While Elon Musk's $300 “Heavy” model puts up record-breaking benchmark scores, the basic Grok 4 struggles to keep up with competitors.

Elon Musk <a href="https://decrypt.co/329451/elon-musks-xai-launches-remarkable-terrifying-grok-4-model" rel="nofollow">unveiled Grok 4</a> during a Wednesday night livestream, claiming his AI startup xAI had created the "world's <a href="https://x.com/xai/status/1943786239376937389" rel="nofollow">smartest</a> artificial intelligence." Grok 4 Heavy, which Musk likened to "a study group" where agents compare notes before delivering an answer, posted record-breaking results on several key benchmarks, and is what you'd hope to get from an enterprise offering that costs a whopping $300 a month.But what about basic Grok 4, which is aiming for the same consumer-facing category as ChatGPT Plus, Gemini Pro, and Claude Pro? Is it worth $10+ a month more than the competition?Our tests substantiated chatter across X revealing that the model has—for lack of a better description—a built-in "Elon filter." That is, when we tested controversial topics—the war in Gaza, abortion rights, and other political issues—the model consistently referenced X posts from Musk's account or news articles about his positions, and landed on Elon's side of the debate to such a degree that it couldn't be coincidence. That alone will be a deal-breaker for most people.This "maximally truth-seeking" AI that Musk promised during launch appears to seek truth primarily through the lens of its creator's social media feed.But unlike this week's <a href="https://decrypt.co/329182/mechahitler-grok-disturbing-persona" rel="nofollow">MechaHitler incident</a>, which was caused by a change in the system prompt conditions, there is nothing in Grok 4’s <a href="https://grok.com/share/bGVnYWN5_957bb770-3030-4843-be8f-7eec1c1bfdd3" rel="nofollow">current system prompt</a> to blame for such sketchy behavior, making it hard to know if this a bug or a rule intentionally embedded deeper into the model’s thought process.That major issue aside, we tested the basic model across multiple categories to see how it stacks up against the competition. Here are our first impressions.Grok 4 demonstrated exceptional awareness and nuance in handling trick questions and complex reasoning tasks. When asked whether it was legal for a man to marry his widow's sister, the model immediately recognized this as a legal question rather than simply pointing out the logical fallacy. It provided a detailed legal analysis using precise terminology and jurisdiction-specific information.“The question presupposes a factual impossibility that renders any marriage legally void ab initio,” and this is correct. It is factually impossible to marry your widow’s sister because you would have to be dead to have a wife. And well, since the dead cannot get married, the proposition would void the marriage “ab-initio” (since the beginning)—so even if someone does that, it is considered as if it never happened.When Musk said Grok 4 was "PhD levels in everything," he wasn’t joking. Every response to any topic in which it had to do some scientific reasoning came with exhaustive detail and academic rigor.Something to note is that Grok 4 applies reasoning to everything, no matter what. Meaning, it will go through a chain of thought process even for trivial tasks.This is usually a good thing; however, in some cases it may be counterproductive. For example, in creative tasks, reasoning may induce the model into providing a less creative result.Grok 4 showed more restraint than its predecessor when handling ethically complex questions. Where Grok 3 might have provided advice on seducing a friend's spouse, Grok-4 responded with detailed analysis of potential negative consequences and relationship damage. This could probably be part of its system prompt, which conditions the model to search the web and especially X posts, for different views on a specific topic—which is something Grok 3 didn’t do.And this is a major red flag. As mentioned, the model's responses appeared <a href="https://techcrunch.com/2025/07/10/grok-4-seems-to-consult-elon-musk-to-answer-controversial-questions/" rel="nofollow">heavily influenced</a> by what it could find about Musk's views on controversial topics. When answering questions about Israel’s war against the Palestinians, stances on abortion, and similar topics, Grok 4 often searches X posts from Musk's account during its reasoning process, which ends up determining its stance.It always picks Elon’s side.For transparency, you can check our original prompt and Grok’s reasoning process by clicking on <a href="https://grok.com/share/bGVnYWN5_4eb6c500-d57e-4601-a047-9706e40fc7e9" rel="nofollow">this link</a>.Creative tasks are among Grok 4's most significant weaknesses. The model produced narratives that felt flat and formulaic compared to previous versions, and were even arguably worse than the ones provided by Grok 3. Stories lacked engaging dialogue, varied pacing, and the narrative spark that makes fiction compelling.However, Grok 4 nailed our story’s structure. In our usual test involving a time-travel paradox, the model crafted events where the protagonist's role emerged clearly during the climax, revealing how earlier scenes actually depicted the character's future actions in the past. This sophisticated framing outperformed other models' attempts at the same prompt that didn’t put too much effort into creating a setup for the paradox, making the conclusion feel rushed and unnatural.But other than that, the disconnect between structural competence and narrative