AI BriefingOpenAI Unveils GPT-5.5
Key point
OpenAI has unveiled GPT-5.5, boosting performance in coding, work tasks, and scientific research.
Details
OpenAI has unveiled GPT-5.5. It was introduced as a model that plans and follows through on coding, online research, data analysis, document/spreadsheet creation, software manipulation, and multi-tool tasks more quickly.
Key points:
- The company explained that it shows higher intelligence than GPT-5.4 while maintaining similar per-token latency in real-world service conditions.
- On Codex tasks, it completes the same work using fewer tokens, improving efficiency as well.
- Safety was reinforced through internal and external red teaming, advanced cybersecurity/biology testing, and feedback from about 200 early partners.
- It is being rolled out sequentially to Plus, Pro, Business, and Enterprise users via ChatGPT and Codex, with API access planned soon.
Key benchmarks:
- Terminal-Bench 2.0 82.7% vs GPT-5.4 75.1%
- Expert-SWE 73.1% vs 68.5%
- GDPval 84.9% vs 83.0%
- OSWorld-Verified 78.7% vs 75.0%
- CyberGym 81.8% vs 79.0%
- BrowseComp 84.4%, FrontierMath Tier 4 35.4%, Tau2-bench Telecom 98.0%
GPT-5.5 Pro recorded BrowseComp 90.1%, FrontierMath Tier 1–3 52.4%, and Tier 4 39.6%. The company said ChatGPT's GPT-5.5 Thinking provides faster, more concise answers, while the Pro version produces more comprehensive and structured results.
On the scientific research side, it showed improvements on GeneBench and BixBench, and the company noted that an internal version was also used to explore new proofs related to Ramsey numbers.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.