GPT-6 Astra 究竟带来了什么：不是更聪明，是第一次能真正帮你干活

OpenAI 在 9 月 3 日凌晨发了一条推特，简短到令人不安：

“This is GPT-6 Astra. Anything you can do on a computer, Astra can do for you. Fast.”

125,164,782 次浏览，328,977 个点赞。

这条推文没说 benchmark，没提参数量，只说了一件事：Astra 可以替你操作电脑，而且很快。

这是 OpenAI 有史以来最直接的产品宣言。它描述的不是一个更聪明的问答系统，而是一个可以执行的 Agent。

---

背景：为什么叫 Astra

Astra 是拉丁语里”星星”的意思。OpenAI 把 GPT-6 命名为 Astra，暗示了一个定位：这不只是上一代的升级版本，而是一个新的起点。

从 GPT-4 到 GPT-5.x，每一代的主要叙事是”更聪明、更准确、更少幻觉”。GPT-6 Astra 的主要叙事变了：从「回答问题」到「完成任务」。

---

核心突破：Computer Use 的质变时刻

发布后 48 小时，推特上讨论最多的不是 benchmark，而是 Computer Use 的真实体验。

开发者测试验收场景（来自 @dotey，18.6万次浏览）：

“目前用下来 GPT-6 Astra 最让我惊艳的还是 Computer Use 的能力，它不像之前那样要稍微等一会才进行各种操作，现在它能很快很精准的帮我测试 App，在旁边看着它点击真的是一种享受。这其实带来一个最大的提升就是让 Agent 从开发到验收形成了完整的闭环。”

10 分钟搭起 7 个 bot 的运营自动化（来自 @知识猫AI实验室）：

“直接丢给 Astra，然后没过 10 分钟它就帮我把数字军团搭建了起来，7 个 bot、2 个群、5 个 skill、还有定时任务。这要是我自己创建不得用大半天时间啊。”

Blender 建模 → Unreal Engine 5 可走动场景（官方演示）：完整房屋建模，含泳池、花园、客厅、厨房，整个流程全自动。

为什么这次 Computer Use 不一样？官方给出的数据：

• OSWorld 2.0：72.6%，而 GPT-5.6 Sol 是 65.7%
• 速度：比 GPT-5.6 Sol 完成同等任务快 47%（75分钟 → 40分钟）
• Mind2Web：端到端任务完成速度 1.9 倍于 GPT-5.6 Sol

速度和准确率同时提升，是 Computer Use 从”实验性功能”变成”可用工具”的关键门槛。之前大家不是不知道 Computer Use 的潜力，而是准确率低、速度慢，真正用起来反而比自己手动做慢。Astra 突破了这个临界点。

---

有价值的用法 vs 浪费 Token 的用法

发布两天后，推特上开始出现一个有意思的讨论：哪些 Computer Use 场景真正有价值，哪些是在浪费钱。

@Jesse Lau 直说：

“用 GPT-6 Astra 的 Computer Use 去调用 PS 画一张现有的图完全属于浪费 tokens，不能测试出其性能的。”

同一个账号后来又发：

“哈哈，我找到 GPT-6 Astra 的正确用法了。”

这个对比值得认真对待。Astra 的 Computer Use 在以下场景表现最佳：

高价值场景

• 复杂多步骤工作流：需要在多个软件之间跳转、输入、读取结果的任务
• 测试和验收：让 Astra 自己测试它写的代码，开发到交付形成闭环
• 重复性专业操作：CRM 更新、日历整理、表单填写——低价值但耗时的知识工作
• 专业软件操作：KiCad（PCB 设计）、Blender、基因数据分析软件——有学习曲线的专业工具
• UI 生成（@MSchwaibold，85.5万次浏览：“GPT-6 Astra is really good at generating UI”）

低价值/浪费场景

• 让它用 Photoshop 画一张你已有的图（用图像生成更快更便宜）
• 单步骤、可以直接用 API 完成的任务（不需要图形界面的事不要用 Computer Use）
• 简单搜索和问答（普通对话模式成本更低）

---

专业工作的五个具体跃升

1. 编码：开发-测试闭环

Terminal-Bench 4.0：57.9%（GPT-5.6 Sol：37.3%）

更重要的是 Codex 的上下文持久化更新：Astra 可以在多个上下文窗口之间保持笔记，早期上下文保持可搜索。这直接解决了长会话里调试信息丢失的问题——每一次 compaction 不再是信息黑洞。

2. CAD 和工程设计

BenchCAD 95.9%（前代 83.3%，提升约 15 个百分点）

KiCad PCB 设计的官方演示：从电路原理图到可制造 PCB，包括元件布置和铜线布线，全程自动。PCB 布局是电子设计流程里最耗时的手工活之一，这个 demo 如果能在真实工作流里复现，影响不小。

3. 文档和演示

AutomationBench 41.4%（前代 18.1%，超过翻倍）

Astra 可以识别你的模板风格并在整份演示文稿里保持一致，而不是每隔几张幻灯片就开始偏离。官方演示里它用几张 OpenAI 模板幻灯片制作了关于虚构模型 GPT-Gaia 的完整演讲——格式、语气、布局全部贯穿始终。

