简而言之 SpaceXAI 于 7 月 8 日以每百万输入代币 2 美元和每百万输出 6 美元的价格发布了 Grok 4.5，不到 Anthropic 和 OpenAI 同类模型价格的一半。 Elon Musk 在 X 上发帖称，该模型“与 Anthropic 之前的旗舰产品 Opus 4.7 大致相当”，现已被 Opus 4.8 取代，同时标榜速度和成本超过了基准性能。 Grok 4.5 尚未在欧盟推出； SpaceXAI 表示预计将于 7 月中旬进入欧洲。 埃隆·马斯克 (Elon Musk) 旗下的 SpaceXAI 于周三发布了 Grok 4.5，这是自 2 月份 SpaceX-xAI 合并结束以及 SpaceX 悬而未决的 600 亿美元收购 Cursor 交易以来，其首款公开模型。它的目标人群是程序员、工程师以及公司所谓的“知识工作者”——这一类别显然涵盖了从软件开发人员到审查合同的律师到构建 Excel 模型的财务团队的所有人。
该公司的宣传并不是说它是最好的模型。它很便宜，至少对于西方型号来说是这样。 Grok 4.5 每百万输入代币的成本为 2 美元，每百万输出代币的成本为 6 美元。 Claude Opus 4.8 是 Anthropic 的主要旗舰产品，运行成本为 5 美元，输出为 25 美元。 GPT 5.6 Sol 是 OpenAI 的新顶级模型，也在周三推出，输入价格为 5 美元，输出价格为 30 美元。 马斯克在 X 上发帖并澄清了他的新模型的实际位置。他称其“与 Opus 4.7 大致相当，但速度更快”。 Opus 4.7 是 Anthropic 之前的旗舰产品； Opus 4.8 自此继承了它。 《克劳德寓言 5》现已成为 Anthropic 的顶级产品。 他将此视为一种深思熟虑的权衡：速度和成本高于原始能力，特斯拉和 SpaceX 的工程师将其视为现实世界实用性的证明。 基准测试实际显示了什么
SpaceXAI 在发射时公布了四项基准测试结果，情况好坏参半。 DeepSWE 1.1 使用标准化测试设置来衡量人工智能关闭开发人员提交的真实软件错误的可靠性，以便可以公平地比较模型，并根据已修复问题的百分比进行评分。 Grok 4.5 的得分为 53%，落后于 Claude Opus 4.8 的 59% 和 GPT 5.5 的 67%。 Anthropic 的前沿模型 Claude Fable 5 以 70% 的成绩位居榜首。 在 SWE Bench Pro（另一个衡量按解决率评分的软件工程问题集合）的基准上，Grok 4.5 的得分为 64.7%，足以在该特定测试中击败 GPT 5.5 的 58.6%。 《Opus 4.8》仍然以 69.2% 的比例领先，《神鬼寓言 5》则以 80.4% 的比例领先。 该公司的基准测试是与 GPT 5.5 进行比较，而不是与 GPT 5.6 进行比较，因为后者也在周三发布，即 Grok 4.5 发布几个小时后。 SpaceXAI 与最近收购的 Cursor AI 合作，在 Colossus 内的数万个 Nvidia GB300 GPU 上训练了 Grok 4.5，Colossus 是孟菲斯超级计算机，总容量超过 200,000 个 GPU。在相同的基准测试中其模型高于其模型的实验室不拥有任何接近该硬件的东西。出来的模型具有竞争力，但不是第一。
Grok 在多个版本中都遵循这种模式。 SpaceXAI 一直以巨大的计算成绩和第三名的成绩出现。 Grok 4.5 改变的是定价和训练信号。 实际情况如何 更好的论点不是原始性能，而是性能。这是效率数学。在 SWE Bench Pro 任务中，Grok 4.5 平均使用 15,954 个输出令牌来完成每项作业。 Opus 4.8 为同一作品烧毁了 67,020 个代币，差距为 4.2 倍。 对于大量运行人工智能的团队来说，除了每个代币的价格已经较低之外，这种差异还可以带来真正的节省。即使 Grok 4.5 在质量基准中得分如此之低，更便宜的代币和更高的使用效率允许更多的迭代，而无需花费太多。 该模型还以每秒 80 个代币的速度运行，这是快速模型领域。 Grok 4.5 使用来自 Cursor 的开发人员会话数据进行训练，包括调试跟踪和实际代码编辑，而不是静态存储库。正如马斯克在法庭上承认的那样，xAI 的训练实践之前曾受到过审查；这次管道穿过 SpaceX 正在直接购买的平台。
对于运行大量编码任务的开发人员来说，数学是有效的：每个输入令牌的能力大约为 Opus 4.7 减少 60%。对于任何追求前沿的人来说，《克劳德寓言 5》在 SpaceXAI 选择发布的每个类别中都处于领先地位。我们使用 Hermes 上构建的 Grok 进行的快速测试对于创意写作来说并不令人印象深刻，但对于简单的编码任务来说是可以接受的。 该模型可通过 API 使用，在 Hermes 和 Grok 上构建，具有 50 万个上下文标记（略少于 400,000 个单词）。欧洲用户需要等待才能使用； SpaceXAI 表示 Grok 4.5 将于 7 月中旬到达欧盟。 每日简报时事通讯 每天从当前的热门新闻报道以及原创专题、播客、视频等开始。
In brief
SpaceXAI released Grok 4.5 on July 8 at $2 per million input tokens and $6 per million output, less than half the price of comparable models from Anthropic and OpenAI.
