大模型
ghacks.net ·
@Arthur Kay
OpenAI发布GPT-6 Astra，宣称在计算机使用、编程和科学领域开辟新前沿
OpenAI宣布推出其最新模型GPT-6 Astra，称其为迄今最智能、最对齐的模型。该报道来自gHacks，但未提供具体技术细节或性能数据。
AI推荐理由 报道OpenAI发布GPT-6 Astra，但内容简短，缺乏技术细节和独立验证，来源为科技博客，时效性高。对AI领域有较高关注价值。
70
正文
OpenAI发布了GPT-6 Astra，其最新模型，并称这是其迄今为止最智能、最对齐的模型。Astra今天开始向一小部分组织推出，更广泛的访问计划很快实现。
据OpenAI称，Astra在计算机使用、浏览、软件工程、网络安全、科学和专业工作方面处于最先进水平。该模型在FrontierMath Tier 4上得分98%，在ARC-AGI-3上得分99.9%，在ExploitBench上得分100%。
它还在OpenAI的准备框架中达到网络安全的关键阈值，这意味着有额外的保障措施。所有基准结果均来自OpenAI。
Astra将在未来几天内向所有ChatGPT Plus、Pro、Business和Enterprise用户开放。它还将通过OpenAI API、Microsoft Azure和AWS Bedrock提供。
GPT-6 Astra基准声明和计算机使用增益
OpenAI表示Astra在多项基准测试中创下新纪录。在Terminal-Bench Science 0.1上，该测试使用代码和终端工具衡量科学研究工作流，Astra得分64.6%。
Claude Fable 5.1在同一测试中得分52.6%，Astra的API成本估计低约31%。在Agents' Last Exam上，该测试关注复杂的专业软件任务，Astra得分59.3%。Claude Opus 5得分55.5%，GPT-5.6 Sol得分53.6%。Astra在此测试中使用的输出令牌也比Opus 5少约65%。
对于编码，Astra在Terminal-Bench 4.0上得分57.9%。GPT-5.6 Sol得分37.3%，Claude Fable 5.1在同一基准上得分55.8%。在科学方面，Astra在GPQA Diamond上达到96.0%，该测试测试研究生水平的科学推理。OpenAI表示这些结果来自其研究环境或API，因此可能与用户在生产ChatGPT中看到的不一致。
OpenAI称Astra是其最好的计算机使用模型。它可以填写在线表单、更新CRM记录、组织日历、进行研究、分析科学数据、生成图表、创建网站以及运行前端QA检查。
在OSWorld 2.0的延迟测试中，Astra完成任务的速度比GPT-5.6 Sol快约47%。Astra得分72.6%，每个任务约40分钟，而GPT-5.6 Sol得分65.7%，约75分钟。
OpenAI还在更新Codex harness以加快计算机使用速度。该公司表示，与GPT-5.6 Sol在Mind2Web基准上相比，这导致任务完成速度快1.9倍。OpenAI展示了Astra处理诸如KiCad中的PCB布局、Blender和Unreal Engine 5中的房屋建模以及游戏开发等任务。
GPT-6 Astra专业工作、编码和网络安全能力
OpenAI表示Astra针对计算机使用和专业工作进行了训练。它可以生成遵循模板并匹配用户风格的文档、电子表格和演示文稿。在BenchCAD上，该测试衡量模型从多个视图重建3D对象的能力，Astra得分95.9%。GPT-5.6 Sol得分83.3%，Claude Fable 5.1在同一测试中得分84.3%。
OpenAI表示Astra是其迄今为止最好的软件工程模型。借助Astra，Codex增加了一种在上下文窗口已满时保留和检索上下文的新方法。它不是将所有内容压缩成一个摘要，而是跨上下文窗口保留可搜索的笔记。
此功能是实验性的，可以在Codex设置中打开。它很快将成为Astra的默认设置。
OpenAI称Astra是科学发现的重大进步，并分享了关于素数间隙的两个新结果。Astra可以将科学推理与计算机使用相结合，并直接在专业软件中工作以检查数据和探索结果。
对于网络安全，OpenAI表示Astra在其准备框架中达到关键阈值。在没有生产保障措施的情况下，Astra在ExploitBench上得分100%，而GPT-5.6 Sol为78.5%。在ExploitGym上，Astra的成功率为42.4%，而GPT-5.6 Sol为30.3%。
在使用最近漏洞的内部测试中，Astra发现并利用了先前未知的两个零日漏洞，OpenAI现在正在向维护者报告。
今天发布的版本不会执行高级任务，例如创建概念验证漏洞利用。OpenAI计划在未来几周内通过OpenAI Daybreak允许对防御性使用采取限制较少的保障措施。
GPT-6 Astra对齐声明、保障措施和剩余担忧
OpenAI将Astra描述为其最对齐的模型。在受Hugging Face事件启发的评估中，该评估测试面对不可能任务的模型是否会超出其预期范围，OpenAI表示Astra在0%的情况下这样做，而GPT-5.6 Sol在没有生产保障措施的情况下为48%。在计算机使用安全压力测试中，OpenAI报告Astra产生错误对齐结果的概率为2.4%，而Fable 5.1为9.5%，Opus 5为11.5%。
OpenAI指出了一个担忧：Astra的书面推理比GPT-5.6 Sol更难监控，尤其是在旨在让模型逃避监控的测试中。
OpenAI表示这是因为Astra对其书面推理有更多控制，并且可以用更少的步骤解决问题。该公司表示正在认真对待这个问题，并提高可监控性是研究优先事项。
