OpenAI 发布 GPT-6 Astra：1.05M 上下文的计算机操作模型，因触及 Critical 网络安全阈值而限制访问

今天，OpenAI 发布了 GPT-6 Astra：https://openai.com/index/gpt-6-astra/。公司称它是最智能、最符合预期的模型，并主要将其定位为计算机使用系统，而不是聊天模型。其宣传点是，Astra 能像人一样操作软件，涵盖浏览器、电子表格、桌面应用程序和终端，并完成多步骤任务，而不是仅描述如何执行这些任务。

它可以部署吗？部分可以，但无法在你的硬件上部署。Astra 是一个封闭的、托管的模型，没有发布权重，因此不能自托管。它目前仅对 OpenAI 的 Trusted Access 和 Daybreak：https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/ 计划中的组织开放。

对开发者来说，主要变化是上下文处理。Codex 以前使用压缩，将上下文满时的早期对话总结。此过程会丢弃代理后续需要的细节：修复失败的原因、运行了哪些测试、早期添加的需求。Astra 改为在上下文窗口中保留笔记，并回溯搜索早期消息和工具输出。该功能通过 config.toml 设置以实验形式提供，并将在未来几周成为 Codex 的默认设置。

Astra 还可以在继续进行不依赖答案的工作时向用户提问。这消除了一个常见的代理失败情况，即一个未解决的决策会阻塞整个任务。

在模型页面：https://developers.openai.com/api/docs/models/gpt-6-astra，Astra 列出了 1,050,000 令牌的上下文窗口、128,000 令牌的最大输出，以及 2026 年 4 月 30 日的知识截止日期。输入为文本和图像，输出仅为文本。reasoning.effort 添加了高（high）之上的两个新级别：xhigh 和 max。工具支持涵盖计算机使用：https://developers.openai.com/api/docs/guides/tools-computer-use、托管 shell、应用补丁、技能、MCP 和工具搜索。不支持微调。

OpenAI 报告 OSWorld V2-Offline 的成绩为 72.6%，而 GPT-5.6 Sol 的成绩为 65.7%，平均任务时间从大约 75 分钟降至 40 分钟。Anthropic 报告 Claude Fable 5.1 的成绩为 77.9%：https://www.anthropic.com/claude/fable，但表示它使用的是不同版本的 OSWorld，因此不应直接比较。

Astra 在 ARC-AGI-3 上得分为 98.6%。这个数据是通过保留每轮推理并对长上下文使用压缩的 Responses API 工具产生的，OpenAI 之前曾展示过：https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/ 那些设置在不改变模型的情况下大幅提升了 ARC-AGI-3 的分数。这个结果衡量的是模型加上代理系统的整体表现。

其他报告的数据：FrontierMath Tier 4 上得分为 97.6%：https://epoch.ai/benchmarks/frontiermath-tier-4-v2，BenchCAD 得分为 95.9%：https://benchcad.com/news.html Vision2Code 对比 Fable 5.1 的 84.3%，Terminal-Bench Science 得分为 64.6%：https://www.tbench.ai/news/terminal-bench-science-0-1 对比 Anthropic 报告的 52.6%。Epoch AI 指出 OpenAI 资助了 FrontierMath，并拥有其部分内容的独家访问权。

编程是当前的弱项。Astra 在 DeepSWE v1.1 上得分为 74.1%，而 Sol 得分为 70.8%。Meta 报告 Muse Spark 1.3 在最大推理能力下得分 75.4%，而公开排行榜：https://deepswe.datacurve.ai/ 将 Gemini 3.8 Flash 和 Claude Opus 5 指在接近 74%。在一个包含 113 项任务的基准测试中，这些差距仅为一两项任务。

Astra 是 OpenAI 首个被认定在其准备框架中达到关键网络安全门槛的模型。在测试中，它开发了针对加固浏览器和操作系统的漏洞利用，并发现了两个此前未知的 V8 漏洞，OpenAI 表示已将其披露给维护者。

其影响具有实际意义。标准访问权限拒绝高级网络安全工作，包括漏洞发现。对于 API 开发者，网络安全检查会直接停止任务，而不是暂停以等待批准。OpenAI 的 Mia Glaese 警告称，非受信访问计划的用户可能会遇到速度变慢、暂停或阻塞，有时甚至是在进行无关工作的过程中。

OpenAI 在 ExploitBench 上报告 100% 分数：https://exploitbench.ai/，这是一个综合能力覆盖评分，而非通过率；在 ExploitGym 上得分为 42.4%，而 Sol 的得分为 30.3%，两者都取消了通常的六小时时间限制。

