Claude Opus 4.7 dropped from 94.7% to 41.0% on NYT Connections in one release. I moved my work over the day it shipped. Three days later, I had my answer.
MRCR long-context at 1M tokens fell 78.3% to 32.2%. 8-needle retrieval at 256k, 91.9% to 59.2%. Anthropic beat 4.6 on 12 of 14 curated benchmarks. NYT Connections is a third-party puzzle benchmark they don't curate. MRCR is one they're phasing out. I think they shipped a regression and called it a launch.
I saw it in the small stuff first. The new tokenizer eats up to 35% more tokens on the same code.
Extended thinking is gone, replaced by adaptive thinking. Anthropic made xhigh the default in Claude Code because high was "leaving quality on the table." That's an admission.
Then the agent work. I run long Claude Code sessions orchestrating Lambda deployments across a serverless platform. 4.7 started assuming the shape of API instead of reading the actual implementation. Three sub-tasks spawned implementing that imaginary API before I caught it. 4.6 would have opened the file first. 4.7 wants the full map before it moves.
It's not only the model. The harness around it keeps cutting into what's left.
4.7 does solve problems 4.6 couldn't. Long sessions on hard stuff, it goes further than 4.6 ever did. I now need to detail my prompts more to prevent these problems.
The model got smarter. The prompts got longer.
Link of the NYT benchmark Github in comment!