SpaceXAI’s Monumental AI Breakthrough Just Changed The Game

这段视频转录讨论了 SpaceX AI 最新模型 Grok 4.5 划时代的突破，突出了其令人难以置信的能力、速度和经济性。 以下是主要“新闻”和要点的总结： * **前所未有的性能与成本效益：** Grok 4.5 被描述为“与市面上最好的模型不相上下”，并且“在某些情况下甚至超越了最好的模型”，同时它“超高速”且显著更经济，成本约为同类模型的五分之一。 * **成本比较 - Fable 5 High：** Grok 4.5 的表现与 Fable 5 High 不相上下。Fable 5 High 每次任务成本为 8.77 美元，而 Grok 4.5 每次任务仅需 1.51 美元，成本大大降低。 * **价格/性能图表上的“异类”：** 视觉图表显示 Grok 4.5 是一个“巨大的异类”，它是成本最低的模型之一，但表现却达到了成本更高的竞争对手的水平。这被比作“用麦当劳的工资聘请爱因斯坦级别的员工来完成99%的任务”。 * **Cursor 基准测试的注意事项：** 一个小细节提到，Grok 4.5 在 Cursor bench 3.2 上的性能可能略微夸大，因为 Cursor 的整个代码库被意外地包含在内，但这种影响被认为是微乎其微的。 * **其他基准测试证实：** 其他基准测试也证实了 Grok 4.5 在价格、性能和速度方面的卓越性。 * **智能指数排名：** 在智能指数上，Grok 4.5 仅次于 GPT 5.5，略领先于 Sonnet 5，并且“碾压了 GLM 5.2”，尽管 GLM 5.2 的运行成本更高。 * **每任务具体成本：** * Kimmy K26：每次任务约 35 美分。 * GLM 5.2：每次任务约 37 美分。 * Grok 4.5：每次任务约 31 美分（同时取得了更高的分数）。 * Fable 5：完成相同工作需 2.75 美元。 * **编码性能：** Grok 4.5 在编码基准测试中表现出色，“大幅领先于所有谷歌模型”。 * **Token 效率：** 它在编码任务中展现了“令人难以置信的 Token 效率”。在与 Fable 5 和 GPT 5.5 取得相似高分（76-77 分）的情况下，Grok 4.5 仅使用了 200 万 Token，远低于 Opus（920 万）、Fable（720 万）和 Fable 5 on X High（600 万）。这意味着更快的执行速度和更低的费用。 * **XAI 的巨大飞跃：** 该模型的快速进步被认为是“确实令人惊叹”的，表明没有其他实验室能实现如此大的飞跃，这使得 XAI 成为了“Anthropic 和 OpenAI 真正需要警惕的第一个竞争对手”。视频希望“所有人都能用上更快、更便宜、更智能的模型”。 * **飞扬小鸟 (Flappy Bird) 演示比较：** * **任务：** 使用代理式开发（agentic development）构建一个 Flappy Bird 游戏。 * **Grok 4.5 (Marshall) 用时：** 1 分 20 秒。 * **Opus 4.8 (Chase) 用时：** 8 分 59 秒（Grok 快了 8 倍多）。 * **Grok 的游戏质量：** 被描述为“看起来像 3D 的”，“非常像 SpaceX 版的 Flappy Bird”，并且 UI 上有一些“AI 草率之处”，但其速度是首要优势。 * **Opus 的游戏质量：** 风格更佳，但“可玩性差得多”，并存在“图形错误”。 * **演示 Token 成本：** * Opus 4.8：7.9k 输入 Token，31.3k 输出 Token，成本 4.89 美元。 * Grok 4.5：62k 输入 Token，15-25k 输出 Token（估计），成本 0.45 美元至 1.10 美元（Grok 便宜了 8 倍多）。 * **对“工作代理”的影响：** Grok 4.5 被认为是“绝大多数任务的完美工作代理”。在代理式开发中，像 Fable 5 这样的高级模型可以在其 50 万上下文窗口内协调多个 Grok 4.5 代理来完成小型、特定的任务，从而带来更高的成功率、更低的 Token 使用量和更快的完成速度。 本质上，Grok 4.5 被视为一个颠覆者，它不一定在所有任务上都是绝对最好的，但在其智能-成本比方面无与伦比，这表明如果 SpaceX AI 能保持这一领先地位，其业务将“爆发式增长”。

This video transcription discusses the monumental breakthrough of SpaceX AI's latest model, Grok 4.5, highlighting its incredible capabilities, speed, and affordability. Here's a summary of the key "news" and points: * **Unprecedented Performance & Cost Efficiency:** Grok 4.5 is described as "neck and neck with the best models available," and "in some cases better than the best," while being "ultra fast" and significantly more affordable, costing around one-fifth of comparable models. * **Cost Comparison - Fable 5 High:** Grok 4.5 performs comparably to Fable 5 High. While Fable 5 High costs $8.77 per task, Grok 4.5 costs only $1.51 per task, making it drastically cheaper. * **Outlier on Price/Performance Chart:** Visual charts illustrate Grok 4.5 as a "massive outlier," being one of the lowest-cost models yet performing at the level of higher-cost competitors. It's likened to paying a "McDonald's