Claude 4: Totally blown up

1. The loneliness of being the best

Claude Opus 4  and  Claude Sonnet 4 , unveiled by Anthropic tonight  , push the big model from "one-time dialogue response" to "long-term, self-managed thinking entity." From parallel tool calls, writable and readable persistent memory, and "extended thinking" that can be suspended for hours - you will find:

These changes allow AI to, for the first time, be able to leave itself a memo, then review it and continue the task.

• • In an environment where local file reading and writing is allowed, Opus 4 can actively generate "memory files" to write key facts, path selection, and even game maps in segments, and then call them. It no longer relies on rote memorization of "context windows", but has learned to "store → retrieve → use". This allows AI tasks to be extended from minutes to hours or even days, and the technology stack is upgraded from a single function call to a persistent process.
• • In the past, the call chain was linear: reasoning → result → tool → answer. Now the model can pause in thinking, fork to call search, execute code or access the file system, and then return to the main line. Multi-threaded scheduling means that several subtasks can exist in the "consciousness" at the same time, and the overall planning ability is close to that of real humans.

This ability may be more significant than any single benchmark score. Brothers, AI is becoming more and more like a person, organizing knowledge by itself, using and reflecting on progress in parallel! How long will it take for humans to be laid off? !

Code, Code, Code

I have to praise the Claude series for its continued and effective investment in coding. This release has made new improvements in coding, not just at the code level, but also a breakthrough in the overall ecology:

• •  Claude Code  has been upgraded from a research preview to a full-fledged version, with VS Code / JetBrains plugins and GitHub Actions. It not only "generates code", but also uses PR comments and CI fixes to automatically loop back, truly embedding the DevOps closed loop. For developers, the role of AI has changed from a "conversational assistant" to a "second committer on the code flow."
• • GitHub just announced at Build 2025 that the new Copilot Coding Agent will call Sonnet 4 by default, and publicly praised its performance in "agentic scenarios." This marks that GitHub no longer regards the OpenAI model as the only cornerstone, which undoubtedly testifies to the technical strength of Claude and also implies that the relationship between Microsoft/GitHub and OpenAI has shifted from "deep binding" to "depth + width" parallelism, which also testifies to how amazing the strength of the Claude 4 series is.

3. If you can’t improve your running score, you must improve your IQ

What’s interesting is that Opus is not stronger than Sonnet in any area. Instead, Sonnet is stronger in the field of code. This information is actually very interesting and can at least explain several things:

• • Large models that are strong in everything can be a huge challenge, both in terms of cost and quality.
• • Segmentation models are not a bad thing, at least Claude Sonnet offers the best cost-effectiveness for AI Coding.
• • Knowledge is near the end, but intelligence is everything.

4. Other Summary