OpenAI 发布 GPT-6 Astra，并首次将其列为 Preparedness Framework 下关键级网络安全模型

OpenAI 已发布 GPT-6 Astra，这是迄今为止最强大的模型。总裁 Greg Brockman 称，它可能已经符合“AGI”（通用人工智能）的标准，或者至少接近实现，这意味着根据 OpenAI 自己的定义，该 AI 系统在大多数经济价值工作上已超越人类。

Astra 曾在德克萨斯州 Stargate 工厂使用超过 100,000 块 GPU 进行预训练。OpenAI 研究员 Aidan Clark 称，这是公司迄今为止最大规模的训练任务。Clark 表示，从 Sol 到 Astra 的飞跃比从早期模型到 Sol 的飞跃能力提升更大，部分原因是以前的 AI 模型在训练监控中发挥了作用。

在 OpenAI 发布的基准测试中，Astra 的评分远高于其前身 GPT-5.6 Sol 和 Anthropic 的 Fable 模型。Astra 在多个学科中均取得最高分：逻辑推理（ARC-AGI-3 上 99.9%，尽管按照其自身测试条件：https://the-decoder.com/openai-claims-gpt-5-6-sol-beats-opus-5-on-arc-agi-3-with-its-latest-api-and-two-additional-settings/）、数学（FrontierMath Tier 4 v2 上 97.6%）、软件工程（DeepSWE v1.1 上 74.1%）、专业知识（GPQA Diamond 上 96%）、工程学（BenchCAD 上 95.9%）以及网络安全（ExploitBench 上 100%）。

BenchCAD 成本：比 Sol 低约 43%，比 Fable 5.1 低约 86%。

Terminal-Bench 4.0 成本：比 Sol 低约 9%，比 Fable 5.1 低约 63%。广告

低成本设置：Terminal-Bench Science 得分 61.1%，成本约低 27%；GPQA Diamond 得分 94.9%，成本约低 37%。Prime 差距从 240 改进到 186，且大型差距边界项 80 多年来首次得到改进。

Fable 5 和 5.1 未包含在 LifeSciBench、GeneBench Pro 和 MedChemBench 中，因为它们拒绝大多数问题。

SRE-Bench 四次尝试内：99.2%，Sol 为 68.7%。Astra 在评估过程中发现了两个此前未知的零日漏洞。

不可能任务范围测试：Sol 超过其授权目标的频率为 48%，Astra 为 0%。Astra 错误陈述自身能力的可能性比 Sol 低 3 倍。

在科学工作中，据报道，该模型在素数间隔的数学结果上有所改进，并在生物学、化学、医学和物理评估中创下了新纪录。OpenAI 还将 Astra 定位为一个能够像人类一样可靠操作计算机的模型，在衡量这种能力的 OSWorld 2.0 上，Astra 的得分为 72.6%，每项任务大约需要 40 分钟，而 Sol 为 65.7%，每项任务大约需要 75 分钟。该公司声称：“你在计算机上能做的任何事情，Astra 都能为你做到。而且快。”：https://x.com/OpenAI/status/2095595741528125780。

除了 Astra，OpenAI 还在更新其 Codex 编码环境。一项新的实验性功能允许模型在长时间的会话中跨多个上下文窗口做笔记，而不是每次都压缩成单一摘要。早期的上下文窗口仍然可以搜索，因此即使之前的信息未被记录在笔记中，Astra 也可以查找先前消息的需求或测试结果。OpenAI 计划在未来几周将其设为默认功能。

通过 API，GPT-6 Astra 在标准模式下的费用为每百万输入令牌 10 美元，每百万输出令牌 50 美元。快速模式承诺速度提高 2.5 倍，但价格也翻倍，使 Astra 比 GPT-5.6 Sol 贵 2.5 倍，并将其置于与 Anthropic 的 Fable 5.1 相同的价格范围：https://the-decoder.com/anthropics-claude-fable-5-1-promises-better-coding-and-research-at-up-to-45-percent-less/。

Brockman 认为，令牌价格正在成为比较模型的糟糕方式，因为 OpenAI 的令牌与竞争对手的不同，甚至在其自家模型系列中也不可比较。他说，重要的是每个完成任务的价格：https://the-decoder.com/frontier-radar-3-how-agentic-ai-is-turning-tokens-into-a-business-metric/，OpenAI 已经在尝试这种定价模式：https://the-decoder.com/openai-starts-charging-some-customers-only-when-its-ai-actually-works/。根据公司称，在 DeepSWE v1.1 上，Astra 的顶级配置将每项任务的 API 预估成本相比 Sol 下降约 57%。

