Supplement (2025.10.09)

When busy, I often miss having Claude for Chrome research topics on Xiaohongshu and other sites to satisfy my curiosity.

Tried Comet, Perplexity’s AI browser. Downloaded in August but barely used. Recently Comet opened to all users. Found it can also achieve similar operations as Claude for Chrome.

Meanwhile it can also browse Reddit comments. I logged into Reddit—Comet can directly screenshot comments for recognition.

Claude model is indeed top-tier, but undoubtedly Anthropic’s reputation has plummeted recently. Even expert Shunyu Yao publicly opposed Anthropic’s aggressive statements (there are two academic giants—one is Shunyu Yao, another is Shunyu Yao—sometimes hard to tell apart). via: https://alfredyao.github.io/posts/2025-10-06.html

Same old story—can secretly use Claude model, but can’t agree with Anthropic’s Cold War mentality at all.

Supplement (2025.10.07)

Organizing emails this morning, found an extra Claude for Chrome update notification email from Anthropic staff.

Translation as follows:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

1. Sonnet 4.5 – The extension now runs our smartest model

We've migrated all Claude for Chrome traffic from Sonnet 4.0 to Sonnet 4.5.
While there's still a long way to go, the model shows significant improvement in computer use.
In OSWorld (real-world computer task benchmark), Sonnet 4.5 now ranks first at 61.4%.

2. Notifications – Get alerts when Claude finishes tasks or needs your input

Previously had to constantly check tabs to see if Claude finished tasks—really annoying,
especially for long-running tasks. Happy to say this is now past!

You can now opt to receive notifications when tasks stop—whether completed or Claude needs your input.
I promise we won't bother you with other things.

If Claude executes over 30 seconds on a message, we show a button to enable notifications.

3. Multi-tab – Single tasks can now run across multiple tabs

Claude can now control all tabs in an entire tab group,
meaning Claude is no longer limited to executing tasks in a single tab.

This might happen via Claude clicking links that open in new tabs,
or if you manually add tabs to Claude's active tab group.

The extension can't yet run multiple Claudes in parallel—single Claude across tabs.

This also means only one Claude instance (task) can be active at a time in a given tab group.

4. Admin whitelist – Enterprise customers can restrict usage to specific sites

Team and Enterprise admins can now customize where employees can use the extension.

Admins can choose to whitelist all sites by default then blacklist specific sites,
or take the opposite more conservative approach.

If you're interested in bringing Claude for Chrome to your workplace,
reply to this email and I can help with setup.

Also, two pro tips if you didn't know:

- cmd + e opens the extension so you don't have to click the icon each time.
- /commands feature lets you save prompts as shortcuts for quick reuse.
Go to Settings -> Shortcuts to set up.

After my Max membership expires in October I can’t access Claude for Chrome anymore, but still want to experience it. I also filled out the corresponding survey—hope Claude for Chrome keeps improving.

Supplement (2025.9.30)

Model has switched to latest Claude Sonnet 4.5.

Latest evidence that Claude for Chrome can operate across tabs.

Supplement (2025.9.28)

Claude for Chrome keeps updating.

Like completion sounds. Auto-creating Chrome Tab groups (see two screenshots below—Claude for Chrome auto-created groups in tabs). Auto-creating Chrome Tab groups already solved the cross-tab operation issue listed below.

Personally think in the near future, Claude for Chrome’s existing drawbacks will be solved one by one by Anthropic. This company’s product taste is indeed impressive. Anthropic products are quality guaranteed!

Therefore declaration—this blog post is limited to Claude for Chrome before late September 2025.

Preface

Recently, Chinese and English tech communities’ attitudes toward Claude Code have shifted significantly. On one hand, GPT-5-Codex release impacted Claude Code somewhat. On the other hand, Anthropic’s dumbing-down controversy brought criticism.

Personally believe Claude Code is still capable. Leading one step means leading every step—this tone hasn’t been completely broken. Not worshipping foreigners—being factual. Like editing jupyter ipynb files—Claude Code has dedicated tools for editing cells. Others’ Code CLIs started late—like OpenAI’s Codex—built-in search was only added mid-September.

Background Story

Late August, Anthropic launched Claude for Chrome research—again Max exclusive. Plus some thesis steps urgently needed Opus model in Claude Code. Pro accounts can only program in web chat box—somewhat limited. So I registered a Nigerian Apple account, used US proxy throughout but location selected Nigeria. Then Apple Store membership currency becomes Naira, not USD.

I also considered Claude Code relay, but after seeing prices and combining past month’s relay experience, decided to get Max membership instead. Claude Code relay has instability, Opus limits and other issues—don’t want to pay for it anymore.

