Claude Opus 4.7 Released: Anthropic Raises the Bar

On April 16th, Anthropic dropped Claude Opus 4.7.

The official announcement was quite modest: “13% improvement over Opus 4.6 on 93 coding benchmarks.”

But those in the know immediately understood—this isn’t about the 13%, it’s about “solving four problems that both Opus 4.6 and Sonnet 4.6 couldn’t crack.”

In plain terms: bugs that stumped previous models? 4.7 can fix them. That’s what matters.

I tested it immediately. I threw a Python concurrency problem at it that Claude had always struggled with—a boundary case involving asyncio, threading, and semaphore control.

The result? Not only did it produce correct code, but it also explained in comments “why the previous approach would cause a race condition.”

Honestly, this exceeded my expectations. I was hoping for “just make it work”—instead I got “make it work, and explain why it works.”

But 4.7’s smartest feature isn’t its coding ability—it’s “knowing what it doesn’t know.”

Low-thinking-mode 4.7 roughly equals medium-thinking-mode 4.6. What does this mean? You can get similar results with fewer tokens and faster speed.

On API costs, Anthropic is finally getting smart. Prices haven’t increased—still $5 per million input tokens.

The 4.7 release takes Claude further down the professional path.