Astra 的价格为每百万输入令牌 10 美元，每百万输出令牌 50 美元，缓存输入令牌价格为 1 美元。超过 272K 输入令牌的请求按整个请求的输入 2 倍、输出 1.5 倍计费。Batch 和 Flex 模式按 50% 收费，Fast 模式按 2 倍收取。Pro、Business 和 Enterprise 用户还可获得 Astra Pro。

需要与我们合作推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等吗？请联系我们：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一位富有远见的企业家和工程师，Asif 致力于利用人工智能的潜力为社会公益服务。他最近的努力是推出人工智能媒体平台 Marktechpost，该平台以深入报道机器学习和深度学习新闻而闻名，这些报道既技术可靠，又易于广大受众理解。该平台每月浏览量超过 200 万次，显示出其在受众中的受欢迎程度。

Today, OpenAI released GPT-6 Astra：https://openai.com/index/gpt-6-astra/. The company calls it its most intelligent and aligned model, and positions it primarily as a computer-use system rather than a chat model. The pitch is that Astra operates software the way a person does, across browsers, spreadsheets, desktop applications and terminals, and finishes multi-step jobs instead of describing how to do them.

Is it deployable? Partly, and not on your own hardware. Astra is a closed, hosted model with no released weights, so self-hosting is not an option. It is live today only for organizations in OpenAI’s Trusted Access and Daybreak：https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/ programs.

The main change for devs is context handling. Codex previously used compaction, summarizing earlier turns once context filled up. That process discards the detail an agent later needs: why a fix failed, which tests ran, which requirement was added early. Astra instead keeps notes across context windows and searches back into earlier messages and tool output. The feature ships experimental behind a config.toml setting and becomes the Codex default in the coming weeks.

Astra can also ask the user a question while continuing work that does not depend on the answer. That removes a common agent failure where one unresolved decision stalls an entire job.

On the model page：https://developers.openai.com/api/docs/models/gpt-6-astra, Astra lists a 1,050,000-token context window, 128,000 max output tokens and an April 30, 2026 knowledge cutoff. Input is text and image, output is text only. reasoning.effort adds two new levels above high : xhigh and max . Tool support covers computer use：https://developers.openai.com/api/docs/guides/tools-computer-use, hosted shell, apply patch, skills, MCP and tool search. Fine-tuning is not supported.

OpenAI reports 72.6% on OSWorld V2-Offline against 65.7% for GPT-5.6 Sol, with average task time falling from roughly 75 minutes to 40. Anthropic reports 77.9% for Claude Fable 5.1：https://www.anthropic.com/claude/fable but says it used a different OSWorld release and should not be compared directly.

Astra scores 98.6% on ARC-AGI-3. That number was produced with a Responses API harness that retains reasoning between turns and uses compaction for long contexts, and OpenAI has previously shown：https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/ those settings move ARC-AGI-3 scores substantially without changing the model. The result measures the model plus the agent system.

Other reported figures: 97.6% on FrontierMath Tier 4：https://epoch.ai/benchmarks/frontiermath-tier-4-v2, 95.9% on BenchCAD：https://benchcad.com/news.html Vision2Code against 84.3% for Fable 5.1, and 64.6% on Terminal-Bench Science：https://www.tbench.ai/news/terminal-bench-science-0-1 against Anthropic’s reported 52.6%. Epoch AI notes OpenAI funded FrontierMath and has exclusive access to part of it.

Coding is the weak spot in the story. Astra scores 74.1% on DeepSWE v1.1 versus 70.8% for Sol. Meta reported 75.4% for Muse Spark 1.3 at maximum reasoning, and the public leaderboard：https://deepswe.datacurve.ai/ puts Gemini 3.8 Flash and Claude Opus 5 near 74%. On a 113-task benchmark, those gaps are one or two tasks.

Astra is the first model OpenAI has designated as reaching the Critical cybersecurity threshold in its Preparedness Framework. In testing it developed exploits for hardened browsers and operating systems, and found two previously unknown V8 vulnerabilities that OpenAI says it is disclosing to maintainers.

The consequences are practical. Standard access refuses advanced cybersecurity work including exploit discovery. For API developers, a cybersecurity safety check stops a task outright rather than pausing for approval. OpenAI’s Mia Glaese warned that users outside trusted-access programs may hit slowdowns, pauses or blocks, sometimes during unrelated work.

OpenAI reports 100% on ExploitBench：https://exploitbench.ai/, an aggregate capability-coverage score rather than a pass rate, and 42.4% on ExploitGym against 30.3% for Sol, with the usual six-hour time limit removed for both.

Astra costs $10 per million input tokens and $50 per million output, with cached input at $1.00. Requests above 272K input tokens bill at 2x input and 1.5x output for the full request. Batch and Flex run at 50%, Fast mode at 2x. Pro, Business and Enterprise users also get Astra Pro.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.