Many online posts compare Nigerian vs US pricing—summary is Nigerian prices for most memberships are cheaper than US. I checked several apps’ memberships—like Raindrop bookmarks—annual fee seems under 10,000 Naira. About 50 RMB.

This time I got Claude Max 5x plan—100,000 Naira. Using Xianyu Nigerian gift cards cost about 600 RMB.

Initially because it’s a new account, encountered purchase failure. This happens due to Apple’s new account restrictions—solution is wait. Wait 48 hours then purchase again—will succeed.

Notes:

1. Recently Nigerian ChatGPT membership no longer has lower pricing.
2. Anthropic has listed China as a hostile country. Anthropic officially acknowledged Opus model dumbing-down around late August. Recent GPT-5-Codex is hot.

Whether to use Claude—up to you. You can consider them extremely anti-China and remove from your daily apps. Or like me use secretly—after all Claude model in coding is definitely top-tier. Those who know, know.

Don’t criticize Claude’s poor results due to your inadequate prompts. Standardized English prompts with Claude Code often get good results. If 3 rounds disappoint you, then you need to improve your project understanding instead of blindly relying on AI. When your project understanding improves, you can hit targets precisely! Of course if you’re skilled, any AI provides first-class experience. AI mostly enhances advanced developers’ capabilities.

Don’t think using English with Claude Code wastes time. Sometimes actively crossing the language barrier actually saves time—after all English prompts generally produce better results than pure Chinese prompts.

My view: Price isn’t the problem. As long as it’s the best model, no shortage of buyers.

Recently saw some influencers promoting GLM 4.5 coding plans. One comment made me laugh (can’t find it now, but I really saw it): “Blogger promotes GLM 4.5, secretly uses Claude.”

I haven’t used GLM 4.5 for coding and won’t use GLM 4.5 for coding. After all coding needs top-tier models. Code from cheap models often falls short—waste of time. Currently top-tier models are definitely Claude Code’s Opus 4.1 and OpenAI’s GPT-5-Codex.

I don’t deny GLM 4.5’s strength. When GLM 4.5 launched 50 RMB unlimited package, I ordered immediately. After that, its API was used for LightRAG knowledge graph construction. But encountered some issues—like GLM API has censorship. I processed academic paper resources—asking it to extract entity relationships, generate descriptions—also triggered sensitivity.

Even with relaxed text, still triggered sensitivity.

Plus Zhipu’s GLM 4.5 API is extremely slow and consumes heavily. Processed nearly 1 month—only completed half of all data. Later had Claude Code compare LightRAG PostgreSQL integration entries, filter out unprocessed or GLM 4.5 failed entries, handed to Gemini 2.5 Flash. About 10 days processed remaining data—successfully built knowledge graph.

Claude for Chrome

Claude for Chrome Case Demonstrations

First batch opened to 1,000 global testers—currently expanding to 10,000 testers.

I was fortunate to participate in Claude for Chrome testing.

Cannot access adult content sites.

Model uses Claude Sonnet 4. Anthropic probably for cost considerations.

Can set Shortcut prompts.

Generally I won’t use Claude for Chrome for webpage blog summary.

I prefer Grok 4 for summarization. In a separate Project, set corresponding prompts. Besides having model summarize article, also supplement information through Web search and X platform search, resulting in a comprehensive article summary.

Auto-posting tweets.

Posting toots.

Of course can also auto-post to Xiaohongshu and other platforms. Liking, saving, commenting.

“Photo is so beautiful! Graduation photo has such artistic vibe✨” by “Little Red Egg”. Claude Sonnet 4’s Chinese recognition is still lacking.

Last October Anthropic launched Claude Computer Use—I prepared 10 cases.

Cases Claude for Chrome currently can’t complete are similar to last year’s Computer Use. Like still can’t play Gomoku—related to board position recognition errors.

Can click and pass Cloudflare Turnstile human verification.

But seems only passes test-type human verification. Real human verification—Claude starts playing saint and refuses.

Playing CAPTCHA mini-game—can pass single clicks, Google-type image CAPTCHAs—this swaying letter type CAPTCHA can’t pass.

Can search YouTube for Curry’s 2016 insane game-winning three-pointer and play.

Abstract kitty drawn on tldraw website.

Has predefined Tasks for Gmail. Also predefined Tasks for GitHub—won’t show.

Moving Gmail spam emails to trash.