在新闻发布会上，Brockman 承认没有明确界定的 AGI 时刻，他表示团队最初认为在 OpenAI 成立时会有一个显而易见的门槛，所有人都会认出。但事情并非如此发展，过渡比预期更加渐进，这也是 OpenAI 去年春天所阐述的立场：https://the-decoder.com/openai-shifts-away-from-sudden-agi-breakthrough-theory/。Brockman 在发布会结束时说：https://venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra，“欢迎来到 AGI 时代。”OpenAI CEO Sam Altman 此前曾表示，如果按他的定义，他预计到今年年底会有一个他会称之为 AGI 的模型：https://the-decoder.com/sam-altman-says-openai-will-have-agi-by-the-end-of-2026-if-you-accept-his-definition/。

Astra 是 OpenAI 根据其《准备框架》首次归类为“关键”的模型：https://the-decoder.com/openai-says-its-pacing-model-development-as-ai-cybersecurity-risks-grow-too-dangerous/。这意味着在提供正确工具和访问权限的情况下，该模型能够发现以前未知的漏洞，并在无需人工逐步指导的情况下在防御良好的系统中构建利用链。OpenAI 还承认，Astra 的书面推理比 GPT-5.6 Sol 的更难以监控：https://the-decoder.com/openai-calls-astra-its-most-dangerous-model-yet-watching-what-it-does-is-only-getting-harder/。

在 ExploitBench 上：https://the-decoder.com/openai-calls-astra-its-most-dangerous-model-yet-watching-what-it-does-is-only-getting-harder/，这是一种衡量模型发现和利用真实软件漏洞能力的基准测试，Astra 获得了满分 100%。OpenAI 表示，该模型在评估过程中发现了两个以前未知的漏洞，公司随后将其报告给受影响的厂商。人类专家确认，Astra 可以在多个软件类别中识别新的零日漏洞，包括浏览器和操作系统。

最先进的网络安全能力目前仅限于 Daybreak Blue 项目中的受信任防御者使用。OpenAI 之前曾推迟 Astra 的发布：https://the-decoder.com/openai-reportedly-slows-research-after-its-own-models-secretly-coordinated-hacks-for-weeks-undetected/ 以进行更多的安全测试。这些双用途能力有利有弊：一个能够自主发现漏洞的代理可以帮助防御者轻松修补漏洞，也同样可以帮助攻击者利用漏洞。

保持对 AI 的关注。内容清晰、有用，无废话。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

解码器：https://the-decoder.com/

OpenAI has shipped GPT-6 Astra, its most capable model to date. President Greg Brockman says it might already qualify as "AGI" or is at least within reach, meaning an AI system that outperforms humans at most economically valuable work by OpenAI's own definition.

Astra was pretrained on more than 100,000 GPUs at the Stargate facility in Texas. OpenAI researcher Aidan Clark called it the company's largest training run ever. The jump from Sol to Astra represents a bigger capability gain than the jump to Sol from earlier models, Clark said, in part because previous AI models played a role in monitoring training. Ad

In benchmarks OpenAI published, Astra scores well above its predecessor GPT-5.6 Sol and Anthropic's Fable models. Astra hits top marks across a range of disciplines: logical reasoning (99.9 percent on ARC-AGI-3, though under its own test conditions：https://the-decoder.com/openai-claims-gpt-5-6-sol-beats-opus-5-on-arc-agi-3-with-its-latest-api-and-two-additional-settings/), math (97.6 percent on FrontierMath Tier 4 v2), software engineering (74.1 percent on DeepSWE v1.1), expert knowledge (96 percent on GPQA Diamond), engineering (95.9 percent on BenchCAD), and cybersecurity (100 percent on ExploitBench). Ad

BenchCAD cost: ~43% below Sol, ~86% below Fable 5.1.

Terminal-Bench 4.0 cost: ~9% below Sol, ~63% below Fable 5.1. Ad

Lower-cost settings: Terminal-Bench Science 61.1% at ~27% lower cost; GPQA Diamond 94.9% at ~37% lower cost. Prime gaps improved from 240 to 186, and a large-gap bound term improved for the first time in over 80 years.

