesis, and document-heavy assignments, especially when utilizing plugins. The elite GPT-5.5 Pro further enhances this, offering substantial latency improvements and consistently more comprehensive, accurate, and useful responses, particularly strong in business, legal, education, and data science.
GPT-5.5 achieves state-of-the-art performance across numerous benchmarks reflecting this caliber of work. It scores an impressive 84.9% on GDPval for knowledge work across 44 occupations, and reaches 78.7% on OSWorld-Verified for operating real computer environments autonomously. On Tau2-bench Telecom, testing complex customer-service workflows, it achieves a remarkable 98.0% without any prompt tuning.
Pioneering Scientific Breakthroughs
The impact of GPT-5.5 extends powerfully into scientific and technical research workflows, which demand more than just answering difficult questions. Researchers need to explore hypotheses, gather evidence, test assumptions, interpret complex results, and strategically decide on subsequent steps. GPT-5.5 exhibits a superior ability to persist throughout this iterative research loop, accelerating discovery like never before.
Notably, GPT-5.5 shows a clear improvement over GPT-5.4 on GeneBench, a new evaluation focusing on multi-stage scientific data analysis in genetics and quantitative biology. Its striking performance is particularly significant given that such tasks often correspond to multi-day projects for human scientific experts. Similarly, on BixBench, a benchmark designed around real-world bioinformatics and data analysis, GPT-5.5 achieved leading performance among all models with published scores, truly establishing it as a bona fide co-scientist.
In a groundbreaking example, an internal version of GPT-5.5 helped discover a new proof about Ramsey numbers—central objects in combinatorics. This field explores how discrete objects fit together, dealing with graphs, networks, sets, and patterns. GPT-5.5 found a proof of a longstanding asymptotic fact concerning off-diagonal Ramsey numbers, a discovery later verified in Lean, a formal proof assistant. This demonstrates GPT-5.5 not merely generating code or explanations, but actively contributing a surprising and highly useful mathematical argument within a core research area, signifying a new era of AI-powered intellectual partnership and scientific advancement.
Source: OpenAI Newsroom
