GPT-5.5

Recent developments

• 2026-11-05 — OpenAI releases GPT-5.5 with 40% token savings and improved benchmark scores
• 2026-11-05 — Security flaws found in combined AI agent skills, even when individual skills are safe
• 2026-11-05 — LLM code generation fails on verification, new paper finds
• 2026-11-05 — Cheaper AI models like Kimi show cost advantage over GPT-5.5 in bug detection
• 2026-11-05 — LLM cost-effectiveness hinges on task success, not just token price
• 2026-10-19 — OpenAI's GPT-5.5 retired from ChatGPT, but remains available via API

Frequently asked

What are the key features and improvements of the newly released GPT-5.5?

The latest GPT-5.5 model represents a full retraining, bringing significant advancements. It offers a 40% reduction in token usage for comparable tasks, enhancing efficiency. Furthermore, it achieved an impressive 82.7% score on the Terminal-Bench 2.0 benchmark, indicating improved performance. However, users should note that its output token price is approximately double that of GPT-5.4, requiring a careful evaluation of cost-effectiveness based on specific use cases and scale.

How does GPT-5.5 perform in code generation and security vulnerability detection?

In code generation, GPT-5.5 is part of a broader trend where LLM-generated code often fails verification despite high acceptance rates, suggesting current evaluation metrics are insufficient. For security vulnerability detection, while cheaper models like Kimi K2.6 can find more bugs on a given budget, GPT-5.5 demonstrates superior consistency, successfully identifying issues that other models might miss. It also plays a role in automating critical tasks like backporting Linux security patches, enhancing system security.

Is GPT-5.5 a cost-effective choice compared to other available models?

The cost-effectiveness of GPT-5.5 is nuanced. While its output token price is higher than its predecessor, GPT-5.4, studies emphasize that true cost-effectiveness should be measured by "cost per successful task," accounting for retries and failures. GPT-5.5, alongside models like Kimi K3, is considered a premium model due to its higher pass rates and reliability. This suggests that for tasks requiring high accuracy and consistency, its overall value can outweigh a higher per-token cost. Additionally, third-party API providers offer competitive rates.

What are the current reliability issues and challenges associated with GPT-5.5 as an AI agent?

GPT-5.5, when deployed as an AI agent, faces notable reliability challenges. Users have reported frustrating experiences in development environments like Cursor IDE, where agents powered by GPT-5.5 may exhibit unpredictable behavior, such as deleting code or failing to complete refactoring tasks correctly. Furthermore, recent analyses indicate a concerning trend of increasing hallucination rates across flagship models, including GPT-5.5 and Gemini 3-Pro, suggesting they may prioritize generating helpful-sounding responses over factual accuracy, leading to fabricated information.

Related