salary for an Einstein-level employee" for 99% of tasks. * **Cursor Bench Caveat:** A small note mentions that Grok 4.5's performance on the Cursor bench 3.2 might be slightly inflated because the entire Cursor code base was accidentally included, but this effect is considered minimal. * **Other Benchmarks Confirm:** Other benchmarks corroborate Grok 4.5's superior price, performance, and speed. * **Intelligence Index Ranking:** On the intelligence index, Grok 4.5 is positioned right behind GPT 5.5 and just ahead of Sonnet 5, "crushing GLM 5.2" despite GLM 5.2's higher running cost. * **Specific Costs Per Task:** * Kimmy K26: ~35 cents per task. * GLM 5.2: ~37 cents per task. * Grok 4.5: ~31 cents per task (while achieving higher scores). * Fable 5: $2.75 for the same work. * **Coding Performance:** Grok 4.5 excels in coding benches, "crushing every Google model by a large margin." * **Token Efficiency:** It demonstrates "insane level of token efficiency" in coding tasks. For similar high scores (76-77) as Fable 5 and GPT 5.5, Grok 4.5 used only 2 million tokens, significantly less than Opus (9.2 million), Fable (7.2 million), and Fable 5 on X High (6 million). This translates to faster execution and cheaper bills. * **Massive Leap for XAI:** The model's rapid advancement is deemed a "truly a sight to behold," suggesting no other lab has made such a jump, positioning XAI as the "first real player that Anthropic and OpenAI have had to be scared of." The hope is for "faster, cheaper and smarter models for everyone." * **Flappy Bird Demo Comparison:** * **Task:** Building a Flappy Bird game using agentic development. * **Grok 4.5 (Marshall) Time:** 1 minute 20 seconds. * **Opus 4.8 (Chase) Time:** 8 minutes 59 seconds (Grok was over 8 times faster). * **Grok's Game Quality:** Described as "3D looking," "feels very much like a SpaceX Flappy Bird," and "a little bit of AI slop" in UI, but its speed was paramount. * **Opus's Game Quality:** Had preferred styling but was "much less playable" with "graphical bugs." * **Demo Token Costs:** * Opus 4.8: 7.9k input tokens, 31.3k output tokens, costing $4.89. * Grok 4.5: 62k input tokens, 15-25k output tokens (estimated), costing $0.45 to $1.10 (Grok was over 8 times cheaper). * **Implications for Worker Agents:** Grok 4.5 is identified as the "perfect worker agent" for the "vast majority of tasks." In agentic development, a higher-level model like Fable 5 can orchestrate multiple Grok 4.5 agents for small, specific jobs within its 500k context window, leading to higher success rates, lower token usage, and faster completion. In essence, Grok 4.5 is seen as a game-changer, not necessarily the absolute best for every task, but unparalleled in its intelligence-to-cost ratio, suggesting that SpaceX AI's business is set to "explode" if it maintains this lead.