为了提供帮助，OpenAI正在为Astra级模型在生产中添加错误对齐监控。这些分类器检查模型的推理和行动，并可以自动停止看起来未经授权的活动。
OpenAI表示这些检查有时会减慢或停止合法工作，包括防御性网络安全。
定价以及关于新GPT-6需要记住的事项
GPT-6 Astra今天开始向一小部分组织推出。它将在未来几天内向所有ChatGPT Plus、Pro、Business和Enterprise用户开放，以及通过OpenAI API、Microsoft Azure和AWS Bedrock提供。使用包含在当前订阅计划中，额外使用有积分。Pro、Business和Enterprise用户还可以访问GPT-6 Astra Pro。Enterprise管理员可以为其工作区打开Astra，但它在发布时默认关闭。
对于开发者，As
tra 在 OpenAI API 中作为 gpt-6-astra 提供，也可通过 Amazon Bedrock 获取。标准 API 定价为每百万输入 token 10 美元，每百万输出 token 50 美元，缓存读取和写入有单独的费率。
快速模式提供高达标准处理两倍的速度，但价格翻倍。Astra 支持符合条件 API 客户的零数据保留。OpenAI 还在测试私有安全处理，以在保护客户隐私的同时改进监控。
本公告中的所有基准测试结果均来自 OpenAI，并在其研究或 API 环境中运行。这些结果可能与用户在生产环境 ChatGPT 中看到的不一致，不应被视为独立验证。
OpenAI 尚未给出所有层级全面可用的具体日期。通过 OpenAI Daybreak 提供的限制较少的网络安全保障措施计划在未来几周内推出，但目前尚不可用。OpenAI 也承认 Astra 的可监控性降低是一个局限，并表示仍在研究这个问题。继续研究。
查看原文（中文正文为 AI 翻译）
OpenAI has announced GPT-6 Astra, its latest model, which it calls its most intelligent and aligned so far. Astra is rolling out today to a small group of organizations, with wider access planned soon.
According to OpenAI, Astra is state-of-the-art for computer use, browsing, software engineering, cybersecurity, science, and professional work. The model scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench.
It also meets the Critical threshold for cybersecurity in OpenAI's Preparedness Framework, which means extra safeguards are in place. All benchmark results come from OpenAI.
Astra will become available to all ChatGPT Plus, Pro, Business, and Enterprise users in the next few days. It will also be offered through the OpenAI API, Microsoft Azure, and AWS Bedrock.
GPT-6 Astra Benchmark Claims and Computer Use Gains
OpenAI says Astra sets new records in several benchmarks. On Terminal-Bench Science 0.1, which measures scientific research workflows using code and terminal tools, Astra scores 64.6%.
Claude Fable 5.1 scores 52.6% on the same test, with Astra's API cost estimated to be about 31% lower. On Agents' Last Exam, which looks at complex professional software tasks, Astra scores 59.3%. Claude Opus 5 scores 55.5% and GPT-5.6 Sol scores 53.6%. Astra also uses about 65% fewer output tokens than Opus 5 on this test.
For coding, Astra scores 57.9% on Terminal-Bench 4.0. GPT-5.6 Sol scores 37.3% and Claude Fable 5.1 scores 55.8% on the same benchmark. In science, Astra reaches 96.0% on GPQA Diamond, which tests graduate-level scientific reasoning. OpenAI says these results come from its research environment or API, so they may not match what users see in production ChatGPT.
