OpenAI GPT-5.5 Release Notes

OpenAI has released GPT-5.5 and GPT-5.5 Pro, marking a shift toward agentic AI capable of planning and executing multi-part tasks across various software tools. These models deliver higher intelligence and improved token efficiency in coding and knowledge work without increasing per-token latency compared to GPT-5.4.

Agentic Coding and Software Engineering

GPT-5.5 is designed as a high-autonomy coding model capable of managing complex, long-horizon engineering tasks. It demonstrates a stronger ability to understand system architecture, reason through ambiguous failures, and maintain context across large codebases.

Performance Benchmarks

• Terminal-Bench 2.0: Achieves 82.7% accuracy in complex command-line workflows.
• SWE-Bench Pro: Reaches 58.6% in resolving real-world GitHub issues.
• Expert-SWE (Internal): Outperforms GPT-5.4 on tasks with a median human completion time of 20 hours.

Real-World Application

Early testers report a step-change in conceptual clarity. For instance, the model successfully performed a complex branch merge involving hundreds of frontend and refactor changes in 20 minutes. Users from NVIDIA and Cursor have noted its persistence and ability to predict testing and review needs without explicit prompting.

Knowledge Work and Computer Use

GPT-5.5 integrates better intent understanding with computer-use capabilities, allowing it to navigate interfaces, click, type, and move across tools to complete knowledge work loops.

Professional Benchmarks

• GDPval: Scores 84.9% in producing well-specified knowledge work across 44 occupations.
• OSWorld-Verified: Reaches 78.7% in operating real computer environments autonomously.
• Tau2-bench Telecom: Achieves 98.0% on complex customer-service workflows without prompt tuning.

Internal Deployment

OpenAI reports that over 85% of its staff use Codex weekly. Use cases include the Finance team reviewing over 71,000 pages of K-1 tax forms (accelerating the task by two weeks) and the Communications team automating low-risk speaking request handling via a Slack agent.

Scientific Research and Mathematical Discovery

GPT-5.5 functions as a "co-scientist," persisting through the iterative loop of exploring ideas, gathering evidence, and interpreting results.

Technical Achievements

• Mathematical Proof: An internal version of GPT-5.5 discovered a new proof regarding off-diagonal Ramsey numbers in combinatorics, which was later verified in Lean.
• Bioinformatics: The model achieved leading performance on BixBench and showed significant improvement over GPT-5.4 on GeneBench, specifically in multi-stage scientific data analysis in genetics.
• Custom Tooling: Researchers have used the model to build complex algebraic-geometry apps from single prompts in minutes, visualizing quadratic surface intersections.

Inference Efficiency and Infrastructure

GPT-5.5 was co-designed and trained on NVIDIA GB200 and GB300 NVL72 systems. To maintain GPT-5.4 latency levels, OpenAI optimized the inference stack using the models themselves.

Codex was used to analyze production traffic patterns and write custom heuristic algorithms for load balancing and partitioning. This optimization increased token generation speeds by over 20%.

Cybersecurity and Safety Frameworks

OpenAI has categorized the biological, chemical, and cybersecurity capabilities of GPT-5.5 as "High" under its Preparedness Framework, though it did not reach the "Critical" level for cybersecurity.

Safety Measures

• Stricter Classifiers: Deployment includes tighter controls around high-risk activity and sensitive cyber requests.
• Trusted Access for Cyber: Verified users and organizations defending critical infrastructure can apply for "cyber-permissive" models (such as GPT-5.4-Cyber) to reduce friction for legitimate defensive work.
• Governance: The model underwent full safety processing, including red-teaming and targeted testing for advanced biology and cybersecurity.

Availability and Pricing

Access

• ChatGPT: GPT-5.5 is available to Plus, Pro, Business, and Enterprise users. GPT-5.5 Pro is available to Pro, Business, and Enterprise users.
• Codex: Available for Plus, Pro, Business, Enterprise, Edu, and Go plans with a 400K context window.
• API: GPT-5.5 and GPT-5.5 Pro are available in the API (as of April 24, 2026).

API Pricing

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context Window
gpt-5.5	$5	$30	1M
gpt-5.5-pro	$30	$180	1M

Batch and Flex pricing are available at 50% of the standard rate, while Priority processing is 2.5x the standard rate.