What a model list never tells you
Where it started
SuperGrok ran a seven-day free trial, and Grok 4.5 was the model everyone was talking about. That was reason enough — I subscribed to see what the fuss was about.
So it's day one of a trial, and the only sensible question is: what can this thing actually do?
grok models
Default model: grok-4.5 Available models: * grok-4.5 (default) - grok-composer-2.5-fast
Two models. Both text. If you stop reading here — and this is exactly where most people stop — the answer is obvious: it's a coding agent, it writes and it reasons, and that's the shape of it.
That answer is wrong. Not slightly wrong. It misses half of what the thing does.
從七天免費試用開始
SuperGrok 有七天免費試用，而 Grok 4.5 是那陣子大家都在講的模型。理由夠了，我就訂下去看看在紅什麼。
試用第一天，該問的問題只有一個：這東西到底能做什麼？
跑 grok models，兩個模型，都是文字模型。看到這裡就停——而大多數人正是停在這裡——結論很明顯：它是 coding agent，會寫程式、會推理，就這個形狀。
這個結論是錯的。不是差一點，是漏掉了它一半的本事。
The list nobody checks
Every agent CLI keeps two inventories, and they answer different questions.
The model list tells you which brain handles your conversation. That's grok-4.5's job.
The tool list tells you what the agent is holding. And that's where capability actually lives.
Almost everyone checks the first one. It takes one line to check the second:
list your available tools
grok -p "list your available tools"
The answer ran past twenty lines — web search, X search, terminal, subagents, a scheduler. Then the last three:
image_gen image_edit image_to_video
It generates images. It edits them. It turns them into video. None of that appears in the model list, and it never will — because generating an image isn't a brain, it's a hammer the agent picks up when it needs one. Two inventories, two questions.
沒人看的那一份清單
每個 agent CLI 都有兩份清單，回答的是不同的問題。
模型清單說的是「你的對話交給哪顆腦袋處理」——那是 grok-4.5 的工作。
工具清單說的是「這個 agent 手上拿著什麼」。而能力住在第二份裡面。
幾乎每個人都只看第一份。看第二份只要一行指令。
回來的清單跑了二十幾行——網頁搜尋、X 搜尋、終端機、子 agent、排程器。然後是最後三行：image_gen、image_edit、image_to_video。
它會生圖、會改圖、會把圖變成影片。這些在模型清單裡一個都看不到，而且永遠不會出現——因為生圖不是一顆腦袋，是這個 agent 需要時會拿起來的一把鎚子。兩份清單，兩個問題。
Same sentence, three agents
What Grok drew
One prompt: a sprig of mint on cream paper, Morandi palette, classical botanical study.
864×1152. I asked for none of the framing — it drew the ruled border itself, and wrote Mentha spicata L. / July 1783 / Herb. Hort. Reg. in the corner in a period hand. An invented provenance, from a royal herbarium that never held this leaf.
Then image_to_video turned the still into six seconds of footage: 480×640, h264, with an AAC audio track I hadn't asked for either. I ran ffprobe on it rather than take the agent's word — it's a habit worth keeping.
Grok 畫的
同一句話：奶油紙上的一枝薄荷，莫蘭迪配色，古典植物標本。
864×1152。外框是它自己加的，右上角那行手寫的 Mentha spicata L. / July 1783 / Herb. Hort. Reg. 也是——一段它自己編的來歷，來自一間從沒收藏過這片葉子的皇家標本館。這些我一項都沒要求。
然後 image_to_video 把靜圖變成六秒影片：480×640、h264，還附了一條我同樣沒要求的 AAC 音軌。我沒有信它的自我宣告，是用 ffprobe 驗過才敢寫——這個習慣值得留著。
What Codex drew
The same question at Codex turned up image_gen__imagegen. So it got the same sentence.
1024×1536. No frame, no invented label. The leaf veins hold up close and the paper has a tooth to it. Between the two, this is the one I'd put on a page.
Then I checked how it signs in:
codex login status # → Logged in using ChatGPT
Not an API key — my ChatGPT subscription. The image generation was inside something I already pay for, and I'd never once asked it what it was holding.
Codex 畫的
同一個問題丟給 Codex，翻出 image_gen__imagegen。於是它拿到同一句話。
1024×1536。沒有外框，沒有杜撰的標籤。葉脈禁得起放大看，紙面有粗糙的質地。兩張擺在一起，我會把這張放進頁面。
然後我看了它是怎麼認證的：不是 API key，是我的 ChatGPT 訂閱。生圖能力一直包在我早就付錢的東西裡面，而我從來沒問過它手上有什麼。
Gemini: this time it really can't
For fairness, the same question went to Gemini CLI. Different kind of surprise:
IneligibleTierError: This client is no longer supported for Gemini Code Assist for individuals. reasonCode: 'UNSUPPORTED_CLIENT' tierId: 'free-tier'
The login was fine. The version was 0.50.0 — the latest on npm. With not logged in and version too old both ruled out, the answer stands: Google has closed the individual free tier for the CLI and wants everyone on Antigravity.