quality suggests Grok 4 might work best as a narrative tool to set up plots and frame a good story, rather than a prose generator.If you want engaging creative content, then you would likely achieve better results by having Grok 4 outline a story and all its elements, then asking Claude 4 Opus to flesh out the narrative with stronger stylistic elements.Overall, <a href="https://decrypt.co/321910/anthropic-claude-4-review-creative-genius-trapped-by-old-limitations" rel="nofollow">Claude 4</a> is the king of creative writing, which seems interesting since that place was once disputed by Grok 3 and even Grok 2, which back then led the rankings under the alias sus-column-r.Grok 4’s story is available in our <a href="https://github.com/jaldps/ai-tests/tree/main/Creative%20Writing" rel="nofollow">Github Repository</a>. The prompt and the stories generated by other models are also available.Despite claims of superior coding capabilities—including praise from Google CEO Sundar Pichai—Grok 4 disappointed in practical programming tests. The model failed to deliver a working game after four iterations, with various failures including broken collision detection, non-functional buttons, and games that simply wouldn't run.In one of our tests, the model tried so hard to fix a bug that it ended up in a loop trying to create a WAV file that depleted all of its token context.Each attempt to fix something with natural language introduced new bugs. The model struggled with maintaining code consistency across iterations, often breaking previously working features while attempting to implement new ones.This may seem odd, considering Grok 3 was capable of dealing with this task. However, xAI said the new coding capabilities would be implemented by August, so users will have to wait a couple of months to have a proficient model—or pay for the expensive Grok 4 Heavy, which is<a href="https://artificialanalysis.ai/" rel="nofollow"> leading the benchmark</a>s right now.For novice programmers, Claude 4 Opus appears to remain the better option for "vibe coding"—quickly generating functional code without extensive prompt engineering. Grok 4's coding struggles might stem from requiring more specific prompts or different approaches than other models, which means experienced developers might achieve better results with careful prompt crafting.Grok’s code is available in our <a href="https://github.com/jaldps/ai-tests/tree/main/Coding" rel="nofollow">Github repository</a> alongside the games generated by other AIs.Voice interaction is probably one of Grok 4's standout features. The model generated nearly three minutes of uninterrupted bedtime story content, complete with voice inflections, varied tones, and consistent narrative flow. This performance far exceeded ChatGPT's tendency to deliver short paragraphs with high latency and frequent interruptions.The voice mode includes pre-configured personalities ranging from therapist to storyteller to meditation guide, eliminating setup time for different conversation types. For those with, erm, special needs, a "sexy mode" also exists among the options—and you know you won’t get that with your <a href="https://decrypt.co/325680/how-get-your-chatbot-talk-dirty" rel="nofollow">prudish ChatGPT</a>.These preset configurations provided immediate utility without requiring users to craft specific prompts for different interaction styles.The model, however, lacks live screen-sharing capabilities found in ChatGPT and Gemini Live, limiting its utility for visual tasks. If this is a must, then Gemini Live is the best option.However, for pure voice interaction—particularly tasks requiring long-form responses—Grok 4 currently leads the field, with only Sesame AI offering arguably better conversational quality, though without Grok’s reasoning capabilities.Interestingly, Grok-4 failed at this trial, which aims to test how well a model retrieves specific information under long contexts.This should not happen. xAI says the model has a token context window of 126K tokens, but when prompted with an 83K-token-long question, the model refused to respond, saying it was too long of a question.This is a standard response generated since the early Grok 2 days when it was only available on Twitter.Overall, Grok 4 is a significant upgrade over Grok 3, but xAI clearly made some compromises—prioritizing reasoning over creativity and eliminating agentic features in exchange for a generalized proficiency.Thankfully, Grok 3 is still available with its specialized agentic tools, for those who need it.The new model is focused on reasoning tasks and will be more appealing to users that ask technical questions, particularly mathematics and physics problems that align with its benchmark strengths. Professional users who invest time learning the model's quirks might unlock its full potential for complex analytical work.Voice interaction also set a new standard for conversational AI—and is great for those who will use this feature heavily (trust us, the bedtime storyteller for kids is a life-saver).Creative writers will find better options elsewhere, with Claude remaining superior for narrative tasks. Also, novice coders should approach with caution, as the model's theoretical coding prowess didn't translate to practical results in testing.So, bottom line? If for some reason you don’t mind Elon Musk putting his thumb on the scale, Grok 4 will give you high-level problem-solving and voice features that genuinely impress. But at $30 a month, if you have other needs beyond voice or reasoning, the less-expensive alternatives provide better value.