4. 科学研究

Terminal-Bench Science 0.1：64.6%（Claude Fable 5.1：52.6%）

Astra 已经帮助解决了素数间隔的两个长期开放数学问题（今天继续披露了两个新结果）。对科研工作者来说更直接的是：它可以在专业软件里直接操作，检查测序质量、可视化基因变异——而不只是”提供建议”。

5. 网站和游戏

通过 ChatGPT 里的 Sites 功能，Astra 可以直接生成、托管并分享网站、Web App 和游戏——从 prompt 到上线，全程在 ChatGPT 里完成。

---

安全性：第一个达到 Critical 阈值的模型

这是 GPT-6 Astra 里另一个不能略过的话题。

Astra 是 OpenAI 第一个在网络安全能力上达到 Critical 阈值 的模型（根据 OpenAI 的 Preparedness Framework）。

具体数字：

• ExploitBench：100%（GPT-5.6 Sol：78.5%）
• SRE-Bench（二进制逆向工程）：88% 单次成功，99.2% 四次内成功
• 在评估中发现并使用了两个此前未知的零日漏洞，已向维护者披露

OpenAI 的处理方式：上线版本拒绝更高级的网络安全任务（生成 PoC 利用代码），计划通过 OpenAI Daybreak 项目在未来几周逐步开放给防御用途。

对齐方面的数字同样显著：

• 计算机使用安全基准：2.4%（GPT-5.6 Sol：22%）——越低越安全
• 在无法完成任务的情况下，Astra 越权行动的比例：0%（GPT-5.6 Sol：48%）

---

价格和可用性

API 定价：$10/百万输入 token，$50/百万输出 token。Fast 模式 2 倍速、2 倍价格。

ChatGPT 用户：Plus、Pro、Business、Enterprise 用户将在接下来几天内陆续获得访问权限，包含在现有订阅额度内，可购买额外额度。

平台：OpenAI API（gpt-6-astra）、Microsoft Azure、Amazon Bedrock。

Enterprise 版本默认关闭，管理员需要手动启用。

---

回到那条推文

“Anything you can do on a computer, Astra can do for you.”

这句话有一个隐含的前提没说出来：需要你告诉它做什么，以及验证它做对了。

但这已经是一个质的变化。过去几年，AI 助手在”更好地回答问题”这个方向上持续迭代。GPT-6 Astra 切换了轨道：它的目标不是给你更好的答案，而是直接替你完成工作。

这对效率工具的竞争格局意味着什么，对知识工作者的日常流程意味着什么，现在还早。但两天的推特观察已经清楚地显示：那些愿意把复杂重复性工作流交给 Astra 的人，正在获得真实的时间回报。

---

相关链接

• 官方博客：openai.com/index/gpt-6-astra/
• 安全更新：Path to Astra
• 系统卡：deploymentsafety.openai.com/gpt-6-astra

OpenAI posted a tweet at 2:32am on September 3rd. It was short enough to be unsettling:

“This is GPT-6 Astra. Anything you can do on a computer, Astra can do for you. Fast.”

125,164,782 views. 328,977 likes.

No benchmarks. No parameter counts. Just one claim: Astra can operate your computer on your behalf, and it’s fast.

This is the most direct product statement OpenAI has ever made. It describes not a smarter question-answering system, but an executable agent.

---

Background: Why “Astra”

Astra is Latin for “stars.” OpenAI naming GPT-6 as Astra signals a positioning: not just an incremental upgrade, but a new starting point.

From GPT-4 through GPT-5.x, the main narrative was “smarter, more accurate, less hallucination.” GPT-6 Astra changes the narrative: from “answer questions” to “complete tasks.”

---

The Core Breakthrough: Computer Use Crosses the Threshold

In the 48 hours after launch, Twitter’s most active discussion wasn’t about benchmarks — it was about real Computer Use experiences.

Developer test-and-validation use case (@dotey, 186K views):

“What impresses me most about GPT-6 Astra is Computer Use. Unlike before where you’d wait a bit for operations, now it can help me test apps very quickly and precisely. Watching it click through things is actually enjoyable. The biggest improvement is that it closes the loop from development to acceptance testing.”

10 minutes to build 7 bots (@知识猫AI实验室):

“Just handed the task to Astra, and in under 10 minutes it built the whole digital fleet — 7 bots, 2 groups, 5 skills, plus scheduled tasks. Something that would have taken me half a day.”

Blender to Unreal Engine 5 walkthrough (official demo): Complete house modeling — pool, garden, living room, kitchen — fully automated.

Why is Computer Use different this time? Official numbers:

• OSWorld 2.0: 72.6% vs. GPT-5.6 Sol’s 65.7%
• Speed: 47% faster per task than GPT-5.6 Sol (75 min → 40 min)
• Mind2Web: end-to-end task completion 1.9× faster than GPT-5.6 Sol

Speed and accuracy improving together is the threshold that turns Computer Use from an “experimental feature” into a “working tool.” The potential was always understood; low accuracy and slow speed meant it was often slower than doing things yourself. Astra crossed that threshold.