Fable 5 and 5.1 are not included in LifeSciBench, GeneBench Pro, and MedChemBench because they reject most questions. Ad

SRE-Bench within four attempts: 99.2% versus 68.7% for Sol. Astra found two previously unknown zero-days during evaluation. Ad

Impossible-task scope test: Sol exceeded its authorized target 48% of the time, Astra 0%. Astra is 3x less likely to misstate its own capabilities.

In scientific work, the model reportedly improved a mathematical result on prime gaps and set new records in biology, chemistry, medicine, and physics evaluations. OpenAI also positions Astra as a model that can reliably operate a computer the way a human would, and on OSWorld 2.0, which measures that ability, Astra scored 72.6 percent at about 40 minutes per task compared to Sol's 65.7 percent at roughly 75 minutes. "Anything you can do on a computer, Astra can do for you. Fast," the company claims：https://x.com/OpenAI/status/2095595741528125780.

Alongside Astra, OpenAI is also updating its Codex coding environment. A new experimental feature lets the model take notes across multiple context windows during long sessions, rather than compressing everything into a single summary each time. Earlier context windows remain searchable, so Astra can look up requirements or test results from previous messages even if they weren't captured in the notes. OpenAI plans to make this the default in the coming weeks.

GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens in standard mode through the API. Fast mode, which promises 2.5x speed, doubles the price, making Astra 2.5x more expensive than GPT-5.6 Sol and putting it in the same price range as Anthropic's Fable 5.1：https://the-decoder.com/anthropics-claude-fable-5-1-promises-better-coding-and-research-at-up-to-45-percent-less/.

Brockman argued that token prices are becoming a poor way to compare models, since OpenAI's tokens aren't the same as a competitor's and aren't even comparable across its own model families. What matters, he said, is the price per completed task：https://the-decoder.com/frontier-radar-3-how-agentic-ai-is-turning-tokens-into-a-business-metric/, and OpenAI is already experimenting with that pricing model：https://the-decoder.com/openai-starts-charging-some-customers-only-when-its-ai-actually-works/. On DeepSWE v1.1, Astra's top configuration cuts estimated API costs per task by about 57 percent compared to Sol, according to the company.

During the press briefing, Brockman acknowledged there's no clearly defined AGI moment, saying the team originally thought there would be an obvious threshold everyone would recognize when OpenAI was founded. That's not how it played out, and the transition has been more gradual than expected, a position OpenAI laid out last spring：https://the-decoder.com/openai-shifts-away-from-sudden-agi-breakthrough-theory/. Brockman closed the briefing by saying：https://venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra, "Welcome to the AGI era." OpenAI CEO Sam Altman had previously said he expects a model he'd call AGI by the end of the year：https://the-decoder.com/sam-altman-says-openai-will-have-agi-by-the-end-of-2026-if-you-accept-his-definition/.

Astra is the first model that OpenAI classifies as "critical" under its Preparedness Framework：https://the-decoder.com/openai-says-its-pacing-model-development-as-ai-cybersecurity-risks-grow-too-dangerous/. That means the model can find previously unknown vulnerabilities and build exploit chains across well-defended systems when given the right tools and access, all without a human guiding it step by step. OpenAI also admits that Astra's written reasoning is harder to monitor：https://the-decoder.com/openai-calls-astra-its-most-dangerous-model-yet-watching-what-it-does-is-only-getting-harder/ than GPT-5.6 Sol's.

On ExploitBench：https://the-decoder.com/openai-calls-astra-its-most-dangerous-model-yet-watching-what-it-does-is-only-getting-harder/, a benchmark that measures a model's ability to discover and exploit real software vulnerabilities, Astra scores a perfect 100 percent. OpenAI says the model discovered two previously unknown vulnerabilities during evaluation, which the company then reported to the affected vendors. Human experts confirmed that Astra can identify novel zero-day vulnerabilities across several software categories, including browsers and operating systems.

The most advanced cybersecurity capabilities are restricted for now to trusted defenders in the Daybreak Blue program. OpenAI had previously delayed Astra's release：https://the-decoder.com/openai-reportedly-slows-research-after-its-own-models-secretly-coordinated-hacks-for-weeks-undetected/ to run more safety testing. These dual-use capabilities cut both ways: an agent that can autonomously find a vulnerability helps a defender patch it just as easily as it helps an attacker exploit it.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/