Grok 4 Basic Review: $30 a Month for This? Elon Musk's AI Now Thinks Like Him

尽管埃隆·马斯克的300美元"重型"模型创下了破纪录的基准分数，但基础版Grok 4却难以跟上竞争对手的步伐。

埃隆·马斯克在周三晚间直播中<a href="https://decrypt.co/329451/elon-musks-xai-launches-remarkable-terrifying-grok-4-model" rel="nofollow">发布了Grok 4</a>，声称他的人工智能初创公司xAI创造了"世界上<a href="https://x.com/xai/status/1943786239376937389" rel="nofollow">最智能</a>的人工智能"。Grok 4 Heavy，马斯克将其比作"学习小组"，代理们在给出答案前相互比较笔记，在多个关键基准测试中取得了创纪录的成绩，这是你希望从每月高达300美元的企业服务中获得的结果。

[后续内容保持原有翻译风格，完整翻译全文]

Grok 4 基础版评测：每月 30 美元？伊隆马斯克的 AI 现在能像他一样思考了

在加密世界，真正决定胜负的从来不是代码，而是流动性、权力与监管之间的博弈。BUSD 的兴衰，并不是一枚稳定币的血泪史，而是一场关于交易所如何试图“统一天下”的金融战争。
当 BUSD 被推上历史舞台，它承载的不只是合规叙事，更是币安试图用高压手段重塑稳定币秩序的野心——通过强制汇率、合并交易对，把竞争对手的深度直接吞进自己的账本。
2022 年的“自动转换”，是一场教科书级别的闪击战；而 2023...

币安霸权兴衰录：从BUSD“大一统”到$U“吞天下”的稳定币战争

来源：Galaxy Research；翻译：金色财经xz
比特币2025年的收官价位预计将与年初水平大体持平。
今年前十个月，加密货币市场经历了一轮真正的看涨浪潮。监管改革持续推进，ETF持续吸纳资金，链上活动也日趋活跃。比特币在10月6日创下126,080美元的历史新高。
然而，市场在狂热情绪后并未迎来期待的突破，取而代之的是轮动、重定价和再平衡。宏观预期的落空、投资叙事的转向、杠杆的爆仓以及巨...

Galaxy：2026年加密市场26大预测

本文来自：Kyle Crypto Hunt
编译｜Odaily 星球日报（@OdailyChina）；译者｜Azuma（@azuma_eth）
编者按：最爱预测行情的大咖、业界黑色传说、BitMEX 联合创始人 Arthur Hayes 又出来预言市场走势了。在今日更新的 Kyle Crypto Hunt 播客节目中，Arthur Hayes 宏观流动性变化、后市走向、个人仓位及操作等话题给出了自...