SpaceXAI 发布 Grok 4.5，Cursor 收购后首个 Opus 级模型

随着 GPT 5.6 确认将于明天发布，今天几乎是人们会对 GPT 5.5 级别模型发布感到兴奋的最后一天，而 SpaceXAI 恰好做了这件事：

[我们已与 SpaceXAI 合作训练 Grok 4.5。这是我们迄今最强大的模型，也是首个为软件工程之外构建的模型。5:57 PM · Jul 8, 2026 · 324 万查看

605 回复 · 1.26K 转推 · 14.4K 喜欢](https://x.com/cursor_ai/status/2074915744999969059)

新发布的 Grok 4.5 与 Composer 系列（1.5T）属于不同的重量级，尽管评测扎实，其性能仍与当前主力 Opus 和 GPT 相当，不过 OpenAI 的评估团队指出强大的 SWE-Bench Pro 现已饱和/存在根本性缺陷，留下的可能是一小部分后继基准，包括 FrontierCode。

关于训练与数据披露，我们掌握的信息仅此而已。

2026年7月7日至8日 AI 新闻。我们查阅了12个subreddit、544个 Twitter 账户 且未涉及其他 Discord。AINews 网站 可搜索往期所有内容。提醒：AINews 现为 Latent Space 的一个栏目。你可以选择加入/退出邮件推送频率！

头条：Grok 4.5 发布

xAI/“SpaceXAI” 公开发布 Grok 4.5，作为一款聚焦编程和智能体的新型前沿模型，定位以单位成本能力取胜，而非追求绝对基准霸主地位。

• 伊隆·马斯克最早表示 Grok 4.5 将在“明天”公开，基于良好的 beta 反馈，称其为“Opus 级”，但速度更快、token 效率更高、成本更低 @elonmusk
• 马斯克随后将 Grok 4.5 内部定性为“与 Opus 4.7 大致相当，但快得多”，强调对特斯拉和 SpaceX 工程师的实用性胜过追逐基准分 @elonmusk
• 官方发布来自 xAI 账户，将 Grok 4.5 描述为“我们首个专为编程和智能体训练的模型”，与 Cursor 合作训练，提供“领先速度与成本效率的前沿智能” @SpaceXAI
• Cursor 称其与 xAI 合作训练了 Grok 4.5，称之为“我们迄今最强大的模型”，并强调这是“首个我们为软件工程之外构建的模型” @cursor_ai
• Cursor 还宣布产品内可用，并“首周用量翻倍” @cursor_ai
• Cursor 澄清“Grok 4.5 与 Composer 属于两种不同的模型重量级”，Composer 2.5 将继续保持可用，后续该较小级别还会推出新模型 @cursor_ai
• 早期生态系统支持立刻出现：Grok 4.5 在 Grok Build/API/Cursor 中可用 @milichab，宣布对 Hermes Agent 提供首日支持 @Teknium，稍后确认在 Hermes Agent/Portal/OpenRouter/Grok 订阅中上线 @Teknium
• 马斯克称上下文窗口可能从 500K 恢复至 1M，“下周前” @elonmusk

官方传递的信息并非“最佳综合模型”，而是接近 Opus 品质，同时在经济效益和速度上明显更优：

• “Opus 级模型，但速度更快，token 效率更高，成本更低” @elonmusk
• “首个专为编程和智能体训练的模型” @SpaceXAI
• “领先速度与成本效率的前沿智能” @SpaceXAI
• “我们迄今最强大的模型”以及“首个我们为软件工程之外构建的模型” @cursor_ai

这一定位很重要：xAI 明确瞄准了近期由 Anthropic/OpenAI/Cursor 式工具使用系统主导的编程智能体工作流市场，而不只是通用聊天。

浮出水面的具体数字：

• 官方定价：每百万输入 token 2 美元，每百万输出 token 6 美元 @scaling01
• Artificial Analysis 重复了同一价格点并补充：
  • 缓存命中折扣 75%，降至每百万 token 0.5 美元
  • 超过 200k token 的长输入价格翻倍
  • 500k 上下文窗口，低于 Grok 4.3 的 1M
  • 保留视觉输入
  • 保留可配置推理 @ArtificialAnlys
• 马斯克随后表示上下文窗口可能很快升级回 1M @elonmusk

用户引用的相对定价对比：

通过第三方对马斯克披露的报道浮出了一个重要规格：

这是一次显著的跃升，很可能解释了为何众多观察者将 4.5 解读为 xAI 首次进入真正旗舰级编程智能体梯队，而非迭代更新。

Artificial Analysis 在推文集中提供了最实质性的外部评测。

关键结果：

• 在 Artificial Analysis 智能指数上排名第4，得分 54，仅落后于 Fable 5、GPT-5.5 和 Opus 4.8 @ArtificialAnlys
• 同一指数上较 Grok 4.3 提升 +16 分 @ArtificialAnlys
• GDPval-AA v2 Elo 1543，同样排名第4，位居 Anthropic 最新 Claude 版之后 @ArtificialAnlys
• τ³-Banking 上取得最高分 33%，高于 GPT-5.5（xhigh）的 31% @ArtificialAnlys
• 在 Grok Build 中 Artificial Analysis 编程智能体指数得分 76，“与 Codex 中的 GPT-5.5 相当”，低于 Claude Code 中的 Fable 5 @ArtificialAnlys
• 每项智能指数任务成本：0.31 美元 @ArtificialAnlys
• 每项 GDPval 任务成本：0.49 美元 @ArtificialAnlys
• 每项编程智能体指数任务成本：2.59 美元 @ArtificialAnlys
• 每项智能指数任务平均输出 token：约 1.4 万，比 Opus 4.8 低 60% 以上 @ArtificialAnlys
• 每项编程智能体指数任务平均总 token：190 万，对比 Claude Code 中 Fable 5 的 720 万和 Codex 中 GPT-5.5 的 620 万 @ArtificialAnlys