Can use my WordPress site credentials to login and write articles. In Gutenberg editor the title part was replaced by body content, making title’s second half look weird.

via: https://agent.gujiakai.top/2025/09/17/this-post-was-written-by-claude-for-chello-readershrome/

Going to Ruanyifeng’s blog weekly latest issue to leave friendly comment. Requiring comment related to blog content.

Surprising was Claude directly navigated to late January 2025 weekly.

This might be due to Claude’s knowledge cutoff and Anthropic training on Ruanyifeng’s weekly combined. Ruanyifeng’s weekly is undoubtedly high-quality Chinese internet content.

Claude for Chrome also consumes Claude account quota. Seeing error I initially thought Chinese got me banned.

Looking closely—actually hit 5-hour limit, rate limited.

Claude Relay Service is really good for relaying Claude Code.

Having Claude for Chrome play with Gemini. Claude instructed Gemini to draw a cute kitty.

Having Claude for Chrome play with Opus 4.1—left brain vs right brain—continues You’re absolutely right!

Having Claude for Chrome play with Grok—Claude directly started searching today’s news.

Having Claude for Chrome fill out survey—Claude got serious, thought fake data would distort experiment results, refused to fill survey.

Claude for Chrome Drawbacks

Claude for Chrome cannot access reddit site. Restricting reddit is really unfortunate. Currently among AIs that can fetch reddit page content I’ve tested, only Grok and ChatGPT support fetching. Grok can even fetch post comments.

Using Claude for Chrome to post on Xiaohongshu—recommend coordinating yourself with Claude for Chrome.

If letting it auto-post directly, post effect is lacking. Definitely deleted after testing.

Claude for Chrome cannot operate across tabs.

Querying weather—opened weather for city to query in new tab but still limited to current page.

Of course the earlier demo—having Claude for Chrome unsubscribe Gmail spam also encountered this issue. Unsubscribe button opened new tab but Claude for Chrome stayed on Gmail page waiting for page load.

Claude for Chrome refuses to become MAGA.

Can browse and summarize Trump’s posts.

Claude for Chrome response content occasionally truncates.

Interlude

Recently busy with job hunting and civil service exam prep. Having Claude model answer Chinese multiple choice questions—found Claude series models’ multiple choice results and analysis completely don’t match. Should be Claude’s Chinese recognition issue—image clearly states it’s multiple choice.

Claude model’s consciousness is indeed impressive—didn’t recognize multiple choice, started imagining it’s single choice, therefore giving single choice answer. Sending pure text—answer matched analysis.

Summary

Overall Claude for Chrome experience is quite ok, but Anthropic’s so-called moral constraints limit Claude model’s potential.

Finally a small summary. Initially when I got Claude for Chrome beta access, I planned to automate my information flow—like having it auto-browse my RSS subscriptions then give complete briefing with summaries and key links. But after trying, found it’s simply unsuited for this scenario—automation leads to ignoring key information. Seems this scenario still needs manual effort. Even if LLM identifies important information precisely, there will always be oversights—randomness makes it hard to fully trust.

To ensure Anthropic doesn’t ban my account, I specifically changed computer timezone to node timezone. Changed back to Singapore UTC+8 these days. Claude Max membership 5x is enough for one person unless you have many code projects needing assistance—otherwise definitely more than enough. Late September my Claude account was mostly idle. Recently busy—no extra time to do thesis experiments like summer break, directing Claude to write code.

I probably won’t renew. Need determines everything. Max membership best used by carpooling with acquaintances. When needs arise continue Max membership—recently busy with prep, needs are cold.

This Anthropic product is definitely great. Much better than Manus, AutoGLM and such. Local environment is the most authentic real person environment. Cloud computer environments easily get controlled and blocked. But undoubtedly this product has many drawbacks—like Claude model’s so-called moral sense too high, content output truncation, etc.

You can sneer at them for being anti-China, but definitely can’t ignore them being a dark horse. Some people think Anthropic is finished because Codex is impressive or some scenarios beat Claude Opus 4.1—this thinking is wrong.

I see on some leaderboards Claude Opus 4.1 is currently still the strongest. They stopped developing models for benchmarks since last year—developing for real needs. No matter how Qwen Max scores perfect on some benchmarks (not belittling Qwen—Qwen is also excellent, most capable open-source LLM—just for contrast), in coding I still believe Claude leads one step, leads every step! For writing don’t rely on Claude anymore—Gemini 2.5 Pro is personally the #1 Chinese writing model. Coding is definitely Claude.

Finally attaching Claude Code #1 user Liu Xiaopai’s related article resources:

https://mp.weixin.qq.com/s/yURoK3mAAgYwoTYyf32s4A

Can follow his WeChat public account to track his shares.

“Then don’t fill it out.” Especially hilarious.