OpenAI releases GPT-5.5: 82.7% Terminal-Bench refreshes SOTA, Agent programming capabilities have greatly improved

Core conclusion

OpenAI released GPT-5.5 on April 23, 2026, claiming it to be the "smartest and easiest to use" model. The new model achieves 82.7% SOTA accuracy on Terminal-Bench 2.0 and 73.1% in Expert-SWE internal evaluation, while significantly reducing token consumption while maintaining the same inference latency as GPT-5.4. The model has begun to be gradually pushed to Plus/Pro/Business/Enterprise users in ChatGPT and Codex, and the API price is set at $5/ million input tokens and $30/ million output tokens.

Background and trigger events

Event Highlights:

• GPT-5.5 is the successor version of GPT-5.4, focusing on improving agentic coding, computer use and knowledge work capabilities
• Same latency as GPT-5.4 but smarter and more token efficient
• Analyze several weeks of production traffic patterns through Codex, write custom heuristic algorithms to optimize GPU load balancing, and increase token generation speed by 20%+
• Based on NVIDIA GB200/GB300 NVL72 system training and inference
• The system card has been released, and the biological/chemical and cyber security capabilities have been evaluated as High (not Critical)

Key Impact

Technical Impact Table

Industry Impact Table

Adaptation suggestions

To Developers:

• GPT-5.5 can generate tokens 1.5x faster in Codex via Fast mode (cost 2.5x)
• API pricing $5/ million inputs, $30/ million outputs, Batch/Flex half price, Priority 2.5x
• GPT-5.5 Pro API pricing $30/ million inputs, $180/ million outputs, for higher accuracy requirements

For users:

• GPT-5.5 Thinking in ChatGPT has been opened to Plus/Pro users
• GPT-5.5 Pro is only available to Pro/Business/Enterprise users

Points worthy of attention:

• OpenAI is building a "global infrastructure for agentic AI"
• GPT-5.5 is used to improve its own inference infrastructure - the model helps improve the system that serves it, forming a positive cycle
• Mathematical discovery ability: GPT-5.5 assisted in discovering a new proof of Ramsey number, which was later verified by Lean. It was the first public example of its participation in core mathematics research.

Tool entry

This release involves multiple AI ecological tools:

• OpenAI — publisher of GPT-5.5, models available via ChatGPT and Codex
• Anthropic — Claude Opus 4.7 has been compared in several benchmarks
• NVIDIA — Provides GB200/GB300 NVL72 system to support GPT-5.5 training and inference
• Codex — OpenAI’s AI programming agent, GPT-5.5 demonstrates the strongest agentic programming capabilities in Codex
• n8n — Workflow automation tool, closely related to the AI agent ecosystem

Industry reaction

• Dan Shipper (Founder and CEO, Every): "The first programming model with serious conceptual clarity."
• Justin Boitano (VP of Enterprise AI, NVIDIA): "GPT-5.5 delivers the sustained performance needed to perform intensive work...enabling teams to move from natural language prompts to end-to-end feature releases."
• Brandon White (co-founder and CEO of Axiom Bio): "If OpenAI continues to develop like this, the foundation of drug discovery will have changed by the end of the year."
• HN Community: GPT-5.5 topped the HN homepage with 1416 points. Developers are concerned about its substantial lead of 82.7% on Terminal-Bench (compared to 69.4% for Claude Opus 4.7), but are concerned that it still lags behind Claude Opus 4.7 on SWE-Bench Pro (64.3% vs 58.6%). Several users noticed that ARC-AGI-3 scores were not published.

Action guidance

Want to learn more about practical applications? Please check out:

• Tutorial: How to configure GPT-5.5 in Codex for efficient programming
• Case: A real-life case of using GPT-5.5 to automate and optimize development workflow