While the competition in large language models (LLMs) is still focused on "who answers more accurately," OpenAI has once again taken a step towards defining industry rules. OpenAI officially announced...The new generation models "GPT-5.5" and "GPT-5.5 Pro"The official definition is that it is a key upgrade towards a "completely new way of working with computers". It is no longer just a chatbot, but an "Agentic AI" that can autonomously plan, operate software, debug, and even conduct scientific research.
At the same time, NVIDIA also released a message confirming the deep cooperation between the two parties at the underlying hardware and software level. More than 10,000 NVIDIA employees were the first to experience the powerful computing power that has revolutionized traditional workflows on the latest GB200 NVL72 system.
Programming and Operating Computers: From Debugging to "One-Click Solution"
The biggest leap forward in GPT-5.5 lies in its "Agentic" capability for performing complex, multi-step tasks.
According to data released by OpenAI, GPT-5.5 is their most powerful code proxy model to date. In Terminal-Bench 2.0, which tests complex command-line workflows, GPT-5.5 achieved an astonishing accuracy of 82.7%; it also set a new high of 58.6% in SWE-Bench Pro, which solves real-world GitHub project problems.
Early tester feedback indicated that GPT-5.5 exhibited an unprecedented "systems perspective." When faced with a massive codebase, it could accurately understand the interconnected architectural logic.
Previously, engineers might have spent days tracking down and refactoring system errors. GPT-5.5 can now complete hundreds of front-end and refactoring changes in a "one-shot" manner within 20 minutes and successfully merge branches. In addition, it can now be integrated with Codex to demonstrate powerful "computer operation" capabilities, as if a real person is sitting next to you, clicking, typing, and switching between applications for you.
Expanding into scientific research: More than just searching, it's about becoming "co-scientists."
Besides programming, GPT-5.5 also performs remarkably well in knowledge work and scientific research.
It has achieved industry-leading results in benchmark tests for bioinformatics and genetic data analysis (such as GeneBench and BixBench). OpenAI even demonstrated a highly indicative case: the internal version of GPT-5.5 successfully helped discover a new mathematical proof about Ramsey numbers (a core problem in combinatorics), showing that GPT-5.5's computational logic has gone beyond simple "data integration" and has the ability to provide substantial mathematical inferences and new insights in top academic fields.
NVIDIA's Deep Endorsement: Enterprise-Level Practices Under the GB200 Architecture
Another major highlight of this GPT-5.5 release is its deep partnership with NVIDIA, the AI chip giant.
NVIDIA官方新聞稿指出，GPT-5.5是一套專為NVIDIA GB200與GB300 NVL72系統「共同設計、訓練與服務」的模型。在正式發表前，NVIDIA已經讓內部1萬名員工透過Codex方式率先使用GPT-5.5。
To address the most critical cybersecurity concerns for enterprises, NVIDIA adopted a "zero data retention" model and maintained read-only access to production systems. In this extremely secure environment, the NVIDIA team successfully reduced debugging cycles from several days to just "hours," and complex multi-file reconstructions could even be completed automatically by AI overnight.
This is not just a software trial; it's more like NVIDIA demonstrating to its global enterprise customers: "Buy my GB200 server, and you can securely have such a tireless super AI development team."
Performance enhancement, cybersecurity control and billing solutions
Surprisingly, despite the significant improvement in intelligence, GPT-5.5 maintains the same token latency in actual deployment as its predecessor, GPT-5.4, and even significantly reduces the number of tokens used when completing the same Codex task.
In terms of cybersecurity, OpenAI rates its cybersecurity and biohazard capabilities as "high" due to the model's powerful vulnerability finding and patching capabilities. To this end, OpenAI launched the "Trusted Access for Cyber" program, allowing proven critical infrastructure defenders priority access to higher-privilege models (such as GPT-5.4-Cyber) to help combat the increasingly rampant cyberattacks.
目前，GPT-5.5與GPT-5.5 Pro已經陸續向ChatGPT與Codex的Plus、Pro、Business與Enterprise訂閱用戶開放。針對開發者的API定價，GPT-5.5為每百萬輸入詞元收取5美元、輸出費用為30美元；而追求極致精準度的GPT-5.5 Pro則採高達每百萬輸入收取30美元、輸出為180美元價格。
Analysis of viewpoints
The release of GPT-5.5 heralds the end of the era of "AI as a tool" and officially ushers in a new era of "AI as a workforce".
In the past, using ChatGPT was like consulting an extremely intelligent encyclopedia; but using GPT-5.5 is more like contracting a project to a seasoned engineering team. This will have a devastating yet creative impact on the overall software development, biopharmaceutical, and data analytics industries.
What's even more intriguing is the symbiotic relationship between OpenAI and NVIDIA. OpenAI uses NVIDIA's hardware to train models, while in turn, GPT-5.5 helps NVIDIA write better load balancing algorithms, further increasing the AI generation speed of GPUs by more than 20%.
This positive spiral of "AI helping to upgrade computing infrastructure, and computing power further advancing AI intelligence" has built an extremely difficult-to-cross hardware and software ecosystem barrier for other competitors (such as Google Gemini or Anthropic Claude).
