本文来源:https://www.163.com/dy/article/LC09PQ3R0511BLFD.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
OpenAI正式发布两款全新人工智能模型GPT-6.2 Astra和GPT-6.2 Sol,并宣称新模型在编程、科研、专业办公以及计算机操作等多个领域实现显著进步。与此同时,新模型能够以更少的Token消耗完成更多任务,让用户在现有使用额度下获得更高的工作效率。
今年9月,OpenAI推出GPT-6 Astra,并将其称为一次具有代际意义的能力跃升。然而,随着竞争加剧,其他AI实验室迅速缩小差距。其中,Anthropic的Claude Fable 5.2在多个编程基准测试中超越Astra,而Google的Gemini 4 Carbon则凭借更低的价格提供了极具竞争力的性能表现。
面对市场挑战,OpenAI此次推出GPT-6.2 Astra,目标是在重新争夺性能领先地位的同时,进一步提高成本效益。
根据OpenAI公布的数据,在评估复杂多步骤专业工作流程能力的Agents' Last Exam测试中,GPT-6.2 Astra的成绩较GPT-6 Astra提升了5.2%,而执行成本不到后者的三分之一。与Claude Fable 5.2相比,GPT-6.2 Astra成绩高出5.5个百分点,而单项任务成本仅为对方的约5%。在Terminal-Bench-Science 0.1科学测试中,新模型成绩也比GPT-6 Astra高出9%,成本依然不足后者的三分之一。
除了基准测试成绩提升之外,OpenAI还表示,GPT-6.2 Astra在专业办公场景中的实用性进一步增强。该模型能够生成设计质量更高的PowerPoint演示文稿、Excel电子表格以及各类文档,并能更精准地遵循现有模板和行业规范。同时,其计算机操作能力和视觉推理能力也获得升级。
针对软件开发领域,OpenAI称GPT-6.2 Astra能够编写结构更清晰的代码,做出更合理的软件架构决策,并提供更有针对性的测试方案,从而减少完成复杂开发任务所需的迭代次数。
与主要面向复杂专业工作的GPT-6.2 Astra不同,GPT-6.2 Sol则瞄准日常办公与通用生产力场景。在FrontierCode 2.0编程基准测试中,GPT-6.2 Sol的成绩仅落后GPT-6.2 Astra 0.5个百分点,但每项任务成本仅约为后者的五分之一。在OpenAI自研的LifeVal基准测试中,Sol的表现甚至超过GPT-6 Astra,而成本约仅为后者的七分之一。
在定价方面,OpenAI宣布将GPT-6.2 Astra标准API的输入和输出Token价格较GPT-6 Astra下调20%,缓存输入Token价格则下降60%。GPT-6.2 Sol则维持与GPT-6.1 Sol相同的价格水平。
OpenAI表示,两款新模型即日起向ChatGPT Plus、Pro、Enterprise以及Edu订阅用户开放,并可通过ChatGPT Work和Codex使用。其中,Pro 500订阅用户可在两款模型上启用Ultrafast超高速模式,而Pro 200订阅用户现已能够在GPT-6.2 Sol上使用该功能。
随着Anthropic Claude Fable 5.2和GoogleGemini 4 Carbon带来的竞争压力不断增加,GPT-6.2 Astra与GPT-6.2 Sol的推出被视为OpenAI在性能、效率与成本三方面发起的新一轮竞争攻势,力图重新确立其在高端人工智能模型市场中的领先地位。
本文来源:https://www.163.com/dy/article/LC09PQ3R0511BLFD.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
OpenAI has officially released two new AI models, GPT-6.2 Astra and GPT-6.2 Sol, claiming that the new models have achieved significant advancements in various fields, including programming, scientific research, professional office work, and computer operations. At the same time, the new models can complete more tasks with fewer tokens, allowing users to achieve higher productivity within their existing usage limits.
In September of this year, OpenAI launched GPT-6 Astra, describing it as a generational leap in capabilities. However, as competition intensifies, other AI labs are rapidly closing the gap. Among them, Anthropic's Claude Fable 5.2 outperformed Astra in several programming benchmarks, while Google's Gemini 4 Carbon delivered highly competitive performance at a lower price.
In response to market challenges, OpenAI has launched GPT-6.2 Astra with the goal of regaining its performance lead while further improving cost-effectiveness.
According to data released by OpenAI, in the Agents' Last Exam test, which assesses the ability to handle complex, multi-step professional workflows, GPT-6.2 Astra outperformed GPT-6 Astra by 5.2%, while its execution cost was less than one-third of that of the latter. Compared to Claude Fable 5.2, GPT-6.2 Astra scored 5.5 percentage points higher, while the cost per task was only about 5% of that of its counterpart. In the Terminal-Bench-Science 0.1 science test, the new model also outperformed GPT-6 Astra by 9%, while still costing less than one-third as much.
In addition to improved benchmark scores, OpenAI also stated that GPT-6.2 Astra's practicality in professional office settings has been further enhanced. The model is capable of generating PowerPoint presentations, Excel spreadsheets, and various other documents with higher design quality, and can adhere more precisely to existing templates and industry standards. At the same time, its computer operation capabilities and visual reasoning capabilities have also been upgraded.
In the field of software development, OpenAI claims that GPT-6.2 Astra can write code with clearer structure, make more reasonable software architecture decisions, and provide more targeted test plans, thereby reducing the number of iterations required to complete complex development tasks.
Unlike GPT-6.2 Astra, which is primarily aimed at complex professional tasks, GPT-6.2 Sol targets everyday office work and general productivity scenarios. In the FrontierCode 2.0 programming benchmark, GPT-6.2 Sol scored only 0.5 percentage points lower than GPT-6.2 Astra, but the cost per task was only about one-fifth of that of the latter. In OpenAI's proprietary LifeVal benchmark, Sol even outperformed GPT-6 Astra, at about one-seventh the cost.
In terms of pricing, OpenAI announced that the input and output token prices for the GPT-6.2 Astra standard API will be reduced by 20% compared to GPT-6 Astra, while the cached input token price will be reduced by 60%. GPT-6.2 Sol maintains the same price level as GPT-6.1 Sol.
OpenAI stated that the two new models are now available to subscribers of ChatGPT Plus, Pro, Enterprise, and Edu, and can be accessed through ChatGPT Work and Codex. Among them, Pro 500 subscribers can enable Ultrafast mode on both models, while Pro 200 subscribers can now use this feature on GPT-6.2 Sol.
With increasing competitive pressure from Anthropic's Claude Fable 5.2 and Google's Gemini 4 Carbon, the launch of GPT-6.2 Astra and GPT-6.2 Sol is seen as OpenAI's new competitive offensive in terms of performance, efficiency, and cost, as it seeks to re-establish its leading position in the high-end AI model market.
