The Days Around the Opus 4.7 Launch
Yesterday around midnight, Anthropic dropped Opus 4.7. I had already lain down to sleep, but the news kept me up. I got up and tried it out.
First Impressions of 4.7
There was no "miracle moment" where something I previously couldn't solve suddenly worked on the first try. I happened to lack a tough problem to crack, so I couldn't verify that. But after using it in my daily workflow for a day, a few things stood out.
Instruction following is noticeably better than 4.6. Document requirements and prompt instructions—4.7 is more willing to follow them to the letter. Anthropic's official announcement put it as "takes the instructions literally." You can feel it after switching over and using it for a while.
The previous pattern of "haha, done" only to find the code doesn't run is also much less common. 4.6 moves fast, but often it's hollow speed—too optimistically declaring completion, and it still fails when you run it. 4.7 will verify things itself before handing them over. The release notes mention this: "devise ways to verify its own outputs before reporting back."
There's another small change I didn't expect. I had a medium-sized project that previously took many back-and-forth rounds with 4.6 and Codex to converge. I switched to 4.7 and asked it to go through it: "see if there were any pitfalls back then." It identified several serious issues that no one had realized before, and started fixing them. This improvement in "looking back" ability is quite noticeable even over a short period.
By the way, 4.7 is not the Mythos that was rumored internally. Anthropic has locked Mythos away in the Project Glasswing consortium due to safety concerns, only giving Amazon, Apple, Google, and Microsoft internal access. 4.7 is officially described as a "less broadly capable" version, most likely a smaller, safer branch pruned from the same-generation capability tree. The exact training details haven't been disclosed—we can only guess.
But looking at the pace, I'm somewhat struck. Opus 4.6 was released on February 5th, and 4.7 on April 16th—a new generation in just over two months. That's faster than before. The last time I had this "damn, another generation already?" feeling was in late 2022, shortly after ChatGPT 3.5 came out, when GPT-4 followed in March 2023.
So one judgment habit needs adjusting: don't define the boundary of a problem by what the current model can do. A problem you struggle with today and think impossible might just work with the same approach three months from now, when the next generation drops. This isn't saying you can lie back and wait for the models. It just means that if you're stuck, you can set it aside for a while, let it try a few more times—no need to write it off just because it doesn't work today.
OpenClaude: First Time Merging Code into an External Open-Source Project
A while back, the Claude Code source code leaked. People in my circle quickly started wondering: the CLI's prompt interface and workflow are quite smooth—could we hack it to support multiple models?
At first, I wanted to do it myself. I had Codex scan the Claude Code source code and asked it, "How much work to abstract this layer into a multi-provider interface?" After evaluating, it told me there was a lot to change, and asked if I was sure. I thought about it and decided to skip it—too much trouble.
Two days later, someone in my circle open-sourced exactly this project, called OpenClaude, built on top of the leaked source code, with multi-model support already in place. I downloaded and tried it immediately.
It was a mess of issues right away.
First, the model effort was hardcoded to high—no xhigh. I use Codex on xhigh
Originally published at https://guanjiawei.ai/en/blog/parallel-with-agents