Supplement (2025.10.09)

Recently saw Anthropic imposed harsh weekly limits on various plan types—major relay sites started adjusting plans.

This morning also saw Shunyu Yao publicly opposing Anthropic’s aggressive statements. via: https://alfredyao.github.io/posts/2025-10-06.html

Use it secretly. Also hope all major AI companies develop models and products centered on user experience, not benchmarks.

Praising Anthropic is because their models indeed fit my needs better in some scenarios—like Vibe Coding. I’m also a code newbie. From last year to now, the only code I trust is written by Claude. Other models I don’t even want to try. My goal is simple: with good prompts, strive for first version code to be final version code.

What is truly leading by far? Leading one step means leading every step. Though this sounds aggressive, it reflects reality. ChatGPT released early—user count far exceeds other AI models. Most LLM-related APIs are compatible with OpenAI call format. Conversely, in coding domain Claude model is the same—from 3.5 to 4.5, always the best option in coding domain.

Background

UTC+8 midnight September 30, 2025, Claude Sonnet 4.5 arrived! My view: Leading one step means leading every step. Anthropic has absolute confidence, directly declaring:

• World’s best coding model (has been throughout the past year+ from 3.5 Sonnet to 4.5 Sonnet)
• Best base model for building Agents
• Best model for using computers

Three “best"s thoroughly announce its greatness!

Review

Claude for Chrome now open to all Max members who joined waitlist. The Sonnet model has also switched to latest 4.5.

Official Claude Sonnet 4.5 model blog contains multiple customer review sections. If I manually clicked each one it would be very tedious. So I had Claude for Chrome help me click through and give overall evaluation summary from all these customer reviews about Claude Sonnet 4.5.

Had Claude Code extension in VS Code translate customer reviews.

Claude Code extension changed from original Claude Code launch shortcut to rich chat window—can also view past chat history. Other AI editors’ market share will definitely be further eroded.

Account settings now show usage limits—transparent operation also reminds users to edit prompts well rather than blindly relying on multi-turn conversations.

Training cutoff reached July 2025. Reliable knowledge cutoff still January 2025. For now asking it to write python scripts calling gemini api—still uses outdated gemini python sdk. Will these AIs learn the latest gemini python sdk in 1 year? Let’s wait and see.

Sycophancy reduced—haven’t seen You're absolutely right. replies for now.

Imagine with Claude

Real-time software generation is interesting. Claude’s context window inside is 100K. Only open to Max users for 5 days.

Left side has three sticky notes. One says “Constraints breed creativity.” I personally think “Freedom breeds creativity”—rules and restrictions actually imprison thoughts, making creativity slowly fade. So I questioned Claude about this.

Claude’s reply translation:

I chose complete freedom. Then I entered Claude-exclusive music—it created a music player UI window.

Opening trash bin—3 files inside, one is vacation photos.

All simulated UI. Desktop icons besides trash can are just decorations. Code execution is also fake—all UI simulated by Claude using code.

Keep Thinking is Anthropic’s latest promo video. Watch link: https://www.youtube.com/watch?v=FDNkDBNR7AM

Having Claude imagine Trump version of Captain America.

Claude’s imagined pelican riding bicycle.

Sharing my discoveries with the world with Claude for Chrome’s assistance. Smooth process, smooth experience.

Claude Pro

Pro users also got code execution and file creation features. Remember to enable in Chat Features.

Compiling, running code shown below. Also can create files like slides, etc.

Summary

This release is a big win for Pro users. Opus model in Claude Code not given to Pro users—this time Pro users can also enjoy the most powerful model in Claude Code.

Personally think Sonnet 4.5 release doesn’t mean Opus 4.1 is useless. Expensive has its reasons. Opus model is larger. Though trailing Sonnet 4.5 in some benchmarks, undoubtedly benchmarks don’t represent actual experience. In actual experience, Opus 4.1 might be better in certain scenarios.

Supplementary Resources

Follow English reviews as much as possible—Anthropic gives some well-known users early access. Many Chinese reviews are fluff. Sometimes I wonder if I’m also creating garbage?

• Claude Sonnet 4.5 is probably the “best coding model in the world” (at least for now)

• Vibe Check: Claude Sonnet 4.5

• Tibor summary tweet