
IT之家 9 月 22 日消息,SpaceXAI 今日宣布推出新一代 Grok 4.7,主打编程和知识工作情形。
官方称其使用全新更大规模基础模型,并借助更长时间的强化学习训练提高复杂任务处理能力,是该公司当下最强的编码和知识处理模型 —— 速度是同类模型的两倍,售价却只有一半。
如前所述,Grok 4.7 针对编程和知识工作开展了优化。与 Grok 4.6 相比,新模型可以在困难任务上持续更长时间,并加强了对自身输出结果的检查能力。
在 CursorBench 4.0 测试中,Grok 4.7 面向长时间运行的编程任务,在售价与表现维度处于同类模型的前列。官方同时表示,该模型的运行速度和定价与 Grok 4.6 相同。
Grok 4.7 使用了比 Grok 4.6 更大的基座模型,并接受了更长时间的强化学习训练。训练数据更加侧重需要数小时完成的复杂任务,同时针对模型的自我验证和长上下文管理能力开展了优化。
SpaceXAI 还针对 Grok Bot harness 对 Grok 4.7 开展了原生训练,使其在对话以及一般知识工作任务中的表现得到提高。另外,该模型在文档和演示文稿生成维度也有所改进。
在 GDPval 和 AA Briefcase 两项基准测试中,Grok 4.7 均较 Grok 4.6 有所提高。这些测试要求 AI 完成律师、护士、金融分析师等专业人士日常工作中的相关任务。官方称,Grok 4.7 在两项测试中的表现已达到与其他前沿模型相近的水平。
安全维度,SpaceXAI 为 Grok 4.7 使用了全新的安全防护体系。官方表示,该模型在拒答和抵抗越狱攻击维度达到了公司当下测试中的最高水平。
在网络安全和生物领域等具有双重用途的情形中,Grok 4.7 试图兼顾合法任务的可用性与危险任务的安全拒答。在 LatchBio 生物安全基准测试中,该模型取得 62.4% 的成绩。
针对网络安全情形,SpaceXAI 表示,Grok 4.7 在 HackerBench v0.3 测试中仅允许 3.3% 的高风险双重用途提示借助,同时较少阻止合法的安全工作。另外,公司已经开始向部分网络安全合作伙伴给出邀请制访问权限,用于红队能力和防御研究。
当下,Grok 4.7 已借助 Cursor 和 Grok Build 给出,同时开放 Grok API,并兼容第三方编程工具、模型路由服务及云平台。
售价维度,Grok 4.7 每百万词元输入起价为 2 美元(IT之家注:现汇率约合 13.4 元人民币),输出起价为 6 美元(现汇率约合 40.2 元人民币)。SpaceXAI 同时还给出了高速版本,输出速度翻倍,但售价也翻倍。
本文来源:https://www.ithome.com/1/005/494.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Home > Intelligent Age > Artificial Intelligence
Thanks to IT Home readers HuTao29, Remote Lead CDM, Blackn, Xiaoxing_14, Wu Yan-zu South China, Domado, Soft Media User 1053564, and Yu Gong Qima for the tips!
IT Home, September 22 — SpaceXAI today announced the release of the next-generation Grok 4.7, which focuses on programming and knowledge work.
According to the official statement, it uses a new, larger base model and has undergone longer reinforcement learning training to improve its handling of complex tasks. It is the company’s strongest coding and knowledge processing model to date — twice as fast as its peers with half the price.
As mentioned earlier, Grok 4.7 is optimized for programming and knowledge work. Compared to Grok 4.6, the new model can run longer on difficult tasks and has enhanced its ability to check its own output.
In the CursorBench 4.0 test, Grok 4.7 ranks at the top in both price and performance for long-running programming tasks. The official representative also stated that the model’s running speed and pricing are consistent with those of Grok 4.6.
Grok 4.7 utilizes a larger base model than Grok 4.6 and has been subjected to longer reinforcement learning. The training data is more focused on complex tasks that require hours to complete, while also optimizing the model’s self-validation and long-context management capabilities.
SpaceXAI has also nativized the training of Grok 4.7 for the Grok Bot harness, improving its performance in conversations and general knowledge work tasks. Additionally, the model has seen improvements in document and presentation generation.
In both the GDPval and AA Briefcase benchmark tests, Grok 4.7 has outperformed Grok 4.6. These tests require AI to complete tasks relevant to the daily work of professionals like lawyers, nurses, and financial analysts. The official report indicates that Grok 4.7’s performance in both tests has reached the level of other leading models.
Regarding safety, SpaceXAI has implemented a new guardrails system for Grok 4.7. The company claims that the model achieves the highest levels of refusal and resistance to jailbreaking attacks in their current testing.
In dual-use scenarios such as cybersecurity and biology, Grok 4.7 attempts to balance the usability of legitimate tasks with the secure refusal of dangerous ones. In the LatchBio biosecurity benchmark, the model scored 62.4%.
For cybersecurity scenarios, SpaceXAI reports that Grok 4.7 allows only 3.3% of high-risk, dual-use prompts to pass through the HackerBench v0.3 test, while minimizing interference with legitimate security work. The company is also beginning to provide invite-only access to selected cybersecurity partners for red-teaming and defense research.
Currently, Grok 4.7 is available via Cursor and Grok Build, and its API is open for integration with third-party programming tools, model routing services, and cloud platforms.
The pricing for Grok 4.7 starts at $2 per million input tokens (approximately 13.4 RMB) and $6 for output (approximately 40.2 RMB). SpaceXAI also offers a high-speed version with double the output rate but double the price.