OpenAI calls Astra its best model for computer use. It can fill out online forms, update CRM records, organize calendars, conduct research, analyze scientific data, generate plots, create websites, and run frontend QA checks.
In latency tests on OSWorld 2.0, Astra completes tasks about 47% faster than GPT-5.6 Sol. Astra scores 72.6% at around 40 minutes per task, while GPT-5.6 Sol scores 65.7% at about 75 minutes.
OpenAI is also updating the Codex harness to make computer use faster. The company says this leads to 1.9 times faster task completion compared to GPT-5.6 Sol on the Mind2Web benchmark. OpenAI shows Astra handling tasks like PCB layout in KiCad, house modeling in Blender and Unreal Engine 5, and game development.
GPT-6 Astra Professional Work, Coding, and Cybersecurity Capabilities
OpenAI says Astra is trained for both computer use and professional work. It can produce documents, spreadsheets, and presentations that follow templates and match a user's style. On BenchCAD, which tests how well a model can reconstruct 3D objects from multiple views, Astra scores 95.9%. GPT-5.6 Sol scores 83.3% and Claude Fable 5.1 scores 84.3% on the same test.
OpenAI says Astra is its best model yet for software engineering. With Astra, Codex adds a new way to keep and retrieve context when the context window is full. Instead of compressing everything into one summary, it keeps searchable notes across context windows.
This feature is experimental and can be turned on in the Codex settings. It will become the default for Astra soon.
OpenAI calls Astra a major step forward for scientific discovery and shares two new results on gaps between prime numbers. Astra can combine scientific reasoning with computer use and work directly in specialized software to inspect data and explore results.
For cybersecurity, OpenAI says Astra meets the Critical threshold in its Preparedness Framework. Without production safeguards, Astra scores 100% on ExploitBench, compared to 78.5% for GPT-5.6 Sol. On ExploitGym, Astra has a 42.4% success rate, while GPT-5.6 Sol scores 30.3%.
In an internal test using recent vulnerabilities, Astra found and used two previously unknown zero-day vulnerabilities, which OpenAI is now reporting to the maintainers.
The version released today will not perform advanced tasks like creating proof-of-concept exploits. OpenAI plans to allow less restrictive safeguards for defensive use through OpenAI Daybreak in the coming weeks.
GPT-6 Astra Alignment Claims, Safeguards, and Remaining Concerns
OpenAI describes Astra as its most aligned model. In an evaluation informed by the Hugging Face incident, which tested whether a model facing an impossible task would go beyond its intended scope, OpenAI says Astra did so in 0% of cases, compared with 48% for GPT-5.6 Sol without production safeguards. On a computer-use safety stress test, OpenAI reports Astra produced misaligned outcomes 2.4% of the time, versus 9.5% for Fable 5.1 and 11.5% for Opus 5.
OpenAI points out one concern: Astra's written reasoning is harder to monitor than GPT-5.6 Sol's, especially in tests designed to make the model evade monitoring.
OpenAI says this is because Astra has more control over its written reasoning and can solve problems with fewer steps. The company says it is taking this issue seriously and that improving monitorability is a research priority.
To help, OpenAI is adding misalignment monitoring in production for Astra-class models. These classifiers check the model's reasoning and actions and can automatically stop activity that looks unauthorized.
OpenAI says these checks can sometimes slow down or stop legitimate work, including defensive cybersecurity.
Pricing And What to Keep in Mind About The New GPT-6
GPT-6 Astra is rolling out today to a small group of organizations. It will be available to all ChatGPT Plus, Pro, Business, and Enterprise users in the next few days, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock. Usage is included in current subscription plans, with credits for extra use. Pro, Business, and Enterprise users also get access to GPT-6 Astra Pro. Enterprise admins can turn on Astra for their workspace, but it is off by default at launch.
For developers, Astra is available in the OpenAI API as gpt-6-astra and through Amazon Bedrock. Standard API pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes.
Fast mode offers up to twice the speed of standard processing at double the price. Astra supports Zero Data Retention for eligible API customers. OpenAI is also testing Private Safety Processing to improve monitoring while protecting customer privacy.
All benchmark results in this announcement are from OpenAI and were run in its research or API environment. These may not match what users see in production ChatGPT and should not be considered independent verification.
OpenAI has not given a specific date for full availability across all tiers. The less-restrictive cybersecurity safeguards through OpenAI Daybreak are planned for the next few weeks, but are not available yet. OpenAI also acknowledges that Astra's reduced monitorability is a limitation and says it is still researching this issue.uing to research.