OpenAI下一代旗舰模型GPT 6.2 Astra疑似已进入小范围灰度测试。近日，多名海外开发者在Codex使用量统计中发现名为“gpt-6.2-astra”的新模型标识，部分账户的调用占比达到19.8%。与此同时，OpenAI官方公布了一款尚未发布的内部模型在三维工程设计基准测试中的成绩，较现款GPT 6 Astra提升约47.3%，引发外界对新一代模型即将发布的猜测。
据科技媒体TestingCatalog报道，从10月31日起，部分Codex用户的使用量分析页面出现了GPT 6.2 Astra的独立记录，但模型选择界面仍显示GPT 6 Astra。
X平台博主Token Gremlin发布的截图显示，某账户约19.8%的模型调用被计入GPT 6.2 Astra，约80%仍归属于GPT 6 Astra。不同用户截图中的调用比例存在差异，有开发者据此推测，OpenAI可能正在通过后台请求路由开展A/B测试。
此外，11月3日，海外模型追踪账号LuminaBench声称，OpenAI API模型列表中曾短暂出现“rc-ada-or-vm”和“rc-ada-sh-vm”两个新标识，随后消失。部分开发者猜测，这些代号可能与新模型的发布候选版本有关，但目前尚无证据能够确认其对应GPT 6.2 Astra。
值得注意的是，OpenAI尚未正式公布GPT 6.2 Astra，也未确认上述灰度测试及新模型的发布时间。社交平台流传的部分模型演示视频同样存在真实性争议，不能作为新模型性能已经得到验证的依据。
就在灰测传闻持续发酵之际，OpenAI于11月3日发布题为《加速面向物理世界的工程发展》的技术博文，披露了AI模型在机械设计、电子工程及实体制造领域的最新进展。
其中最受关注的是一组尚未发布模型的测试成绩。
在BenchCAD 2.0工程基准测试中，现款GPT 6 Astra在最高推理设置下的平均得分为0.594，而一款被OpenAI称为“更新的内部模型”的平均得分达到0.875，相对提升约47.3%。
BenchCAD 2.0主要评估模型重建三维机械零件及装配体、识别电路板连接关系等能力。OpenAI认为，这一结果初步表明，新一代模型有望进一步提高专业工程设计的迭代效率。
不过，OpenAI并未披露这款内部模型的正式名称，也没有说明其是否就是近期传闻中的GPT 6.2 Astra。上述成绩属于特定工程基准测试结果，尚不能直接代表真实制造任务的整体成功率。
除了公布模型跑分，OpenAI还展示了GPT 6 Astra参与实体硬件设计的案例。
按照官方介绍，两名工程师利用GPT 6 Astra，在不到一周时间内完成了一款便携式AI设备的设计工作。该设备可通过语音与Codex交互，模型参与了外壳结构、内部机械组件、PCB电路板、电子元器件、用户界面及固件软件的设计。
整个设计过程中，工程师没有直接手工修改模型生成的设计。随后，代工厂在不到一个月的时间里制造出了一小批实体原型机。
但OpenAI也承认，将设计转化为可制造产品仍需要多轮人工审查与迭代。GPT 6 Astra尚不能始终准确考虑生产工艺限制、制造公差和装配细节。
OpenAI同时披露，自今年2月以来，其硬件团队使用AI模型消耗的Token数量已经增长23倍。Codex和ChatGPT正被用于计算机辅助设计、制造图纸生成、工程仿真及设计审查等工作。
在外部合作方面，OpenAI宣布与马自达开展新的研究合作，探索利用AI辅助汽车空气动力学及碰撞安全设计。GPT 6 Astra能够生成不同的车头几何造型，并结合气流仿真帮助工程师评估设计方案。
马自达首席工程师中田明宏表示，借助AI探索和模拟数十种设计变体，工程师能够更清楚地理解哪些造型特征会影响空气动力学表现，以及设计师在哪些区域拥有更大的造型自由度。
OpenAI还扩大了与印第赛车队Chip Ganassi Racing的合作，尝试将现有零部件的三维扫描数据转换为可编辑CAD模型和二维制造图纸，以减少工程师在逆向建模及原型开发中的重复工作。
从Codex后台出现的新模型标识，到官方公布的内部模型工程跑分，近期一系列信息显示，OpenAI正在继续推进大模型的编程、计算机操作及实体工程设计能力。
不过，GPT 6.2 Astra是否已经进入正式灰测、内部新模型是否对应这一版本，以及新模型何时发布，目前仍有待OpenAI官方确认。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Open AI ' s next-generation flagship model GPT 6.2 Astra is suspected to have entered the small greyscale test. In recent days, several overseas developers have found in COdex usage statistics a new model label called gpt-6.2-astra, with a transfer rate of 19.8 per cent for some accounts. At the same time, the official Open AI published the results of an internal model that had not yet been publish in a three-dimensional engineering design benchmark test, which increased by approximately 47.3 per cent over the cash GPT 6 Astra, prompting speculation about a new generation of models to be publish.
