GPT-6 Astra: ?One Step Closer to AGI

What's happened/new

On 3 Sep 2026, OpenAI released GPT-6 Astra, positioned as its most intelligent and most aligned model. Day one access was limited to Trusted Access and Daybreak enterprise customers, with rollout in coming days to ChatGPT Plus, Pro, Business and Enterprise, the API, Microsoft Azure, and Amazon Bedrock. Astra has a 1.05 million-token context window, 128k max output and a 30 Apr 2026 knowledge cutoff. Pricing is US$10/US$50 per million input/output tokens, 2.5x GPT-5.6 Sol (US$4/US$20) and level with Claude Fable 5.1. Astra was trained on more than 100k GPUs at Stargate, Texas.

Comments

Astra operates software and understands user intent better. Agentic tool use has been the standard across the major models for the best part of the year, so generation-to-execution is not by itself the story. What separates Astra is that it finishes much more tasks before a human intervention and operates tightly within the authorized scope. On OSWorld 2.0, it scores 72.6% (GPT-5.6 Sol: 65.7%), at roughly 40min/task (GPT-5.6 Sol: 75 min/task), about 47% faster. It scores 41.4% on AutomationBench (GPT-5.6 Sol: 18.1%). Astra understands user intent better. OpenAI built an overreach test to evaluate whether a model facing a difficult or impossible task will go beyond its intended scope. GPT-6 did not breach in any test case (GPT-5.6 Sol: 48% of the time).

Astra’s outperformance is harness-dependent. The model delivers outstanding results according to OpenAI. On ARC-AGI-3, Astra scores 99.9% on OpenAI's own continuous conversation harness at roughly US$19k, against 62.7% on ARC’s standard harness at roughly US$26k. OpenAI tuned its own continuous conversation harness and custom compaction for the test, and with that Astra scores higher and cheaper. But OpenAI’s proprietary harness is not publicly available.

Astra’s improved computer-use capabilities underscore incumbent software providers’ value. An agent cannot execute in the abstract: it needs authenticated access to the system where the work is recorded, and the proprietary data that defines what finished work looks like. Both system access and data sit with the incumbent software vendors. Owners of systems of record are paid for holding that access to customer’s data.

Frontier models’ accelerating iterations may unlock more AI application use cases. It takes about three months for Anthropic to make consecutive releases (Fable 5: 9 June; Fable 5.1: 1 September) and two months for OpenAI (GPT-5.6: 9 July; GPT-6: 3 September). Frontier labs’ iteration pace is faster than ever with substantial improvement at each new release.

As model capabilities continue to improve, we see the launch of Astra as a broad-based tailwind for all AI applications, with the potential to unlock a range of new use cases.

Risks: Slower-than-expected enterprise adoption and monetization; model vendors moving further into the application layer.

本报告分析师：

王宪策 Nathan    

SFC HK执业证书编号：BKS605

宋琦

SFC HK执业证书编号：BTF521

王晴

SFC HK执业证书编号：BUL117

免责声明

本公众号涉及的内容仅供参考，并不构成对所述证券的推荐、邀约或邀约邀请。同时亦不构成任何合约的成立，不构成买卖任何证券或交易的建议， 亦不构成投资建议、财务、法律、税务或其他意见。本公众号对于因此信息和内容或者据此进行投资而直接或间接引起的任何损失不承担任何责任亦不作出任何保证。

本公众号中所发布的信息、观点以及数据有可能因发布后的市场走势或其他不可预见之因素而影响其有效性。文中的意见预测和估计反映了我们截至制作本文之日期的评估，本公众号并不承担更新信息、观点以及数据的责任，而同时在对相关信息进行任何更新时亦不作另行通知。

本公众号及其关联机构可能会对于上述信息、观点的标的公司所发行的证券进行交易，还可能为这些公司提供其他服务。本公众号及其关联机构可能购买、持有或出售相同或相关产品。

版权及相关权益归海通国际及相关内容提供方所有。未经海通国际书面许可，任何人不得以任何方式或方法修改、复制、分发或转载官方微信中之任何内容。

《研究报告》免责声明

本公众号不是海通国际研究报告的发布平台。本公众号涉及的研究报告信息由海通国际证券有限公司于香港特别行政区发布，内容从于发布之时从认为可靠的渠道获得的信息或基于海通国际已发布的研究报告编制而成。研究报告信息中的观点及评级的发布取决于报告作者所知悉的各种市场因素及公司内在因素，以及海通国际对其之盈利预测。目标价格是基于一系列的假设和前设条件，投资者需在准确了解相关研究报告中的全部信息的基础下，方可对研究报告所表达的观点形成比较全面的理解。本公众号仅为对相关研究报告部分内容的引用或者复述。受法规技术或其它客观条件所限，无法同时提供观点、评级的依据或盈利预测及目标价格背后的假设等相关信息。因此其内容可能无法完整或准确表达相关研究报告的观点或意见。海通国际并不保证本公众号涉及的研究报告所包含的信息、观点以及数据的准确性及完整性。如欲了解完整观点，应参见海通国际研究网站(www.equities.htisec.com)所载完整报告，或与海通国际销售人员联系。

投资涉及风险，海通国际并没有考虑投资者的具体投资目标，财务状况等的特定需求，任何人不得依赖研究报告包含的信息、观点以及数据的准确性及完整性作出投资决定。