Artificial Analysis 的解读很明确：Grok 4.5 的能力接近前沿，但在效率方面异常强大，使其处于成本/性能的帕累托前沿。

马斯克明确转发了 Artificial Analysis 的评估 @elonmusk。

As GPT 5.6 is confirmed to launch tomorrow, today is pretty much the last day anyone will be excited about a GPT 5.5 equivalent model launch, and that is exactly what SpaceXAI did:

[We've partnered with SpaceXAI to train Grok 4.5. It’s our most powerful model yet and the first we've built for more than software engineering. 5:57 PM · Jul 8, 2026 · 3.24M Views

605 Replies · 1.26K Reposts · 14.4K Likes](https://x.com/cursor_ai/status/2074915744999969059)

The new Grok 4.5 is a different weight class than the Composer series (1.5T) and despite the solid evals still performs very comparably to the current workhorse Opus and GPTs, although per OpenAI’s evals team even the mighty SWE-Bench Pro is now saturated/terminally flawed - leaving presumably a small list of successors including FrontierCode.

As for training and data disclosures, this is all the information we have.

AI News for 7/07/2026-7/08/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Top Story: Grok 4.5 release

xAI/“SpaceXAI” publicly launched Grok 4.5 as a new coding-and-agents-focused frontier model, positioned on capability-per-dollar rather than absolute benchmark supremacy.

• Elon Musk first said Grok 4.5 would be made public “tomorrow” based on strong beta feedback, calling it “Opus-class,” but faster, more token-efficient, and lower cost @elonmusk.

• Musk later framed Grok 4.5 internally as “roughly comparable to Opus 4.7, but much faster,” emphasizing usefulness to Tesla and SpaceX engineers over benchmark chasing @elonmusk.

• The official launch came from xAI’s account, describing Grok 4.5 as “our first model trained specifically for coding and agents,” trained with Cursor, and offering “frontier intelligence at leading speeds and cost efficiency” @SpaceXAI.

• Cursor said it partnered with xAI to train Grok 4.5, called it “our most powerful model yet,” and stressed that it was “the first we’ve built for more than software engineering” @cursor_ai.

• Cursor also announced in-product availability with “double usage for the first week” @cursor_ai.

• Cursor clarified that “Grok 4.5 and Composer are two different model weight classes,” and that Composer 2.5 would remain available with future models in that smaller class @cursor_ai.

• Early ecosystem support appeared immediately: Grok 4.5 became available in Grok Build/API/Cursor @milichab, day-0 support was announced for Hermes Agent @Teknium, and later live availability in Hermes Agent/Portal/OpenRouter/Grok subscriptions was confirmed @Teknium.

• Musk said the context window would likely move from 500k back to 1M “by next week” @elonmusk.

Officially, xAI’s message was not “best overall model,” but near-Opus quality with materially better economics and speed:

• “Opus-class model, but faster, more token-efficient and lower cost” @elonmusk

• “First model trained specifically for coding and agents” @SpaceXAI

• “Frontier intelligence at leading speeds and cost efficiency” @SpaceXAI

• “Most powerful model yet” and “first we’ve built for more than software engineering” @cursor_ai

This framing matters: xAI is explicitly targeting the coding-agent workflow market that has recently been dominated by Anthropic/OpenAI/Cursor-style tool-using systems, not just general chat.

The concrete numbers that surfaced:

• Official pricing: $2 / 1M input tokens, $6 / 1M output tokens @scaling01

• Artificial Analysis repeated the same price point and added:

  • cache hits discounted by 75% to $0.5 / 1M tokens

  • long inputs over 200k tokens cost double

  • 500k context window , down from Grok 4.3’s 1M

  • vision input retained

  • configurable reasoning retained @ArtificialAnlys

• Musk later said the context window would probably upgrade back to 1M soon @elonmusk.

Relative pricing comparisons cited by users:

One important spec surfaced via third-party reporting of Musk’s disclosure:

That is a notable jump, and likely central to why multiple observers interpreted 4.5 as xAI’s first entry into the true flagship coding-agent tier rather than an iterative refresh.

Artificial Analysis provided the most substantive external evaluation in the tweet set.

Key results:

• #4 on Artificial Analysis Intelligence Index , score 54 , behind only Fable 5, GPT-5.5, and Opus 4.8 @ArtificialAnlys

• +16 points vs Grok 4.3 on the same index @ArtificialAnlys

• GDPval-AA v2 Elo 1543 , also ranking #4 , behind Anthropic’s latest Claude releases @ArtificialAnlys

• Top score on τ³-Banking: 33% , above 31% for GPT-5.5 (xhigh) @ArtificialAnlys

• Artificial Analysis Coding Agent Index score 76 in Grok Build, “on par with GPT-5.5 in Codex” and below Fable 5 in Claude Code @ArtificialAnlys

• Cost per Intelligence Index task: $0.31 @ArtificialAnlys

• Cost per GDPval task: $0.49 @ArtificialAnlys

• Cost per Coding Agent Index task: $2.59 @ArtificialAnlys

• Average output tokens per Intelligence Index task: ~14k , over 60% lower than Opus 4.8 @ArtificialAnlys

• Average total tokens per Coding Agent Index task: 1.9M , versus 7.2M for Fable 5 in Claude Code and 6.2M for GPT-5.5 in Codex @ArtificialAnlys

Artificial Analysis’ interpretation was clear: Grok 4.5 is near-frontier on capability, but unusually strong on efficiency, making it sit on the Pareto frontier for cost/performance.

Musk explicitly amplified the Artificial Analysis assessment @elonmusk.