Elon Musk posted on X that the model is "roughly comparable to Opus 4.7," Anthropic's previous flagship, now superseded by Opus 4.8, while touting speed and cost over benchmark performance.
Grok 4.5 is not available in the EU yet; SpaceXAI says European access is expected in mid-July.
Elon Musk’s SpaceXAI released Grok 4.5 on Wednesday, its first public model since the SpaceX-xAI merger closed in February and SpaceX's pending $60 billion deal to acquire Cursor. It targets coders, engineers, and what the company calls "knowledge workers"—a category that apparently covers everyone from software developers to lawyers reviewing contracts to finance teams building Excel models.
The company’s pitch isn't that it's the best model. It's that it's cheap, for a western model at least. Grok 4.5 costs $2 per million input tokens and $6 per million output. Claude Opus 4.8, Anthropic's primary flagship, runs $5 input and $25 output. GPT 5.6 Sol, OpenAI's new top-tier model that also launched Wednesday, is priced at $5 input and $30 output.
Musk posted on X and clarified where his new model actually sits. He called it "roughly comparable to Opus 4.7, but much faster." Opus 4.7 is Anthropic's previous flagship; Opus 4.8 has since succeeded it. Claude Fable 5 is now Anthropic's top of the line offering.
He framed that as a deliberate tradeoff: speed and cost over raw capability, with engineers at Tesla and SpaceX as the proof of real-world utility.
What the benchmarks actually show
SpaceXAI published four benchmark results at launch, and the picture is mixed. DeepSWE 1.1 measures how reliably an AI can close real software bugs submitted by developers, using a standardized testing setup so models can be compared fairly, scored by percentage of issues fixed. Grok 4.5 scored 53%, behind Claude Opus 4.8 at 59% and GPT 5.5 at 67%. Claude Fable 5, Anthropic's frontier model, topped the chart at 70%.
On SWE Bench Pro, another benchmark that measures a collection of software engineering problems scored by resolution rate, Grok 4.5 posted 64.7%, enough to beat GPT 5.5's 58.6% on that particular test. Opus 4.8 still leads at 69.2%, and Fable 5 sits at 80.4%.
The company's benchmarks compare against GPT 5.5, not GPT 5.6, because the latter also launched Wednesday, hours after Grok 4.5's announcement.
SpaceXAI trained Grok 4.5 in collaboration with the recently acquired Cursor AI on tens of thousands of Nvidia GB300 GPUs inside Colossus, the Memphis supercomputer with total capacity across more than 200,000 GPUs. The labs whose models sit above it on those same benchmarks don't own anything close to that hardware. The model that came out is competitive, just not first.
That pattern has followed Grok across multiple releases. SpaceXAI has consistently turned up with enormous compute and third-place scores. What changed with Grok 4.5 is the pricing and the training signal.
Where the case for it actually holds
The better argument isn't raw performance; it's efficiency math. On SWE Bench Pro tasks, Grok 4.5 used an average of 15,954 output tokens to complete each job. Opus 4.8 burned through 67,020 tokens for the same work, a 4.2x gap.
For teams running AI at volume, that difference compounds into real savings on top of the already-lower price per token. Even with Grok 4.5 scoring so low in the quality benchmarks, cheaper tokens and more efficiency in usage allow for more iterations without spending so much.
The model also runs at 80 tokens per second, which is fast-model territory. Grok 4.5 was trained on developer session data from Cursor, including debugging traces and real code edits rather than static repositories. As Musk admitted in court, xAI's training practices have drawn scrutiny before; this time the pipeline runs through a platform SpaceX is in the process of buying outright.
For developers running high-volume coding tasks, the math works: roughly Opus 4.7 capability at 60% less per input token. For anyone chasing the frontier, Claude Fable 5 leads every category SpaceXAI chose to publish. Our quick test using Grok build on Hermes was underwhelming for creative writing and acceptable on a simple coding task.
The model is available via API, on Hermes and Grok build with half a million tokens of context (a little bit less than 400,000 words). European users will have to wait before using it; SpaceXAI says Grok 4.5 reaches the EU in mid-July.
