GPT-5.5 - OpenAI's latest flagship large-scale model

What is GPT-5.5?

GPT-5.5 is OpenAI's latest flagship model, positioned as the most intelligent and intuitive AI system to date. The model achieves significant leaps in four major areas: agent programming, computer use, knowledge work, and scientific research, reaching 82.7% in Terminal-Bench 2.0 and 58.6% in SWE-Bench Pro. GPT-5.5 is co-designed with NVIDIA GB200/GB300 NVL72 systems from training to deployment, achieving comparable latency to GPT-5.4 while using fewer tokens to accomplish the same tasks. The model is now available to ChatGPT Plus/Pro and Codex users.

Main functions of GPT-5.5

• Intelligent agent programmingIt supports end-to-end code implementation, refactoring, debugging and testing, and can independently plan and execute complex development tasks.
• Computer useIt has visual perception and GUI operation capabilities, and can automatically click, input, and navigate across tools to complete workflows.
• Knowledge workGenerate documents, tables, and PPTs; conduct operational research and data modeling; and transform disorganized business inputs into actionable plans.
• Scientific research: To assist in multi-stage scientific data analysis, gene expression research, and mathematical proof exploration, acting as a "collaborative scientist".
• Tool callIt can independently use plugins, browse web pages, and analyze data, checking output and self-correcting in multi-step tasks.

Technical Principles of GPT-5.5

• Intelligent agent reasoning architectureIt adopts an end-to-end task planning and execution framework, which allows the model to autonomously decompose multiple tasks, call tools, verify outputs, and continuously iterate in fuzzy scenarios without the need for manual intervention.
• Hardware and software co-designIt is designed in conjunction with the NVIDIA GB200/GB300 NVL72 system for the entire stack from training to inference, achieving bidirectional optimization of model architecture and hardware infrastructure.
• Dynamic reasoning optimizationBy treating inference as an integrated system rather than an isolated optimization, and replacing fixed static partitioning with dynamic load balancing and partitioning heuristics, GPU utilization is improved by more than 20%.
• Context-aware mechanismSupports ultra-long context windows ranging from 400K (Codex) to 1M (API), and employs an efficient attention mechanism to maintain long-term memory and cross-file reasoning in large-scale codebases and documents.
• Multimodal sensing and manipulationIt integrates visual understanding capabilities, can parse screen content and perform precise GUI operations, enabling autonomous computer use across applications.

Performance of GPT-5.5

• Intelligent agent terminal operationTerminal-Bench 2.0 achieved a completion rate of 82.7%, significantly outperforming GPT-5.4's 75.1% and Claude Opus 4.7's 69.4%, resulting in a substantial improvement in the completion rate of complex command-line workflows.
• Real code fixesSWE-Bench Pro achieves a 58.6% success rate, surpassing its predecessor in end-to-end problem-solving capabilities on GitHub, allowing for more tasks to be completed in a single pass.
• Knowledge Work AssessmentGDPval reached 84.9%, and its performance on knowledge work tasks covering 44 occupations was better than Claude Opus 4.7's 80.3% and Gemini 3.1 Pro's 67.3%.
• Computer autonomous operationOSWorld-Verified rating reaches 78.7%, demonstrating industry-leading ability to autonomously perform complex operations in real-world computer environments.
• Customer service workflowTau2-bench Telecom achieves a 98.0% success rate, handling complex customer service processes without the need for prompt word optimization.
• Financial modelingFinanceAgent achieved a completion rate of 60.0%, and internal investment banking modeling tasks reached 88.5%, demonstrating outstanding professional financial analysis capabilities.
• Mathematical reasoningFrontierMath reached 51.7%, and high-difficulty math problems (Tier 3 & 4) reached 35.4%, far exceeding Claude Opus 4.7's 22.9%.
• CybersecurityCyberGym achieved a score of 81.8%, outperforming its competitors in advanced cybersecurity capability assessments.
• Reasoning efficiencyOn the Artificial Analysis Intelligence Index, it achieves higher scores with the same token consumption, or requires fewer tokens for the same score, thus achieving both intelligence and efficiency.
• Delay controlIn a real production environment, the per-token latency is on par with GPT-5.4, breaking the scaling law that "stronger always means slower".

How to use GPT-5.5

• ChatGPT usersVisit the ChatGPT official website,Plus/Pro/Business/Enterprise users can now switch to GPT-5.5 Thinking in the model selector; Pro users can also use GPT-5.5 Pro.
• Codex usersVisit the Codex website,Plus/Pro/Business/Enterprise/Edu/Go plan users can select GPT-5.5 in Codex, which supports 400K context and Fast mode.
• API developersThe model will soon be available on the Responses and Chat Completions APIs.
• Secure accessTrusted organizations engaged in cyber defense can apply for Trusted Access for Cyber, which reduces usage restrictions for defensive work through chatgpt.com/cyber.

Key information and usage requirements of GPT-5.5

• Release statusOfficially released on April 24, 2026, it has been pushed to subscribers, and the API will be available soon.
• Subscription RequirementsChatGPT requires Plus or higher; Codex covers all tiers from Plus to Go; GPT-5.5 Pro is only available to Pro/enterprise users.
• Context windowCodex supports 400K tokens; the standard API version supports 1M tokens.
• Security MechanismDeploy the strongest protective measures to date, including a cybersecurity classifier, biological/chemical capability assessment, and external red team testing.
• Hardware dependencyInference is based on the NVIDIA GB200/GB300 NVL72 system. Fast mode requires additional payment (2.5 times the cost, 1.5 times the speed).

The core advantages of GPT-5.5

• Performance leapTerminal-Bench 2.0 achieved a score of 82.7%, SWE-Bench Pro achieved 58.6%, and GDPval achieved 84.9%, all of which are ahead of their predecessors and competitors.
• Efficiency BreakthroughIt has comparable latency to GPT-5.4, consumes significantly fewer tokens to complete the same Codex task, and is more cost-efficient.
• Deep reasoningIt possesses "conceptual clarity," enabling it to understand system architecture, predict failure causes, and anticipate testing requirements without needing to break down tasks step by step.
• Hardware CollaborationDesigned in collaboration with NVIDIA GB200/GB300 NVL72, it achieves deep software and hardware optimization from training to inference.
• Sustained autonomyMaintain focus during long-running, complex tasks, reduce interruptions, and support 400K context windows.

GPT-5.5 project address

• Project official websitehttps://openai.com/index/introducing-gpt-5-5/

Comparison of GPT-5.5 with similar competing products

Application scenarios of GPT-5.5

• Software developmentIt can directly generate runnable applications from natural language requirements and handle large-scale codebase refactoring and cross-branch merging.
• Business OperationsIt automatically analyzes six months of business data, builds scoring frameworks, and reviews tax forms, accelerating finance and communication processes.
• Scientific research: Analyze tens of thousands of sample gene datasets, assist in combinatorial mathematical proofs, and build bioinformatics visualization tools.
• CybersecurityProvides advanced security capabilities for trusted defenders, supporting code auditing, vulnerability patching, and infrastructure protection.
• Daily officeChatGPT enhances the efficiency of knowledge workers by handling complex research, information synthesis, and document analysis.