November 20 news.OpenAI Yesterday (19 November) published a book announcing the launch of GPT-51-Codex-Max smartsProgramming ModelThis has significantly enhanced long-term reasoning, efficiency and real-time interaction。In addition, the model will replace GPT-51-Codex as the default model on the Cordex integration interface。
1AI quotes a Boven presentation, which follows Google Gemini 3 Pro, but in several key programming benchmarking tests, Codex-Max showed greater strength. For example, in the SWE-Bench Verified test, which measures the solution of actual software problems, Codex-Max leads 76.2% with a small accuracy of 779%。
Codex-Max also took the lead in the Terminal-Bench 2.0 test, with an accuracy rate of 58.1%, while Gemini had an accuracy rate of 54.2%; in the LiveCodeBench Pro (a competitive coding Elo benchmark test) it scored the same score as Gemini ' s 2439。
A major structural upgrade of GPT-51-Codex-Max was the introduction of a mechanism called “compression”. The mechanism allows the model to keep the key context information intelligently and discard irrelevant details when it approaches its context window limitations, thus enabling continuous work across millions of tokens without a decline in performance。
Thanks to this, the model has been successfully completed in internal tests for complex tasks lasting more than 24 hours, such as multi-step code re-engineering and autonomous debugging. At the same time, this technology has enhanced the efficiency of some 30% token, effectively reducing costs and delays。
The new model is now integrated into the various Codex development environments of OpenAI itself, including its official command line tool (Codex CLI), its internal code review tool and various interactive programming environments。
Through these tools, developers can experience strong real-time interaction of models, such as intensive learning training or simulation of optical laws in visual interfaces. However, GPT-51-Codex-Max has not yet been made available through public API, but official indications are that it will be opened soon. Ordinary users need to subscribe to payment schemes such as ChatGPT Plus, Pro or Enterprise。
OpenAI revealed that its internal 95% engineers use Codex every week and that since their introduction, these engineers have submitted, on average, about 70% pull requests (Pull Requests), significantly increasing the rate of internal development。
Although Codex-Max has a high degree of autonomy, OpenAI emphasizes that it should be used as a coding “assist” rather than a human alternative. To ensure transparency, the model generates detailed terminal logs and test references to enable developers to review and validate all the codes they generate. In addition, the model operates in a strict sandbox environment by default, and network access is disabled to ensure safety。