---

High-Value Uses vs. Token Waste

Within 48 hours, Twitter developed an interesting meta-discussion: which Computer Use scenarios are genuinely valuable, and which are wasting money.

@Jesse Lau said directly:

“Using GPT-6 Astra’s Computer Use to have it draw an existing image in Photoshop is pure token waste — it doesn’t test its actual capabilities.”

The same account later posted: “Ha, I’ve found the right way to use GPT-6 Astra.”

This contrast is worth taking seriously. Astra’s Computer Use performs best in:

High-value scenarios

• Complex multi-step workflows: tasks requiring navigation across multiple applications
• Testing and validation: let Astra test the code it wrote — close the development-to-delivery loop
• Repetitive professional operations: CRM updates, calendar organization, form filling — low-value but time-consuming knowledge work
• Specialized professional software: KiCad (PCB design), Blender, genetic analysis software — tools with steep learning curves
• UI generation (@MSchwaibold, 855K views: “GPT-6 Astra is really good at generating UI”)

Low-value / wasteful scenarios

• Having it recreate an existing image in Photoshop (image generation is faster and cheaper)
• Single-step tasks achievable via direct API (don’t use Computer Use for things that don’t need a GUI)
• Simple search and Q&A (conversational mode is much cheaper)

---

Five Concrete Jumps in Professional Work

1. Coding: Development-Test Loop Closed

Terminal-Bench 4.0: 57.9% vs. GPT-5.6 Sol’s 37.3%

More important: Codex now keeps notes across context windows, with earlier contexts remaining searchable. This directly solves the problem of debugging information disappearing during long sessions — each compaction is no longer an information black hole.

2. CAD and Engineering Design

BenchCAD: 95.9% (up from 83.3%, roughly 15 percentage points)

The official KiCad demo: from electronic schematic to manufacturable PCB, including component placement and copper routing, fully automated. PCB layout is one of the most time-consuming manual steps in electronics design.

3. Documents and Presentations

AutomationBench: 41.4% (up from 18.1%, more than doubled)

Astra can recognize your template style and maintain it throughout a full presentation. The official demo produced a complete deck for fictional model “GPT-Gaia” using a few template slides, keeping format, tone, and layout consistent throughout.

4. Scientific Research

Terminal-Bench Science 0.1: 64.6% (Claude Fable 5.1: 52.6%)

Astra has already helped solve two long-standing open problems on prime number gaps, with two more disclosed today. More practically: it can operate specialized research software directly — inspect sequencing quality, visualize genetic variation — rather than just providing advice.

5. Websites and Games

Via Sites in ChatGPT, Astra can generate, host, and share websites, web apps, and games directly from a prompt — from idea to live URL, entirely within ChatGPT.

---

Safety: The First Model to Hit the Critical Threshold

This isn’t something to skip.

Astra is the first OpenAI model to reach the Critical threshold in cybersecurity under OpenAI’s Preparedness Framework.

Numbers:

• ExploitBench: 100% (GPT-5.6 Sol: 78.5%)
• SRE-Bench (binary reverse engineering): 88% single-attempt, 99.2% within four attempts
• During evaluation, discovered and used two previously unknown zero-day vulnerabilities, now disclosed to maintainers

OpenAI’s approach: the launch version refuses advanced cybersecurity tasks (generating proof-of-concept exploits), with planned expansion through OpenAI Daybreak for defensive use over coming weeks.

Alignment numbers are equally notable:

• Computer use safety benchmark: 2.4% (GPT-5.6 Sol: 22.0%) — lower is safer
• Unauthorized scope expansion with impossible tasks: 0% (GPT-5.6 Sol: 48%)

---

Pricing and Availability

API pricing: $10/M input tokens, $50/M output tokens. Fast mode: 2× speed, 2× price.

ChatGPT: Plus, Pro, Business, Enterprise users rolling out over coming days, included in existing subscription allowances, additional credits purchasable.

Platforms: OpenAI API (gpt-6-astra), Microsoft Azure, Amazon Bedrock.

Enterprise defaults to off; admins must enable.

---

Back to That Tweet

“Anything you can do on a computer, Astra can do for you.”

There’s an implicit premise left unsaid: you still need to tell it what to do, and verify it did it right.

But that’s already a qualitative change. For the past few years, AI assistants iterated on “answering questions better.” GPT-6 Astra shifts tracks: its goal isn’t a better answer — it’s to complete the work directly.

What this means for the productivity tool landscape, and for knowledge workers’ daily workflows, is still early. But two days of Twitter observation have made one thing clear: people who are willing to hand complex, repetitive workflows to Astra are getting real time back.

---

Links

• Official blog: openai.com/index/gpt-6-astra/
• Safety update: Path to Astra
• System card: deploymentsafety.openai.com/gpt-6-astra