So this one is a real no. But it's a no I ran, not a no I read.
Gemini：這次是真的不行
為了公平，同一個問題也丟給 Gemini CLI，撞到另一種驚喜。
登入是好的，版本 0.50.0 就是 npm 上的最新版。「沒登入」和「版本太舊」兩個假設都排除掉之後，答案很清楚：Google 把 CLI 的個人免費額度收了，要大家改用 Antigravity。
所以這個是真的不行。但這個「不行」是我跑出來的，不是我讀來的。
Who gets the job
Not the smartest — the only one
Three inventories on the table, and the division of labour sorts itself out. The deciding question turns out not to be who's smartest. It's who's the only one.
Images go to Codex — higher resolution, better rendering, inside a subscription I already hold. Video only Grok has: with sound, though small.
But the thing that makes Grok irreplaceable isn't in that comparison at all. It's four tools nobody talks about:
x_user_search x_semantic_search x_keyword_search x_thread_fetch
It searches X. Real posts, real like counts, real dates. Nothing else I own comes close — a web search returns blog posts about the conversation; this returns the conversation.
So that's the job it got: choosing what this article should be about. Before I wrote a word, I had it sweep X for what had already been said. It came back with the worn-out angles first — Codex inside Claude, adversarial review, agent teams — which were, inconveniently, the ones I'd have reached for. And then:
No X evidence of posts explaining that agent model lists hide tool capabilities.
Nobody had written it. So I did.
Grok got that job for one reason, and it isn't intelligence. It's that nothing else could have done it.
不是最聰明的，是唯一的那個
三份清單攤在桌上，分工自己就清楚了。而決定的問題不是「誰比較聰明」，是「誰是唯一的」。
生圖歸 Codex——解析度更高、成果更好，而且包在我本來就有的訂閱裡。影片只有 Grok 有：有聲音，雖然小。
但真正讓 Grok 無可取代的東西根本不在這個比較裡，是四個沒人在談的工具：x_user_search、x_semantic_search、x_keyword_search、x_thread_fetch。
它能搜 X。真實的貼文、真實的讚數、真實的日期。我手上沒有任何東西接近得了——網路搜尋撈回來的是討論那些對話的部落格，這個撈回來的是對話本身。
所以它拿到的工作是：決定這篇文章該寫什麼。動筆之前，我讓它掃了一遍 X 上已經被說過的話。它先帶回那些寫爛的角度——「Codex 接進 Claude」「對抗式審查」「agent 團隊」——不巧，那些正是我會伸手去拿的。然後是這一句：沒有任何 X 貼文寫過「模型清單藏住了工具能力」。
沒人寫過。所以我寫了。
Grok 拿到這個工作只有一個理由，而那個理由不是聰明。是別人做不到。
Notes from the floor
Four things that cost me time, in case they'd cost you the same:
grok.exe isn't on the Git Bash PATH. Call it with the full path from ~/.grok/bin/.
codex exec reads stdin. In a batch script it hangs waiting for input unless you feed it < /dev/null. Outside a git repo it also wants --skip-git-repo-check.
Grok writes .claude/napkin.md into your working directory — its own cross-session memory. In a project that already has Claude Code settings, two agents end up sharing one config folder.
X search is slow — five to eight minutes. It dispatches an agent to go and look; it isn't a lookup. Good for map this topic for me, useless for quick check.
現場筆記
四件花了我時間的事，免得也花你的：
grok.exe 不在 Git Bash 的 PATH 上，要從 ~/.grok/bin/ 用全路徑叫它。
codex exec 會讀 stdin。寫在批次腳本裡，不餵它 < /dev/null 就會卡在那裡等輸入。不在 git repo 裡還要加 --skip-git-repo-check。
Grok 會在你的工作目錄寫 .claude/napkin.md——它自己的跨 session 記憶檔。在一個已經有 Claude Code 設定的專案裡，兩套 agent 就共用同一個設定資料夾了。
X 搜尋很慢，五到八分鐘。它是派一個 agent 去看，不是查表。適合「幫我把這個主題摸清楚」，拿來「快速查一下」是浪費。
Day one
The trial has a few days left on it, and I still haven't decided whether to keep paying. But the question that decides it won't be whether Grok 4.5 is smart. Plenty of things are smart.
It'll be whether those four X tools change what I make. That's the only thing on the list nothing else can do.
Which is the same lesson as the rest of this, really. Before you decide what a tool is worth — or what it can't do — ask it what's in its hands. Not the pricing page. Not the model list. The tool.
第一天
試用還剩幾天，我還沒決定要不要繼續付。但決定的問題不會是「Grok 4.5 聰不聰明」——聰明的東西多的是。
會是那四個 X 工具有沒有改變我做出來的東西。那是整份清單上，唯一別人做不到的。
說起來，這跟前面講的是同一件事。在你決定一個工具值多少、或它不能做什麼之前，先問它手上有什麼。不問價目表，不問模型清單。問工具本人。
— TingYu's Deco
Ask the Tool, Not the Docs · 2026 · Taiwan
