OpenAI has introduced GPT-5.5, a new model designed for real-world computing tasks including coding, research, data analysis, document creation, and software interaction. It moves beyond single-response AI toward full workflow execution with planning, tool use, and task completion.
The model focuses on understanding intent faster and handling multi-step tasks with minimal guidance. It is designed for agent-based environments where AI systems can continue working across tools until a task is fully completed.
GPT-5.5
GPT-5.5 is an agentic AI model built to operate across software environments, codebases, documents, and data systems. It can interpret goals, break them into steps, and execute them using tools while maintaining long-context awareness.
It improves reasoning efficiency while maintaining GPT-5.4-level latency in real-world serving. The model is also more token-efficient, producing higher-quality outputs with fewer computational steps, especially in coding and structured workflows.
Key features
GPT-5.5 improves performance across coding, knowledge work, scientific research, and computer-based automation. It is designed to plan, execute, verify, and continue tasks across tools with reduced user intervention.
It is especially effective in long-horizon workflows that require reasoning, adaptation, and context tracking. This includes software engineering, research analysis, and enterprise operations involving multiple tools and systems.
Core capabilities:
- Advanced agentic coding and autonomous software execution
- 82.7% accuracy on Terminal-Bench 2.0 command-line workflows
- 58.6% performance on SWE-Bench Pro real GitHub issue solving
- Strong performance on Expert-SWE long-horizon engineering tasks
- Improved understanding of full codebase structure and dependencies
- Better debugging, refactoring, testing, and validation workflows
- Stronger context retention across large systems
- Reduced token usage for equivalent coding tasks
- High efficiency on Artificial Analysis Coding Index (lower cost per output)
- Improved computer-use actions (navigation, typing, tool interaction)
- Strong document, spreadsheet, and presentation generation
- Better structured business and operational planning support
- Improved reasoning in ambiguous multi-step environments
- GPT-5.5 Thinking mode for faster complex reasoning
- GPT-5.5 Pro for higher accuracy and deeper analysis
Beyond coding, GPT-5.5 supports enterprise workflows such as reporting, financial analysis, communication planning, and structured data interpretation. It converts unstructured inputs into structured outputs across business environments.
Internal usage shows strong adoption, with over 85% of OpenAI employees using Codex weekly. It is used for dataset analysis, risk scoring systems, automated Slack workflows, large-scale document processing, and automated reporting that reduces manual workload.
In ChatGPT, GPT-5.5 Thinking improves speed and clarity in complex tasks such as coding, synthesis, research, and analysis. GPT-5.5 Pro improves structure, depth, and accuracy in business, legal, education, and technical workflows.
Benchmark performance:
- GDPval: 84.9% across 44 occupations
- OSWorld-Verified: 78.7% real computer environments
- Tau2-bench Telecom: 98.0% customer service workflows
- FinanceAgent: 60.0%
- Investment banking modeling tasks: 88.5%
- OfficeQA Pro: 54.1%
Scientific research capabilities
GPT-5.5 improves multi-step scientific workflows that require experimentation, reasoning, and iterative analysis over time. It supports hypothesis testing, data exploration, and result interpretation across long research cycles.
- GeneBench: Improved performance in genetics and quantitative biology tasks involving complex datasets, uncertainty, and statistical modeling, often reflecting multi-day scientific workloads.
- BixBench: Strong performance in bioinformatics and real-world biomedical data analysis tasks.
- Internal testing: Contributed to a new proof in combinatorics related to Ramsey numbers, later verified using formal methods, demonstrating structured reasoning in mathematical research tasks.
Early testers used GPT-5.5 Pro as a research assistant for manuscript review, iterative analysis, hypothesis development, and multi-source reasoning across code and documents. It performs better in long research workflows requiring progressive refinement.
Efficiency and infrastructure
GPT-5.5 was designed to maintain GPT-5.4-level latency while improving intelligence and efficiency. It was co-developed with NVIDIA GB200 and GB300 NVL72 systems.
A major optimization involved dynamic workload balancing instead of fixed chunk processing. Codex analyzed real production traffic patterns and developed improved partitioning methods, increasing token generation speed by over 20%.
The model also contributed to optimizing its own serving infrastructure, improving overall system performance and efficiency.
Safety and cybersecurity
GPT-5.5 introduces stronger safety systems under OpenAI’s Preparedness Framework, especially for cybersecurity-related capabilities. It is classified as “High” in cybersecurity potential.
Safety measures include:
- Stronger classifiers for sensitive cyber-related requests
- Improved misuse detection systems
- Controls for repeated abuse patterns
- Authenticated access for verified users
- Monitoring of impermissible usage
The model does not reach “Critical” cybersecurity level but shows improvement over GPT-5.4.
OpenAI is expanding Trusted Access for Cyber via Codex, allowing verified users and organizations to access advanced capabilities under stricter security controls. The company is also working with government partners to support protection of critical infrastructure such as energy systems, water networks, and public digital services.
Pricing and availability
GPT-5.5 is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as Codex. GPT-5.5 Pro is available for Pro, Business, and Enterprise tiers.
In ChatGPT:
- GPT-5.5 Thinking: Plus, Pro, Business, Enterprise
- GPT-5.5 Pro: Pro, Business, Enterprise
In Codex:
- Available across Plus, Pro, Business, Enterprise, Edu, Go
- 400K context window support
- Fast mode: 1.5x speed at 2.5x cost
API (coming soon):
- gpt-5.5: $5 input / $30 output per 1M tokens
- Batch/Flex: 50% reduced cost
- Priority: 2.5x cost
- gpt-5.5-pro: $30 input / $180 output per 1M tokens
- 1M token context window
OpenAI noted that GPT-5.5 is priced higher than GPT-5.4 but delivers higher capability and improved token efficiency across workloads.