Technological media TestingCatalog reported that, as of 31 October, some of the Cordex users had independent records of GPT 6.2 Astra on their usage analysis page, but the model selection interface still showed GPT 6 astra.
A screenshot of Token GremlinZ94Z, the owner of platform X, shows that about 19.8 per cent of model calls to an account are recorded in GPT 6.2 Astra and about 80 per cent still belongs to GPT 6 astra. There are differences in the call rates in the different user screenshots, according to which developers assume that Open AI may be in the process of conducting the A/B test by the requested route through the pass.
In addition, on 3 November, Lumina Bench, the offshore model tracking account, claimed that two new markings, “rc-ada-or-vm” and “rc-ada-sh-vm”, had briefly appeared in the Open AI API model list and subsequently disappeared. Some developers speculate that these codes may be related to the publish candidate version of the new model, but there is no evidence to confirm its GPT 6.2 Astra.
It is noteworthy that Open AI has not officially published GPT 6.2 Astra, nor has it confirmed publish time for the aforementioned greyscale test and the new model. Part of the model presentation video that is circulated on social platforms is equally genuine and cannot serve as a basis for new model performance to be validated.
On 3 November, at a time when grey-detected rumours continued to ferment, Open AI disclosed the latest progress of the AI model in the areas of mechanical design, electronic engineering and physical manufacturing in a technical op-ed entitled publish: Accelerating Engineering Development for the Physical World.
Of particular concern was a set of test results that had not yet been modelled publish.
In the BenchCAD 2.0 Project benchmark test, the average score for cash GPT 6 Astra under the highest reasoning setting was 0.594, while the average score for a “updated internal model” referred to by Open AI was 0.875, a relative increase of approximately 47.3 per cent.
BenchCAD 2.0 primarily assesses the ability of the model to rebuild three-dimensional mechanical parts and assembly bodies and to identify circuit board connections. In Open AI ' s view, this finding initially indicates that a new generation of models is expected to further improve the iterative efficiency of professional engineering designs.
However, Open AI did not disclose the official name of the internal model, nor did it state whether it was GPT 6.2 Astra in recent rumours. These achievements are the result of a specific engineering benchmark test and do not yet directly represent the whole of success rate, the real manufacturing mission.
In addition to publishing model run points, Open AI presented the case of GPT 6 Astra participating entity hardware design.
According to official presentations, two engineers, using GPT 6 Astra, completed the design of a portable AI device in less than a week. The device interacts with Codex by voice, and the model is involved in the design of the shell structure, internal mechanical components, PCB circuit boards, electrons, user interfaces and solidware software.
Throughout the design process, the engineer did not directly manually modify the design generated by the model. Subsequently, a small group of entity prototype machines were manufactured in less than a month.
However, Open AI also recognizes that the conversion of designs to manufactured products still requires multiple rounds of manual review and iterative. GPT 6 Astra has not always been able to give accurate consideration to process limitations, transport and assembly details.
Open AI also disclosed that the number of Tokens consumed by its hardware team using the AI model has increased 23 times since February this year. Codex and ChatGPT are being used for computer-assisted design, production of drawings, engineering simulations and design reviews.
In terms of external cooperation, Open AI announced new research collaboration with Maada to explore AI-assisted aerodynamics and collision safety designs. GPT 6 Astra is capable of generating different head geometry models and helping engineers to evaluate design options in combination with air flow simulations.
Mada ' s chief engineer, Minami Nakada, stated that, through AI exploration and simulation of dozens of design variants, engineers were able to understand more clearly which configuration features affected aerodynamic performance and where the designer had greater form freedom.
Open AI has also expanded its collaboration with the Indy team Chip Ganassi Racing to try to convert 3D scanning data from existing spare parts into editable CAD models and 2D production drawings to reduce the duplication of engineers in reverse modelling and prototype development.
From the appearance of new model markings in the backstage of Codex to official internal model engineering runs, a series of recent information shows that Open AI is continuing to advance the programming, computer operations and physical engineering capabilities of LLM.
However, whether GPT 6.2 Astra has entered the official ash survey, whether the new internal model corresponds to this version, and when the new model will be publish is still pending official confirmation from Open AI.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.